# Audience, decisions, and rollout for the state of AI-assisted data visualization

Status: audience and distribution strategy, 16 August 2026.

Evidence cut: 2026-08-16. Community routes and event dates should be checked
again immediately before any outreach. This document recommends a rollout; it
does not authorize publication, community posting, submissions, or direct
outreach.

Companions: [editorial architecture](/backstage/reports/editorial-architecture/),
[executive summary](/reports/executive-summary/),
[research field guide](/reports/research-review/), and
[practitioner and reader experience](/reports/practitioner-and-reader-experience/).

## Recommendation

Do not launch this as one large report for one imagined audience. Launch one
canonical evidence hub with several decision-shaped entry points.

The underlying research has at least six credible reader groups. They share an
interest in AI and visualization but arrive at different moments, use different
language, and need to make different decisions. The comprehensive report is the
evidence spine. The useful public units are shorter guides, scorecards, figures,
methods, and source tables that help one group decide something now.

The first four audiences should drive the package:

1. visualization and data practitioners deciding where AI belongs in a real
   workflow;
2. BI and analytics leaders deciding what to enable, buy, govern, or stop;
3. visualization, HCI, and AI researchers choosing studies, benchmarks, and
   claims; and
4. data journalists, graphics editors, and newsroom developers deciding what is
   safe to use and publish.

Product builders and educators are important secondary audiences. General
executives and broad AI readers should receive a concise brief, not set the
structure of the research.

The positioning should be literal:

> What AI chart and dashboard systems can do in 2026, where they fail, and how
> to evaluate them for real work.

Use **AI-assisted data visualization** as the umbrella. Use the audience's own
more specific terms in route titles and descriptions: **AI data visualization
tools**, **Power BI Copilot**, **Tableau Agent**, **conversational analytics**,
**agentic analytics**, **agentic visualization**, **LLM visualization
generation**, **coding agents for data analysis**, **semantic models**, and
**reader trust**. Do not force one vocabulary across every surface.

## What is observed, inferred, and still unknown

This strategy combines three evidence classes:

- **Observed:** the current report files and rendered artifacts; official
  descriptions of community access, editorial terms, conference programs,
  search-measurement and citation mechanics; and bounded public language
  samples. No site-specific query or route-use data entered this synthesis.
- **Inferred:** audience priority, decision moments, suitable artifacts, and the
  sequence in which those artifacts should travel through the observed routes.
- **Unknown until tested:** how many people will read, refer, cite, subscribe,
  invite, or change a decision because of this work.

The audience map is therefore a testable model, not an adoption claim. Community
membership counts do not establish interest in this report. A visible page does
not establish use. A citation does not establish a changed decision.

The source trail for the audience scan is preserved in a private evidence
ledger, while this public-facing strategy links to sources that can be checked
by a reader. No private captures or copyrighted source text should travel into
the public package.

## The actual product

The phrase “report rollout” can hide the wrong unit of work. A reader rarely has
the job “read a state-of-the-field report.” They have a question such as:

- Can this tool be trusted with our semantic model?
- Which parts of my workflow should I delegate, and which must remain mine?
- What evidence would show that an agent made the work better rather than merely
  faster?
- Which benchmark or study would close a real gap?
- What can a newsroom permit without weakening source, editorial, or reader
  accountability?
- What should a student still learn when a model can produce the first chart?

The public offering is therefore a family of **evidence-backed decision guides**
built from one maintained evidence base. Internally, **decision products** is a
clearer name for this family than “automated consulting.” “Automated consulting”
foregrounds the production mechanism and invites a comparison with consulting
labor. “Decision product” names the object and its use. For public language,
“evidence-backed decision guide” or “maintained decision guide” is more legible
and less grandiose.

### One validity result, seven audience questions

The latest configuration audit has one canonical answer: **no benchmark result
or prepared protocol currently authorizes a Vizier configuration winner.** The
translation changes the decision language, not the truth:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Which exact part of my job did this score test, and what still needs me or a real reader? | Evidence-lane checklist beside the workflow evaluator | A high score is not accepted, delivered, readable, or maintainable work. |
| I run analytics | Does the evidence reach our semantic model, permissions, users, incidents, and total cost? | Five-gate card beside the pilot scorecard | Benchmark or vendor success does not clear organizational acceptance. |
| I research visualization | Which edge is measured, which is a cross-literature bridge, and what falsifies it? | Full validity register, benchmark crosswalk, and preregistration | Convergent vocabulary is not a causal architecture result. |
| I publish data journalism | Where are the source, editorial, accessibility, and reader receipts? | Publication checklist that refuses benchmark substitution | Model literacy or rubric agreement is not publication authority or audience comprehension. |
| I build AI systems | Does the critic, skill, or extra agent add unique evidence at equal resources on the current harness? | Fail-closed ablation card with model/harness vintage | An older or unequal-budget lift cannot select the present architecture. |
| I teach or evaluate access | Was unassisted transfer or representative assistive use measured, or only task completion? | Measurement-ladder note beside the learning/access protocol | Human-designed literacy items are not learning, transfer, or accessibility outcomes. |
| I sponsor or edit | Is there a current accepted comparison, or only a plausible theory and prepared protocol? | Direct answer: no winner; five validity gates remain | Evidence plumbing and synthesis do not authorize adoption. |

Every shorter layer should point to the complete threat audit and carry the
evidence date and reopen condition. This is audience translation, not observed
audience demand or proof that any route has helped a reader.

### One system-anatomy result, seven audience questions

The [complete field-guide account](/reports/research-review/#a-reference-anatomy-assigns-information-and-authority)
has one canonical answer: **six of six audited authoring or analysis systems
separate generation, execution, and some critique route; four expose meaningful
human control; zero reaches accepted delivery, intended-reader outcome, or
maintenance.** The translation changes the next decision, not those counts:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Where can I inspect, correct, branch, or stop—and who says this exact artifact is ready? | Workflow overlay from intent through human acceptance | A runtime or critic can reject bounded failures; it cannot accept the chart for the practitioner. |
| I run analytics | Where do authoritative measures, permissions, joins, acceptance, incidents, and later updates enter? | Governance overlay on the eleven-stage map | Five of six systems ground to data; none supplies organizational delivery or maintenance evidence. |
| I research visualization | Which transition is manipulated, which information is visible, and which later gate is merely inferred? | Machine-readable stage matrix plus exact configuration and falsifier fields | Component convergence is not one causal architecture result; downstream zeroes are held-corpus results. |
| I publish data journalism | Where are source custody, transformation, render, editorial, accessibility, delivery, reader, and correction receipts? | Publication path that keeps editorial acceptance separate from model critique | An editor with the source trail accepts the exact artifact; a critic never gains publication authority. |
| I build AI systems | What does each boundary read, write, execute, reject, repair, and escalate? | Instrumented information-and-authority contract with equal-budget ablations | Agent names do not define specialization; do not let an evaluator write its own acceptance rule. |
| I teach or evaluate access | Does the generated interaction preserve source, state, keyboard/assistive paths, unassisted learning, and real reader use? | Learning/access overlay on representation, delivery, and reader stages | ViviDoc's author control and output ratings are not learning, accessibility, transfer, or delivered-reader evidence. |
| I sponsor or edit | Which stages are demonstrated, and where does human or field evidence begin? | Direct answer: 6/6 technical loop; 4/6 human control; 0/6 downstream chain | Invest in the missing owner and audience receipts, not a larger agent diagram. |

Every layer points to the same eleven-stage account and keeps critic authority,
acceptance, delivery, reader evidence, and maintenance separate. This remains a
modeled route design, not observed audience use.

### One acquisition result, seven audience questions

The [complete field-guide account](/reports/research-review/#public-acquisition-is-not-adoption)
has one canonical answer: **twelve named skill listings carry public install
signals, but zero held row exposes real-task invocation, repeat use,
organizational acceptance, or outcome.** The twelve rows collapse to ten parent
repositories and eleven documented lineages. The translation changes the next
decision, not those counts:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Should I try the highly installed package? | Seven-stage trial card beginning with the exact task and ending with retained use | Installs shortlist; a local no-skill comparison decides whether to keep it. |
| I run analytics | Does popularity justify enabling or procuring it? | Pilot ledger joining version, governed data, successful task use, reviewer acceptance, incident path, and total cost | No named public row reaches organizational acceptance. |
| I research visualization | What should an adoption study measure? | Registry-to-outcome cohort design with provenance, exposure, invocation, retention, and outcome denominators | Do not call acquisition or benchmark execution adoption. |
| I publish data journalism | Can a popular package enter a publication workflow? | Intake card for source custody, approved task, exact package, editorial review, reader/access checks, and correction | Repository attention grants no publication authority. |
| I build AI systems | What instrumentation is missing? | Event schema separating attempted install, success, invocation, outcome, return use, removal, and version migration | Privacy-respecting instrumentation needs explicit denominators and opt-out. |
| I teach or evaluate access | Does availability improve learning or access? | Classroom/access trial joining allowed assistance, representative use, unassisted transfer, and accessible delivery | Install and task-completion signals do not establish learning or access. |
| I sponsor or edit | Is this ecosystem being adopted? | Direct answer: 12/12 public install signals; 0/12 behavioral or organizational rows | Say public acquisition signal, not adoption, until later receipts exist. |

Every layer points back to the same provenance-aware answer. These are intended
routes, not evidence that any audience has used or accepted the layer.

### One governance result, seven audience questions

The [complete practitioner account](/reports/practitioner-and-reader-experience/#available-controls-are-not-governed-deployment)
has one canonical answer: **three provider families expose meaningful controls,
two provider stories report named-feature organizational use, two studies add
adjacent governance-process evidence, and zero of seven held rows joins a
complete governed deployment.** The translation changes what each audience
does next, not those counts:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Which permissions, semantic objects, review, refusal, and escalation rules govern this exact task? | Plain-language deployment card beside the workflow evaluator | A product toggle or governed semantic layer does not accept the resulting artifact. |
| I run analytics | Can we produce all nine receipts for this feature, scope, owner, audit path, incident, outcome, and recheck? | Same-deployment governance ledger beside the pilot scorecard | Available controls and a provider customer story cannot be combined into our approval record. |
| I research visualization | Which control was actually configured, which organizational behavior was observed, and what denominator or comparator supports the outcome? | Seven-row matrix, source class, missing-receipt state, and reopen rule | Provider documentation is mechanism evidence; customer stories are attributed use reports, not independent efficacy or prevalence. |
| I publish data journalism | Who owns source authority, editorial acceptance, audience harm, correction, and the later public recheck? | Publication incident chain tied to the exact assistant-assisted artifact | Enterprise audit availability does not grant editorial or reader acceptance authority. |
| I build AI systems | Which identity, data route, event, refusal, escalation, and version coordinates are observable end to end? | Instrumented nine-receipt contract with fail-closed unknowns | Exposed controls are not evidence that a customer enabled, reviewed, or acted on them. |
| I teach or evaluate access | Did representative people use the delivered interaction, report harms or exclusions, and return after a change? | Representative-use and recheck overlay on the governance ledger | Policy review and telemetry cannot substitute for accessibility, learning, or transfer evidence. |
| I sponsor or edit | Is there one complete governed deployment in public custody? | Direct answer: no; fund or request the first nine-receipt case | The 0/7 is bounded to named public surfaces and does not imply private records are absent. |

Every layer should link back to the same nine receipts and retain the evidence
cut. This is a designed route for identified audiences, not evidence that those
audiences have yet used or accepted it.

### One capability-to-practice result, seven audience questions

The [complete practitioner account](/reports/practitioner-and-reader-experience/#three-controlled-bridges-connect-capability-to-human-consequence)
has one canonical answer: **three of ten audited primary cases form controlled
partial bridges from technical evidence to human use, correction, or reader
harm; zero reaches exact-version accepted delivery or a complete episode.**
HAIChart is the broader reinforcement-learning recommender case, interactive
task decomposition supplies the correction case, and ChartAttack supplies the
reader-harm case. The translation changes the next decision, not those counts:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Did people use or correct the same system and artifacts that produced the capability result—and did any output become accepted work? | Three-rung bridge card beside the workflow evaluator | Controlled use and correction are stronger than a benchmark alone; neither is field delivery. |
| I run analytics | Did one frozen configuration clear our semantic and acceptance contract, enter the governed surface, and survive real use? | Capability → acceptance → delivery gate beside the pilot scorecard | A controlled comparison cannot authorize rollout without the organization's own definitions, users, and incident path. |
| I research visualization | Which artifact, task population, configuration, version, technical receipt, and human receipt remain joined? | Ten-case machine-readable ledger plus named search surfaces | Three partial bridges and two zero counts are bounded results, not world-level prevalence or absence. |
| I publish data journalism | Does the same published graphic retain source, editorial, responsive, accessibility, reader, correction, and cost receipts? | Publication bridge card that begins with ChartAttack's measured harm | A controlled chart-QA effect establishes possible reader harm, not its prevalence or a consequential newsroom outcome. |
| I build AI systems | Which exact mechanism improved a technical receipt, what human behavior changed, and where did the chain break? | Versioned trace from system configuration through acceptance and delivery | HAIChart is a reinforcement-learning recommender, not evidence for a current LLM authoring stack; do not transfer its result by category label. |
| I teach or evaluate access | Did representative people use the exact delivered interaction, and were comprehension, correction, transfer, device, and assistive paths measured separately? | Representative-use extension to the bridge card | A participant or reader study is not automatically a learning, accessibility, or release result. |
| I sponsor or edit | How much of the capability-to-practice chain is observed? | Direct answer: 3/10 controlled partial; 0/10 accepted-delivery; 0/10 complete | Fund the first exact accepted-delivery continuation, not another disconnected benchmark or testimonial. |

Every layer should link back to the comprehensive answer and retain its reopen
condition: carry the same artifacts, tasks, configuration, and version through
a frozen acceptance rule, real delivery, representative use, a later event,
and whole cost. This is translation for intended audiences, not evidence that
an audience has used or accepted the public layer.

### One reader-outcome result, seven audience questions

The canonical answer is: **one of ten held rows directly carries selected
AI-generated charts into an independent controlled-reader effect; five rows
measure adjacent human outcomes under different AI roles; zero reaches exact
accepted delivery and later same-lineage reader recheck.** The translation must
name the AI role before naming the outcome:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Did someone other than me understand the exact artifact I accepted? | Eight-receipt artifact-to-reader card | Creator confidence and model QA are not reader acceptance. |
| I run analytics | Did the deployed display improve a real authorized decision, and did that result survive a change? | Same-release decision and recheck register | A simulated chatbot-versus-dashboard study cannot authorize production. |
| I research visualization | Which AI role caused which human outcome on which artifact? | Factorized create/select/label/explain/interact map plus lineage | Adjacent reader-assistant effects cannot fill generated-artifact cells. |
| I publish data journalism | Did intended readers understand the published claim without added harm? | Source-to-publication-to-reader incident chain | ChartAttack establishes possible controlled harm, not newsroom prevalence. |
| I build AI systems | Can the result be reproduced on an immutable release after a named change? | Versioned exposure manifest plus return-reader cohort | Maintenance without return readers does not preserve a human result. |
| I teach or evaluate access | Did assistance improve access, comprehension, and later unassisted use on intended devices? | Separate modality, performance, preference, learning, transfer, and recheck lanes | Preference, spatial understanding, and accuracy can move differently. |
| I sponsor or edit | Is delivered audience benefit established? | Direct answer: 1 direct controlled effect; 0 delivered-recheck chains | Say “controlled effect” and “adjacent outcome,” not “proven audience benefit.” |

These are audience-facing decisions layered over one verified evidence ledger,
not seven new interpretations of the underlying result.

### One development-decision and recheck result, seven audience questions

The canonical answer is: **Lexara adds one two-week longitudinal development-
decision near-miss; adding the later same-feature InfoViP operational line
brings the audit to 15 named reader, production, and evaluation cases, while
zero reaches the complete exact-artifact decision/recheck chain.** Six CVA
developers ran 38 experiments over 57 uniquely authored cases, ten models, and
six prompts using their own data and prompts. The study records real model and
prompt selections, rationale, disagreements, and confidence. It does not bind
an immutable study build and exact configuration to accepted downstream
delivery, a consequential audience decision or calibrated trust, whole cost,
and later post-release same-lineage recheck:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Can I justify the model and prompt behind the artifact I am about to accept? | Evaluation-to-acceptance card with cases, diffs, overrides, selected configuration, and artifact hash | A model comparison does not show that readers understood the released artifact. |
| I run analytics | Did the selected configuration improve an authorized decision after deployment? | Release ledger joining evaluation export, semantic acceptance, user denominator, decision errors, incidents, and recheck | Development selection and reported confidence are not operational outcome or calibrated trust. |
| I research visualization | Which lifecycle receipt does a longitudinal toolkit study actually add? | Twelve-receipt matrix separating development decision, acceptance, delivery, audience consequence, and return | Repeated experiments during one diary study are not a longitudinal outcome study of one release. |
| I publish data journalism | Can an evaluation finding become a reproducible publication gate? | Editorial packet pinning source, model/prompt, expected claims, visual diff, override, published route, reader test, and correction | Backend eligibility metrics do not establish public comprehension or editorial acceptance. |
| I build AI systems | Can the selected configuration and rationale survive model or prompt drift? | Versioned evaluation export tied to release, telemetry, incident, rollback, affected-user check, and owner | A currently responding application shell is not the participant-tested build or a maintenance contract. |
| I teach or evaluate access | Did the system improve learning or access for intended people over time? | Separate assisted performance, modality, preference, transfer, confidence calibration, and later unassisted recheck | Developer confidence and metric-human agreement do not establish learner or accessibility outcomes. |
| I sponsor or edit | Has longitudinal evaluation proved audience benefit? | Direct answer: one stronger two-week development-decision study; zero complete episodes among 15 named cases | Say “development-decision evidence,” not “proven deployment impact.” |

Every layer leads back to the same twelve receipts. The intended route changes
the decision product, never the zero or the distinction between development
selection, contract acceptance, audience consequence, confidence calibration,
and post-release return.

### One artifact-custody result, seven audience questions

The canonical answer is: **a bounded title-cue screen over A208's 122
authoritative stable-key studies surfaces one explicit GPT signal, one pin-able
supplemental-output bundle, and zero audience-delivery or later-recheck
chains.** [Beyond Generating Code](https://arxiv.org/abs/2306.02914) evaluates
91 quiz questions and nine homework assignments. Its [one-commit public
supplement](https://github.com/GPT4VIS/GPT-4-CS171/tree/6e1367a2f38c5c29af9376f7b6012f1fec1112f5)
preserves prompts, results, generated code, and grader-ready bundles across
1,271 blobs. It does not preserve an immutable provider model snapshot or
complete configuration, expose a versioned release or later commit, directly
evaluate the final project, or deliver artifacts to the government, UN, or
primary-school audiences named in prompts:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Can I reproduce and justify the output I am about to accept? | Pinned prompt/output/code bundle plus model/configuration manifest, artifact hash, and acceptance record | A large repository bundle is better custody, not evidence that an intended reader received or understood the artifact. |
| I run analytics | Did this exact accepted output enter the governed decision surface and survive a change? | Release-linked ledger joining output bundle, semantic acceptance, authorized delivery, incidents, outcome, and recheck | Model aliases and homework grades do not identify an immutable exposure or production effect. |
| I research visualization | What does the first stable-frame GenAI signal add to the lifecycle? | Artifact-custody row separating prompt, output, code, grading, release, delivery, audience outcome, and return | The other 121 title nonmatches were not full-text exclusions; five review positions remain unresolved and the denominator stays null. |
| I publish data journalism | Did the named public audience actually receive and evaluate the published claim? | Source-to-prompt-to-output-to-publication-to-reader packet with editorial acceptance and correction path | Naming a government or UN audience inside a prompt is audience-conditioned generation, not publication or reader evidence. |
| I build AI systems | Can I recreate the exact provider behavior and compare it after a release or model change? | Immutable model/configuration receipt, tagged release, retained attempts, later-event diff, and affected-user recheck | A `gpt-4` alias, one root commit, and retry description do not freeze provider state or create a longitudinal comparison. |
| I teach or evaluate access | Did learners or intended access users receive the artifact, learn from it, and retain or transfer the result? | Assignment/final-project exposure record plus representative outcome, device/access path, and later transfer check | Quiz and homework performance, grader identification, and audience-directed prompts are not learner delivery, accessibility, or transfer evidence. |
| I sponsor or edit | What is the shortest defensible result? | Direct answer: 122 titles triaged; one explicit GPT signal; one pinned bundle; zero delivery/recheck chains | Say “artifact custody,” not “audience-tested”; do not turn 121 nonmatches into negatives. |

Every route begins from the same comprehensive answer and changes only the
decision layer. The first valid upgrade is one frozen artifact and
configuration carried through acceptance, intended-audience delivery, measured
outcome, a material later event, and a same-lineage recheck. Designed
translations are still not evidence that any audience has used them.

### One mature-platform result, seven audience questions

The canonical answer is: **exact-DOI discovery covers 114 authoritative A208
keys and 87 available abstracts; it adds no hidden GenAI candidate, but it
surfaces one mature non-AI visualization afterlife and zero complete AI-
assisted lifecycles.** [AIDSVu](https://pmc.ncbi.nlm.nih.gov/articles/PMC7654504/)
reports ten years of public delivery, 501,527 unique users in 2019, named
planning applications, recurring governance, and explicit maintenance limits.
An official [2026 data release](https://aidsvu.org/news-updates/aidsvu-releases-2025-prep-data/)
and current [FAQ](https://aidsvu.org/faqs/) add a later event, annual-update
contract, and careful unavailable/unreleased/suppressed/lagged/estimated/
corrected data states. They do not add an AI visualization role, immutable
build, version-bound measured audience outcome, or affected-audience recheck:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | What must a public visualization preserve to remain useful for years? | AIDSVu worked case joining data authority, audience-specific surfaces, privacy states, governance, recurring updates, and named later release | Broad reach and durable maintenance do not show that one version caused a better reader decision. |
| I run analytics | Does aggregate platform use justify enabling an AI feature? | Delivery-to-outcome ledger separating authorized data, release, active use, decision record, incident, affected-user return, and whole cost | IP-derived organization categories and named uses are proxies, not role-level acceptance or measured outcome. |
| I research visualization | What did abstract enrichment add to the lifecycle denominator? | Reusable `122 stable / 114 DOI / 87 abstracts / 27 no abstract / 8 non-DOI` coverage receipt plus B124 primary audit | The 86 GenAI cue nonmatches are not exclusions, and a non-AI afterlife cannot complete an AI case. |
| I publish data journalism | Which durable public-data practices transfer into publication? | Worked case separating unavailable, unreleased, suppressed, lagged, estimated, and later-corrected values | Transfer the custody method; do not claim its readers understood every release or that an AI system produced it. |
| I build AI systems | Can mature platform delivery and governance validate a new model layer? | Side-by-side receipt card: platform afterlife versus model exposure, accepted artifact, measured outcome, incident, and return | Never borrow AIDSVu's reach, governance, or later data release into an untracked AI feature. |
| I teach or evaluate access | Did representative people understand and use the maintained surface after change? | Exact-release task, modality, device, comprehension, access, and later-return protocol | Users, downloads, engagement, and planning examples do not establish comprehension, learning, accessible use, or transfer. |
| I sponsor or edit | What is the shortest defensible finding? | Direct answer: one mature non-AI visualization afterlife; zero complete AI lifecycles in the bounded abstract screen | Do not say zero of 122/127 or call a non-AI platform an AI success case. |

Every route keeps the mature platform receipts because they are genuinely
useful. Every route also keeps the missing AI and audience-outcome joins. The
translation changes the next decision product, not the evidence class.

### One primary-evidence result, seven audience questions

The canonical answer is: **five exact full texts recovered below the 27
no-abstract DOI rows add four empirical human-evidence cases, but no accepted
deployment.** [B17](https://www.osti.gov/servlets/purl/1108507) and
[B128](https://ir.cwi.nl/pub/24618/24618.pdf) measure controlled tasks;
[B120](https://www.cs.umd.edu/~ben/papers/Sopan2012Community.pdf) records HHS
co-design, a demo, and bounded use; [B90](https://doi.org/10.5281/zenodo.6324996)
records an enterprise demonstrator and production intent. None joins accepted
field release to later affected-audience recheck. All five explicit GenAI cue
screens are empty. That pass left 22 DOI rows unassessed; the subsequent high-
signal recovery adds eight abstract or official-project surfaces and leaves
14 DOI rows without substantive primary content:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Which human evidence should travel with this artifact? | Three-lane card separating controlled task, co-design/demo, and production intent | Keep every receipt attached to its artifact and stage; do not call the ladder deployment. |
| I run analytics | Is a tested prototype ready for operational approval? | Gate joining exact accepted build, named operators, routine use, decision outcome, later event, and affected-user return | Positive usability, a partner test bed, and production intent are not operational acceptance. |
| I research visualization | What did primary-text recovery add below the missing-abstract layer? | Reusable `27 no abstract / 5 full texts / 8 further content surfaces / 14 unassessed` receipt | Five full-text and eight abstract-level cue nonmatches do not classify the remaining 14 DOI rows or the world. |
| I publish data journalism | Which verb does the evidence support? | Literal labels: controlled result, co-designed prototype, bounded demo, intended production, accepted release | Never compress these states into “used in practice.” |
| I build AI systems | What must follow prototype evaluation? | Bind exact artifact/configuration to acceptance, delivery telemetry, outcome, later material change, and return | A roadmap or implementation partner cannot substitute for a production receipt. |
| I teach or evaluate access | Did representative people understand and use it? | Representative task, comprehension, effort, access, and transfer measures by audience and device | A demo audience or convenience sample does not establish learning, access, or transfer. |
| I sponsor or edit | What is the shortest defensible finding? | Direct answer: four empirical human-evidence cases; zero accepted deployments in this bounded screen | Do not say zero of 27, 122, or 127 and do not combine the four artifacts. |

The translation changes the next product—study card, approval gate, methods
register, publication language, release ledger, access protocol, or executive
answer—while preserving one sentence: **co-design, a demo, a human task study,
and production intent are four different receipts—not deployment.**

### One field-use result, seven audience questions

The canonical answer is: **the eight stable non-DOI keys yield three DOI
repairs, one non-AI embedded-use comparator, and zero complete AI lifecycles.**
The [cancer diary](https://pubmed.ncbi.nlm.nih.gov/22874273/) reports increasing
system-log use; an [independent systematic
review](https://pmc.ncbi.nlm.nih.gov/articles/PMC11705737/) recovers an
11-clinician QUIS median of 4.38. Four full texts, two primary abstracts, one
issue excerpt, and one gated primary text remain distinct:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | What is the strongest practice receipt here? | Cancer-diary card separating embedded use logs, 11-clinician usability, and the missing build/outcome/recheck fields | Operational use is stronger than a demo, but remains non-AI and incomplete. |
| I run analytics | Which evidence can justify deployment approval? | Ladder from theory and prototype through simulation, technical performance, embedded use, acceptance, outcome, and recheck | Increasing use, satisfaction, and an industry demonstrator do not establish a governed accepted release. |
| I research visualization | What did recovery below the DOI layer change? | `8 keys / 3 DOI repairs / 2 year corrections / 1 foreign PMID / 4 full texts / 1 gated` authority-and-custody receipt | Seven cue nonmatches do not classify the gated row, 122 keys, or 127 positions. |
| I publish data journalism | Which evidence label should survive publication? | Literal labels: theoretical model, prototype demo, bounded simulation, technical experiment, embedded use, accepted release | Never compress the ladder into “used in practice” or “AI-powered.” |
| I build AI systems | What follows positive usability? | Bind exact build and configuration to acceptance, telemetry, consequential outcome, later material change, and affected-user return | Median satisfaction from 11 clinicians does not supply version lineage or revalidation. |
| I teach or evaluate access | Whose human evidence is present? | One clinician survey and one observed emergency operator, with task, access, patient, transfer, and later-return measures still open | Technical data and conceptual audience labels are not representative human evidence. |
| I sponsor or edit | What is the shortest defensible finding? | Direct answer: a field-use receipt is still not an AI lifecycle | One primary text remains gated and the complete-review denominator is unknown. |

Every route begins from the same repaired authority ledger and preserves the
same missing receipts. The useful translation changes; the evidence does not.

### One prototype-evaluation result, seven audience questions

The canonical answer is: **nine high-signal DOI rows yield eight substantive
primary abstract or official-project surfaces, seven explicit human
evaluations, one production-intent near-miss, and zero complete lifecycles.**
[InfoViP](https://pure.johnshopkins.edu/en/publications/information-visualization-platform-for-postmarket-surveillance-de/)
is the nearest delivery case: seven FDA safety evaluators supplied requirements
and evaluated the prototype, suggestions were addressed, and an [official FDA
page](https://www.fda.gov/science-research/advancing-regulatory-science/improving-efficiency-and-rigor-pharmacovigilance-fda-visualization-multi-source-information-and)
says an enhanced NLP and unsupervised-learning version will be installed in
production. Installation, acceptance, routine use, regulatory outcome, later
change, and evaluator return are not confirmed. Fourteen DOI rows still lack
substantive primary content:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | What does positive human evaluation let me claim? | Receipt card separating participant, task, measured response, artifact state, and missing delivery gates | Usability, willingness, and positive comments support iteration, not release or routine use. |
| I run analytics | Is InfoViP ready to count as operational evidence? | Seven evaluators, requirements, feedback, addressed suggestions, and a separately labeled planned installation | “Will be installed” is not installed, accepted, used, or outcome-bearing. |
| I research visualization | What did this recovery pass change? | `22 prior gap / 9 investigated / 8 substantive surfaces / 7 explicit evaluations / 14 remaining` coverage receipt | The eight surfaces are not full papers and their cue nonmatches are not full-text exclusions. |
| I publish data journalism | Which verb survives publication? | Controlled study, prototype evaluation, formative pilot, production intent, accepted release, routine use, recheck | Do not translate the first four into “deployed” or “used in practice.” |
| I build AI systems | What follows prototype evaluation? | Bind exact build/configuration to authorization, telemetry, outcome, later event, and affected-user return | Addressed feedback is design progress, not versioned production acceptance. |
| I teach or evaluate access | Whose evidence is represented? | Practitioners, students, cross-country users, and real fantasy-sport users, with counts and tasks when exposed | No row establishes representative accessibility, own-device transfer, or post-release learning. |
| I sponsor or edit | What is the shortest defensible finding? | Direct answer: seven human-evaluation signals are still zero complete production lifecycles | The sole AI-adjacent case is non-GenAI, and 14 DOI rows remain unassessed. |

Every route preserves the same grammatical and lifecycle boundary. The
translation changes the decision product; it does not turn `will be installed`
into `was deployed`.

### One residual-recovery result, seven audience questions

The historical answer at that evidence cut was: **all 14 residual DOI rows
were investigated; eleven gained substantive primary content in that pass,
including one full chapter, and three were then metadata-only. The strongest
additions were evaluation-to-redesign and practice context, not AI
deployment.** The [full
chapter](https://link.springer.com/chapter/10.1007/978-3-319-58521-5_20)
connects three UX experts and 25 problems to a third design version. The
[corridor case](https://iris.polito.it/handle/11583/2541289) and [university-
network case](https://link.springer.com/chapter/10.1007/978-3-030-58802-1_4)
involve stakeholders and managers but expose no accepted build, ordinary-use
denominator, consequential outcome, or later return:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | What mechanism can I use now? | Receipt joining evaluated version, evaluator, problem inventory, and resulting version | Expert review supports redesign; it does not prove operator acceptance. |
| I run analytics | Which cases are closest to practice? | Corridor-stakeholder and university-network cases followed into accepted-build, use, and outcome gates | Organizational context and multi-unit evaluation are not deployment denominators. |
| I research visualization | What changed in the evidence frame? | `14 residual / 11 substantive / 1 full chapter / 3 metadata-only / 0 complete AI lifecycles` | Eleven bounded non-GenAI screens are not a whole-review zero. |
| I publish data journalism | Which wording survives publication? | Expert-evaluated, applied with stakeholders, evaluated across units, positive expert feedback | Do not compress those stages into deployed, adopted, or proven impact. |
| I build AI systems | What follows evaluation-to-redesign? | Freeze the changed build, obtain operator acceptance, measure ordinary use and outcome, then recheck after change | A third version is a design receipt, not a production receipt. |
| I teach or evaluate access | Whose understanding is represented? | Separate experts, managers, stakeholders, and non-experts; preserve counts, tasks, access conditions, and transfer | Experts cannot stand in for representative affected audiences or accessibility coverage. |
| I sponsor or edit | What is the shortest defensible finding? | Historical answer: the content gap was three at that pass; no new AI deployment lifecycle was found | Zero complete lifecycles in this batch is not zero of the review. |

Every layer derives from the same comprehensive answer. The translation
changes vocabulary, depth, and decision utility—not the evidence state.

### One public-delivery afterlife result, seven audience questions

The historical answer at that pass was: **B110 had gained a substantive official abstract,
immutable author-linked framework and application source, and a same-named
DiscoverWater interface reachable at KU in August 2026. B92 and B91 then
remained content-unassessed. This was a bounded non-AI public-delivery afterlife, not an
exact-build use or outcome lifecycle.** The [official
abstract](https://agris.fao.org/search/en/providers/122535/records/65df73744c5aef494fe2a5fc)
names stakeholders, policy makers, scientists, educators, resource managers,
and users. The [pinned
framework](https://github.com/mistyblue17/DiscoverFramework_v1.2/tree/10a735526e72b6563749e07ddea762ff457e3061),
[pinned DiscoverWater application](https://github.com/mistyblue17/DiscoverWater/tree/51dc683291e8e730c6b497272f1ff78c0c488da9),
and [live interface](https://interactiveviz.ku.edu/DiscoverWater/) preserve
different receipt stages:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | What makes a public visualization inspectable? | Paper claim, source SHA, named live URL, and dated interaction check as separate receipts | Source availability and reachability do not prove audience usefulness. |
| I run analytics | What must a pilot or procurement decision add? | Live-build digest, owner acceptance, ordinary-use denominator, decision outcome, maintenance owner, and cost ledger | A public demo is not an accepted operational deployment. |
| I research visualization | What changed in the evidence frame? | Historical receipt: `3 residual / 1 newly substantive / 1 live surface / 2 unassessed / 0 complete AI lifecycles` | At that pass only B110 was screened; B92 and B91 were not negative rows. |
| I publish data journalism | Which wording survives publication? | “The named DiscoverWater interface was reachable in August 2026.” | Do not write “used for five years,” “maintained,” “adopted,” or “effective.” |
| I build AI systems | What provenance should the next release retain? | Bind deployed bytes to a commit, acceptance receipt, route checks, telemetry, feedback, changes, and audience return | A pinned source repository is not automatically the deployed build. |
| I teach or evaluate access | Whose experience is represented? | Treat each intended group as a separate evaluation population with tasks, access paths, and acceptance criteria | Named intended audiences do not constitute a representative or accessible evaluation sample. |
| I sponsor or edit | What is the shortest defensible finding? | At that pass, two records remained unassessed; the new evidence was public non-AI delivery, not proven use or impact | Current reachability is not deployment efficacy or an AI success story. |

Every route starts from one maintained evidence state. The live check earns the
word `reachable`; it does not earn `continuously available`, `used`,
`accepted`, `effective`, or `rechecked`.

### One source-lineage answer, seven audience decisions

The canonical answer is: **the dated live DiscoverWater page is not reproduced
by either held public framework state, while selected data assets establish
partial shared lineage. One of three sampled pairs matches bytes, two match
canonically, and one differs materially. No exact build or audience outcome is
held.** The source audit keeps a useful partial answer without promoting a
repository link, prototype demonstration, analytics hook, or intended audience
into deployment or use:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Can I reproduce what readers see? | Dated live digest, published page reconstruction, three asset pairs, and the missing manifest | A repository link and matching data do not reproduce the current page. |
| I run analytics | Is this governed, accepted, used, and maintained? | Release checklist: page digest, manifest, owner acceptance, task denominator, outcome, support, and cost | Public reachability is a delivery state, not operational use. |
| I research visualization | Which artifact did the paper and audience encounter? | Exact and canonical comparisons plus explicit reopen conditions | Partial lineage and intended audiences cannot become a version-bound evaluation. |
| I publish data journalism | Which wording survives publication? | “Reachable in August 2026; selected source data match; the published page source does not reproduce the live page.” | Do not write “current open-source build,” “used for five years,” or “proved awareness.” |
| I build AI systems | What closes the provenance gap? | Manifest joining source tree, dependencies, configuration, route, release time, and owner; telemetry and return stay separate | A source SHA is not a deployed SHA, and analytics code is not telemetry evidence. |
| I teach or evaluate access | Can intended learners and access users complete their tasks? | Exact-build testing on representative devices and assistive paths, with comprehension, correction, and later return | A tutorial, responsive aspiration, and named educators do not establish accessible learning. |
| I sponsor or edit | What is the shortest defensible result? | One live non-AI interface, partial public-source lineage, zero exact-build or audience-outcome receipts | Fund a manifest plus version-bound audience return, not another same-name source link. |

These are translations of one comprehensive answer. They change vocabulary and
decision utility, never the evidence state.

### One publisher-abstract recovery and one stop rule, seven audience decisions

The current canonical answer is: **B92 now has an exact publisher abstract
describing a multicriteria hydrogen-pipeline risk model, Monte Carlo
simulation, Kendall's tau rank comparison, and graphs for ranking sections and
targeting mitigation. It names no human sample or AI visualization role and
supplies no release, delivery, use, outcome, recheck, or whole cost. B91
remains content-unassessed, while generic retrieval over six named surfaces is
paused after three bounded lawful passes.** The 27-row ledger stays `6 full /
20 primary abstract or official / 1 unassessed / 0 complete AI lifecycles`;
the separate access state is `1 access-blocked and unassessed / 0 active
generic B91 targets`. These are
abstract-bounded findings because the full B92 paper is not held:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | What mechanism can I reuse now? | Propagate uncertainty, compare ranking stability, and show the ordered decision rather than only a score | The abstract establishes a method, not a tested implementation or audience result. |
| I run analytics | What would make this operational evidence? | Acceptance rule, intended user and task, version-bound delivery, observed-use denominator, decision outcome, support owner, and cost | A risk-ranking method is not evidence of adoption, impact, or governed operation. |
| I research visualization | What changed in the evidence frame? | `27 investigated / 6 full / 20 abstract or official / 1 access-blocked and unassessed / 0 active generic B91 targets / 0 complete AI lifecycles` | B91 remains unknown. Reopen only on exact new lawful custody; no substantive row supplies a complete AI lifecycle. |
| I publish data journalism | Which wording survives publication? | “The publisher abstract describes a non-AI uncertainty-and-ranking decision-support method.” | Do not write “users adopted,” “AI-assisted,” “effective,” or “deployed.” |
| I build AI systems | What evidence should an implementation retain? | Immutable code and configuration, uncertainty representation, rank-sensitivity tests, acceptance, delivery, telemetry, correction history, and return | An inspectable method description is not an inspectable or accepted build. |
| I teach or evaluate access | What distinction should learners practice? | Separate method, human evaluator, intended user, delivered artifact, access path, outcome, and later return | A graph intended to help a decision does not show who understood or could access it. |
| I sponsor or edit | What is the shortest defensible finding? | One row is assessed at the publisher-abstract layer; the final gap is access-blocked and unassessed, not negative; zero complete AI lifecycles are held | Fund access only if an exact deposit can be obtained; otherwise fund one same-artifact AI delivery-to-recheck episode. |

Every route derives from the same comprehensive answer. Translation changes
depth, vocabulary, and next action; it does not add participants, delivery, or
outcomes that the source does not report.

### One field-use, cost, and governance answer, seven audience decisions

The canonical answer is: **A deduplication component of FDA's InfoViP was used
by 20 unique reviewers in a six-month internal real-work evaluation and is
later reported operating across 29 million historic plus daily incoming FAERS
reports; three related FDA awards total $2.40 million obligated and $2.86
million estimated, while the current inventory still records implementation
`N/A` and no associated AI-system ATO.** Admit one counted field-evaluation and
operating-pipeline near-miss with a bounded award envelope; do not convert it
into accepted whole-platform deployment, current routine use, impact, whole
cost, ROI, or a complete episode:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Which work state can I emulate? | Requirements, internal deployment, real-work submissions, reviewer-confirmed labels, error analysis, operating processing, release, and change history | Twenty evaluation users do not identify a current accepted interface or routine-use denominator. |
| I run analytics | What can the operating and performance numbers justify? | Preserve human confirmation; monitor the 0.71 recall / 0.67 precision field means and later 0.36–0.93 F1 spread by dataset | Bounded set performance and pipeline operation do not establish universal gains or changed safety decisions. |
| I research visualization | Which lifecycle bridge is genuinely new? | One same-feature chain from co-design through a counted field evaluation to operating processing and an explicit award envelope | Voluntary evaluation, post-release ordinary use, authorization, outcome, and later return remain separate variables. |
| I publish data journalism | Which wording survives publication? | “Twenty reviewers used a deduplication component in a six-month internal evaluation; it is later reported operating across historic and daily data; three awards total $2.40M obligated / $2.86M estimated.” | Also report implementation N/A and no ATO; do not write “fully authorized,” “proved impact,” or “whole cost.” |
| I build AI systems | Which engineering and governance receipts must stay joined? | Evaluated and operating code, model/data snapshot, visualization build, acceptance, authorization, cohort, overrides/errors, operations owner, change, return, and cost | At-scale operation and human confirmation do not supply an immutable accepted build or maintenance transfer. |
| I teach or evaluate access | Which evidence states must learners distinguish? | Co-design, voluntary field evaluation, internal deployment, pipeline operation, routine use, authorization, access, outcome, and later return | Each denominator belongs to a different state; no held record measures representative access or accessibility outcomes. |
| I sponsor or edit | What is the shortest defensible result? | 20 field-evaluation users; 29M historic plus daily operation; $2.40M obligated / $2.86M estimated; 15 named cases; zero complete episodes | The three-award envelope omits internal labor, infrastructure, grants, review, operations, maintenance, and failures; it is not ROI. |

These are seven decision layers over one record. None may erase the field-use
denominator, performance variation, award-envelope boundary, or current
governance counterevidence.

### One production-component and QA answer, seven audience decisions

The canonical answer is: **A final 2025 CIOMS use-case report says FDA's
InfoViP deduplication component was approved for historical and live
processing, installed in AWS, integrated with AERS, and processing more than
30 million historical plus about 8,000 daily submissions; it also says a solid
QA plan was not yet in place, routine-use roles were incomplete, and downstream
signal effects had not been investigated, while the current federal inventory
still lists implementation `N/A` and no AI-system ATO.** A 2026 FDA fellowship
record proposes Elsa GenAI API and UI work; it is a change trigger, not an
observed release.

One date-and-relationship repair now travels through every route. An [FDA
page](https://www.fda.gov/drugs/cder-conversations/understanding-cders-postmarket-safety-surveillance-programs-and-public-data)
that appears 2026-dated in search explicitly says its content is current to 3
April 2024, so its future-tense InfoViP language predates the final report. A
[May 2026 biography](https://www.fda.gov/media/192553/download) names Oanh Dang
as serving project lead without assigning operations or maintenance authority.
FDA's [Elsa 4.0/HALO
release](https://www.fda.gov/news-events/press-announcements/fda-expands-ai-capabilities-and-completes-data-platform-consolidation)
does not name InfoViP, so the application-specific integration remains
unobserved. For BI and sponsors, the resulting decision rule is: require the
unit, event date, exact role, and named application join before accepting a
broader platform or GenAI status claim.

The accountability translation is equally literal. The live [InfoViP–Elsa
opportunity](https://www.zintellect.com/Opportunity/Details/FDA-CDER-2026-0089)
is now closed and names Joshua Xu, Leihong Wu, and Oanh Dang as research
mentors. Their official profiles support R2R/advanced-AI integration,
bioinformatics research, and InfoViP project leadership; none names an InfoViP
operator, maintainer, QA executor, release approver, or authorization
authority. A fellow would be a nonemployee barred from inherently governmental
functions. The useful BI layer is therefore **three research contacts, zero
named operating authorities**—not “the closed fellowship shipped.”

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Where is human control real? | Experts can confirm or change reference cases in bounded case-series work; design that inspectable path explicitly. | The full daily stream cannot be human-confirmed, so do not label all processing “human in the loop.” |
| I run analytics | Who can answer the research question, and who can approve the change? | Route research-scope questions to Xu, Wu, and Dang; keep operator, maintainer, QA executor, release approver, and authorization authority as five separate fields. | Three mentors and a closed listing do not assign any of the five operating-authority roles. |
| I research visualization | How should this episode be coded? | One same-project self-reported production component and bounded routine case-series path among 15 named cases. | The current inventory is a conflicting broader record; version, authorization, outcome, return, and cost gaps keep complete episodes at zero. |
| I publish data journalism | Which sentence survives publication? | Report approval, installation, integration, >30M history, ~8K daily, and the QA/role/outcome limits together. | Do not call the whole platform independently audited, fully authorized, or proven to improve signal discovery. |
| I build AI systems | What release trace is still missing? | Join the named mentors and implementer to immutable build/configuration, acceptance, ATO, QA result, operating owner, exposure, override/error, change, and affected-user return. | A fellow would be a nonemployee unable to perform inherently governmental functions; implementation work cannot approve itself. |
| I teach or evaluate access | What distinction should learners practice? | Separate automated all-stream processing, inspectable case-series review, monitoring, planned audit, and observed user outcome. | No held receipt covers assistive paths, comprehension, representative access, or post-change return. |
| I sponsor or edit | What is the shortest defensible status and next gate? | Approved component; three research mentors; zero named operating authorities; require version/sign-off/ATO, completed QA audit, current use, outcome, and return next. | The Elsa/API/UI opportunity is closed, not completed or shipped. |

Every layer points back to the same full account. None may turn a same-project
production statement into independent audit, the current inventory conflict
into a resolved authorization, or future-tense fellowship aims into deployment.

### One review-lineage result, seven audience questions

The canonical answer is: **a 2025 [systematic
review](https://www.frontiersin.org/journals/communication/articles/10.3389/fcomm.2025.1605655/full)
maps 127 studies of data visualization in broadly defined AI-assisted
decision-making, but it does not construct the denominator for complete
same-artifact GenAI lifecycles.** It
retains 118 empirical papers and nine reviews; it explicitly does not focus on
GenAI, leaves GenAI trust and interpretability for future investigation, and
calls longitudinal decision-behavior evidence future work. The qualifying
count is unknown, not zero of 127. Its [publisher
supplement](https://public-pages-files-2025.frontiersin.org/articles/1605655/file/Supplementary_file_1.pdf/1605655_supplementary-file_1/2)
also separates reported count from auditable identity: six domain totals sum to
127, while the “full list” names 126 citations and leaves one Education entry
as a literal `?`. A source-internal repair now identifies that row as
[Hernández-Calderón et al. (2023)](https://doi.org/10.1093/iwc/iwac043): A208
says the domains were assigned after selection, and this is the sole one of its
five Education examples absent from the twelve named entries. The publisher
file remains unchanged. The complete six-domain crosswalk now maps 122 of the
127 supplement positions to distinct A208 review keys. Five supplement-only
labels have exact-looking external DOI candidates—[Jiao
(2022)](https://doi.org/10.1016/j.compeleceng.2022.107737), [Burt et al.
(2017)](https://doi.org/10.1186/s13643-017-0544-1), [Theis et al.
(2018)](https://doi.org/10.2196/medinform.9394), [M. Lu
(2020)](https://doi.org/10.1145/3419635.3419733), and [Islam et al.
(2022)](https://doi.org/10.1007/s10479-021-04465-7)—but no review-controlled
artifact joins them. One admitted key also carries a printed PMID that resolves
to another paper. The frame is 122 authoritative keys, five authority gaps,
and one quarantined identifier conflict. Lifecycle screening remains
unstarted and the count remains null:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Can I cite the review frame without hiding its authority gaps? | `122 authoritative / 5 unresolved` receipt plus the identifier-conflict note | A complete-looking citation is not necessarily the review's confirmed row. |
| I run analytics | Which evidence-ingestion states may enter the governed denominator? | Separate source label, review key, external candidate, identifier validation, and admission authority | Never coalesce a candidate and authoritative key or reuse a conflicted identifier. |
| I research visualization | Is the review frame stable enough to screen? | Reusable 127-position register and severe-test verifier | Resolve five review-side authority joins and the conflicted PMID first. |
| I publish data journalism | Can the five likely papers be named? | Name them as registry candidates while saying A208 does not link them | Do not describe any candidate as the review's confirmed row; quarantine the unrelated PMID. |
| I build AI systems | Can automated review ingestion abstain on high-confidence matches? | Five fail-closed candidate-admission cases plus identifier-owner validation | Confidence and topical fit cannot manufacture provenance. |
| I teach or evaluate access | Why are likely identity matches still not enough? | Worked source/key/candidate/identifier/authority examples | Identity confidence, identifier validity, and admission authority are different evidence dimensions. |
| I sponsor or edit | What is the shortest defensible conclusion? | Direct answer: 122 of 127 positions are keyed; five need review-side evidence; screening has not started | Fund the authority repairs; never say “0 of 127.” |

Each layer leads back to the same review and eight-field audit. The vocabulary
and decision change; the null denominator and reopen conditions do not.

### One accepted-cost result, seven audience questions

The cost audit has one canonical answer: **zero of twelve held fragments
reports both equivalent observed per-arm route-wide cost and a frozen accepted-
artifact denominator.** The peer-reviewed [Selective TTS visual-insights
study](https://aclanthology.org/2026.findings-acl.1724/) is the strongest
fixed-budget near miss: it matches declared LLM calls and, separately, output
tokens inside one pipeline. Its released accounting omits repair and failed-
worker cost, and its final reports are proxy-judge-scored candidates rather
than contract-accepted artifacts. Keep three receipts separate: matched
declared partial budget, observed route-wide use, and cost per accepted
artifact. The translation changes the decision language, not the evidence:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Which budget was promised, what did every attempt actually use, and how many became work I could accept? | Three-receipt attempt-to-acceptance card beside the workflow evaluator | First draft, final-report label, finish signal, and benchmark score are not accepted work. |
| I run analytics | Does the pilot expose planned cap, observed all-in use, and cost per output that passed our semantic, governance, correction, and delivery contract? | Procurement denominator card beside the eleven-lane ledger | A matched call allowance or one arm's API average cannot price a governed comparison. |
| I research visualization | Are declared budgets, observed events, outcome states, acceptance rule, task base, and observation window equivalent across arms? | Released per-task traces, route manifest, and explicit missingness | Calls or completion tokens can control an experiment without constituting total cost. |
| I publish data journalism | How many attempted graphics passed source, editorial, responsive, accessibility, and reader gates, and what did the rejected work cost? | Publication ledger retaining failed, rejected, and repaired work | A generated “final report” is not a publishable graphic. |
| I build AI systems | Do provider requests, retries, repair, failed branches, local compute, evaluation, and human events reconcile through acceptance? | Instrumentation schema plus fail-closed budget and ratio verifier | Assumed stage counters and candidate production are not observed use or acceptance. |
| I teach or evaluate access | Did the output pass the learner or assistive-path contract, and were representative use and correction effort counted? | Representative-user acceptance record beside the learning/access protocol | Proxy-judge preference does not establish accessible or educational acceptance. |
| I sponsor or edit | What did work that passed the same bar actually cost? | Direct answer: 0/12 held comparisons; one matched partial budget; authorize the bounded test | A quality gain under a partial inference budget does not authorize procurement or route selection. |

Every layer should retain rejected, abandoned, quarantined, and no-output work
in the numerator; count only artifacts that passed the frozen contract; and
report eligible, attempted, candidate, accepted, delivered, and reader-
successful ratios separately. This is a protocol result, not evidence that
either route is cheaper.

### One retry-topology result, seven audience questions

The [complete specialist account](/reports/specialized-vision-models/#the-same-execution-endpoint-can-hide-different-retry-burden)
has one canonical answer: **over 888 tasks, VisCoder2-32B and GPT-4.1 both finish
at 732 execution passes after 584 versus 714 conditional revisions; the held
records do not support a cost winner.** The endpoint, attempt topology, resource
receipt, and accepted-work denominator remain separate. The translation changes
the next action, not the evidence:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | When do extra repair rounds stop rescuing enough work to justify another attempt? | Retry-survival card with attempts, new passes, and unresolved work by round | A shared final execution rate does not show effort, visual quality, or accepted work. |
| I run analytics | What did each retry actually consume, and how many outputs passed the governed acceptance bar? | Pilot register joining retry identity to tokens, compute, latency, charges, people, and accepted outcomes | The 130-revision difference does not price different models or authorize procurement. |
| I research visualization | What is the conditional rescue rate among tasks still failing at each round? | Released per-task survival trace with uncertainty, resource telemetry, and stop rule | Aggregate endpoints erase attrition; reconstructed attempts are not causal cost evidence. |
| I publish data journalism | Which repaired graphics passed source, visual, editorial, responsive, accessibility, and reader review? | Publication ledger that preserves each rejected and repaired state | Execution cannot grant editorial or publication authority. |
| I build AI systems | Are usage, time, model/runtime identity, and stop reason persisted on every debug attempt? | Event schema and fail-closed telemetry check | Printed or discarded provider usage cannot support later reconciliation. |
| I teach or evaluate access | Did later retries improve representative understanding or assistive use, and at what correction burden? | Learner/access acceptance record beside retry survival | Mechanical execution and visual-quality scores do not establish learning or access. |
| I sponsor or edit | Did one route reach the same accepted result with less total burden? | Direct answer: same 82.4% execution endpoint; 584 versus 714 revisions; cost winner unknown | Fewer generations are not lower total cost without comparable resource and acceptance receipts. |

These are intended decision layers over one custody record. They are not
evidence that any audience has used, preferred, or accepted the translation.

### One longitudinal co-design result, seven audience questions

The Graphy audit has one canonical answer: **the same three blind co-designers
completed 12 sessions across four workshops and eight months while selected
interaction changes were implemented, but no workshop-version binding, formal
efficacy result, versioned release, or representative release recheck is in
custody.** The public repository has seven commits and no tags or releases.
The translation changes the next decision, not the evidence class:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Which interaction survived repeated use, and did that behavior reach the exact version I can use? | Co-design change history beside the workflow evaluator | Repeated co-design is stronger than a walkthrough; it is not proof that a released build retained the result. |
| I run analytics | Did returning participants shape the pilot, and was the governed release formally rechecked after material changes? | Workshop-to-build-to-release gate beside the pilot scorecard | A useful co-design process does not establish operational efficacy, support, or adoption. |
| I research visualization | Which workshop, commit, model, task, measure, and interval produced each result? | Longitudinal co-design register with immutable session bindings | Reflective ratings and design synthesis are not formal task outcomes or an in-situ effect. |
| I publish data journalism | Which verbs can the evidence support: co-designed, evaluated, released, maintained, independently rechecked? | Five-verb evidence note beside the accessibility protocol | Graphy supports co-designed and openly implemented; the later lifecycle verbs remain unproven here. |
| I build AI systems | Which prompt, query, state, loader, model-default, or crash change triggers another human check? | Versioned return-user gate in the release packet | A commit records a change; it does not record representative recovery. |
| I teach or evaluate access | Can people with different blindness histories, tactile experience, devices, and settings use the exact released interaction? | Representative-use plan preserving the layered and select-confirm-ask-verify mechanisms | Three experienced co-designers do not represent every blind or low-vision learner. |
| I sponsor or edit | What is the shortest defensible result? | Direct answer: eight months of repeated co-design; zero complete release-afterlife episode | Valuable design evidence is not formal efficacy, adoption, maintenance, or whole-cost evidence. |

Every layer should point back to the comprehensive practitioner answer and keep
the same reopen condition: bind a workshop to immutable code and model state,
or formally recheck a versioned release with representative users after a
material change. This is translation for intended audiences, not evidence that
any audience has used the layer.

### One six-receipt null, seven audience questions

The exact-join audit has one canonical answer: **zero of seven leading cases in
the named surfaces through 15 August 2026 joins repeated representative use,
immutable tested build, exact exposure model, versioned same-lineage release,
later material event, and representative or actor-separated post-change
recheck.** Every receipt appears somewhere, but none stays attached to one
artifact and user lineage. The audience layer changes the next action, not the
zero:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Did intended users return to the exact changed release I can use? | Six-receipt release card beside the workflow evaluator | A paper, repository, and changelog can remain three unjoined records. |
| I run analytics | Which exact release cleared representative recheck after material behavior changed? | Pilot-to-rollout gate with a named recheck owner | Maintainer smoke tests and repository activity are not operational acceptance. |
| I research visualization | Can exposure, build, model, release, change, actor, and outcome be joined without crossing case boundaries? | Machine-readable horizontal matrix plus named search surfaces | Zero of seven is a bounded null, not a world-level absence. |
| I publish data journalism | Which literal verbs are supported: studied, versioned, released, changed, rechecked? | One sentence naming the strongest rung and first missing receipt | Do not splice favorable stages from different versions, participants, or projects. |
| I build AI systems | Which immutable exposure receipt triggers the next human regression check? | Code-and-model fingerprint in every study and release packet | A moving model family or current configurable default cannot identify historical exposure. |
| I teach or evaluate access | Did representative people on their own access paths return after the exact AI release changed? | Co-design → evaluation → build → release → return-user states | Participation in one version does not validate an unidentified or rewritten successor. |
| I sponsor or edit | How many cases clear all six receipts? | Direct answer: 0/7; fund the first exact-release return-user receipt | The null should reopen when a qualifying primary record enters custody. |

The strongest next routes are specific: bind Graphy workshops to code and
model state; identify MAIDR's participant-tested version and bring users back
after maintenance; obtain the Prism affected-operator recheck; or carry the
OpenClaw proposal through merge, corrected release, and reporter
reconciliation. These are research targets, not evidence that an audience has
adopted the public layer.

### One production-recovery result, seven audience questions

The production audit has one canonical answer: **three held AI-assisted
dashboard afterlives, two maintainer-verified restorations, zero affected-actor
recoveries, zero transferred maintenance authority, and zero complete twelve-
state rows.** Its [full public account](/reports/practitioner-and-reader-experience/#production-recovery-needs-twelve-actor-separated-states)
keeps issue closure, maintainer recheck, contribution, and authority separate.
The translation changes the audience's decision, not the counts:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Does the exact corrected artifact work for the person and route that failed? | Twelve-state afterlife card with actor, version, route, repair, delivery, recheck, and prevention | A merged patch or closed issue is not recovery. |
| I run analytics | Can the pilot survive an incident without one owner carrying every state? | Recovery-and-authority matrix across semantic, deploy, incident, release, and user-acceptance owners | Accepted contribution does not transfer operational authority. |
| I research visualization | Which actor observed and accepted each transition? | Event-level register retaining partial and missing states | Do not assemble the best cells from three cases into one effect. |
| I publish data journalism | Did the correction reach the published surface and a separate checker? | Publication incident chain: report, repair, rebuild, republish, editor check, reader-surface recheck | A corrected source that never republishes is not a corrected public graphic. |
| I build AI systems | Which state should the product instrument or block? | State machine for proposal, merge, deploy, maintainer check, independent check, and new gate | Ticket status and CI cannot collapse actor and delivery states. |
| I teach or evaluate access | Can another person use the corrected state on the intended device and access path? | Actor-switched recovery exercise with task completion and explanation | Automated checks and maintainer demos are not representative-user results. |
| I sponsor or edit | Is AI-assisted production recovery established? | Direct answer: 3 afterlives, 2 restorations, 0 independent recoveries or authority transfers | Say “maintainer recheck,” not “independent recovery.” |

The first valid upgrades are also audience-specific: a return-user receipt on
the corrected route, or a named second person exercising release or incident
authority. Another project-level AI disclosure, pull-request count, issue
close, or maintainer-only demo does not move either missing state.

### One lifecycle result, seven audience questions

The latest practitioner audit has one canonical answer: **four promising public
joins still produce zero new same-artifact lifecycle episodes.** MAIDR now
contributes a dedicated legacy AI-study surface, a `v2.10.0` version floor,
abstract-level findings for eight blind and low-vision participants, and real
later maintenance, but no tested-build receipt or representative recheck. Its
current TypeScript rewrite is a separate validation target. A public-
health copilot paper describes a 16-person human method but says the experiment
was removed and reports no human result. The translation changes what each
audience does next:

| Audience route | Question at the decision moment | Layer to provide | Boundary to retain |
| --- | --- | --- | --- |
| I make visualizations | Is this the study build, merely the first containing release, or a later rechecked build? | Artifact-version-actor checklist beside the three admitted cases | A version floor is not participant exposure or proof that the studied artifact survived. |
| I run analytics | Where are the reported human results, real decisions, incidents, and whole operating cost? | Pilot card separating method, result, release, and afterlife | A planned human study or technical metric does not establish operational adoption. |
| I research visualization | What join key connects exposure, version, release, model, later event, actor, and outcome? | Full horizontal admission register and partial-lane ledger | Repository containment narrows the build; it does not identify participant exposure. |
| I publish data journalism | Which described methods survived into reportable results? | Methods-results-release integrity note | Participant count and survey design are not outcomes when the final manuscript reports none. |
| I build AI systems | Were representative users re-run after this model, verification-flow, or architecture change? | Versioned return-user gate in the release packet | Maintenance is a trigger for human recheck, not the recheck itself; a rewrite reopens the gate. |
| I teach or evaluate access | Did representative participants test the exact AI behavior on their own assistive setup? | Pre-AI baseline plus a visibly missing AI-release recheck | Representative use of one version does not validate its rewritten successor. |
| I sponsor or edit | How many candidates cleared every key, and what did the strongest near-miss actually add? | Direct answer: zero new episodes; a study surface, version floor, and exact recheck request | Useful evidence in one lane improves the next study without completing the case. |

Every layer should point back to the comprehensive public answer, preserve the
bounded-search language, and carry the exact-version reopen condition. “Zero
admitted” does not mean “no case exists anywhere,” and a translation is not
evidence that an audience has used it.

## Priority audience map

Priority reflects fit with the current evidence, reachability through observed
routes, and likelihood of a concrete decision. It is not a market-size estimate.

| Priority | Audience | Trigger | Job to be done | Decisions this package can improve | Best first asset | Plausible routes |
| --- | --- | --- | --- | --- | --- | --- |
| 1 | Visualization and data practitioners | A new tool or agent is offered; a deadline or unfamiliar stack makes assistance attractive; generated output looks plausible but is hard to finish | Use AI to reduce work without losing analytical intent, precise control, or accountability | Which tasks to delegate; which work surface to use; what to inspect; when to stop or revert | Workflow evaluator plus practitioner guide | DVS, Nightingale, visualization newsletters and podcasts, Tableau and Power BI practitioner groups |
| 1 | BI and analytics leaders | A Copilot or agent license, pilot, governance review, or high-visibility error | Decide whether and how to adopt AI over governed organizational data | Enable, pilot, procure, restrict, or retire; semantic-model readiness; review and incident ownership | BI pilot and procurement scorecard | Tableau AI user group and Slack, Microsoft Fabric or Power BI groups, dbt community, analytics leadership events |
| 1 | Visualization, HCI, and AI researchers | Literature review, study design, benchmark selection, grant or workshop planning | Locate what is measured, what is merely demonstrated, and which evidence would change the synthesis | Reuse a benchmark; design a study; select a gap; avoid an invalid cross-benchmark claim | Benchmark crosswalk plus research-gap ledger | IEEE VIS, VISxGenAI, BELIV, AccessViz, EduVis, labs, reading groups, citation repositories |
| 1 | Data journalists, graphics editors, and newsroom developers | A newsroom policy review, an AI-assisted analysis, a graphics deadline, or a proposal to publish generated work | Use assistance while preserving source custody, reproducibility, editorial authority, and reader trust | Permit or prohibit a use; set disclosure and review; validate an analysis; choose a reader test | Newsroom protocol and one worked case | NICAR/IRE, OpenNews/Source, data-journalism communities, graphics teams, newsroom training |
| 2 | Product, engineering, and tool teams | A roadmap decision, evaluation failure, customer demand, or a plan to add an agent or skill | Choose durable mechanisms and tests rather than copying demo behavior | Context architecture; agent versus deterministic component; eval suite; recovery path; instrumentation | System anatomy plus evaluation and ablation guide | GitHub projects, Observable, engineering blogs, developer events, targeted product-team briefings |
| 2 | Educators, data-literacy, and accessibility specialists | Curriculum redesign, assignment policy, an accessibility review, or evidence of assisted performance | Preserve learning, transfer, critique, and access while using assistance deliberately | Allowed assistance; skills to teach; unassisted transfer test; delivered-reader acceptance | Learning and accessibility protocol | EduVis, AccessViz, DVS education groups, instructors, accessibility communities |
| 3 | Executives, editors, and broad AI readers | Board discussion, investment planning, or a need to interpret a striking claim | Understand what is possible, what remains risky, and where investment is warranted | Sponsor, pilot, require evidence, or decline | Five-page executive brief and two figures | Trusted intermediaries, invited briefings, newsletters, podcasts, LinkedIn |

### Audiences are not amplifiers

Editors, community hosts, newsletter writers, podcast hosts, conference
organizers, instructors, and vendors may carry the work to readers. They are not
automatically the people whose decision the work should serve. Each outreach
pitch should distinguish:

- **reader:** who uses the material;
- **decision:** what they can do differently;
- **carrier:** who already has permission to reach them; and
- **proof object:** the figure, result, protocol, or dataset that makes carrying
  it worthwhile.

This prevents “distribution” from becoming a list of places to paste a link.

## Audience dossiers

### 1. Visualization and data practitioners

**Who this includes.** Independent visualization designers, data analysts,
analytics engineers who also communicate results, BI authors, newsroom graphics
practitioners, and developers who build custom visual interfaces. They vary in
code fluency, but all own some part of the path from question to delivered
artifact.

**Decision moments.** The blank page; a request outside the creator's normal
stack; a recurring transformation; a generated chart that is almost right; a
repair that breaks something else; a handoff or maintenance obligation.

**Jobs.** Get to a useful first representation, understand unfamiliar code or
data, compare alternatives, make precise corrections, prove the numbers and
interactions, deliver in the real environment, and preserve a record another
person can maintain.

**Questions in their language.** “Can it make the chart I mean?” “Can I edit the
result?” “Where did that number come from?” “Does it work with live data?” “How
many prompts will this take?” “What did it take to make this route usable?”
“What did we plan, and what did each task actually use?”
“Which phase made it slow or expensive, and was the route cold or warm?”
“Was the affected route independently rechecked after the repair?” “Which
parts still need me?” “Will I spend the saved time checking it?”

**What to give them.** A surface-by-job map; the possible/accomplished/experienced
threshold; a checklist for data meaning, transformation, interaction,
accessibility, and handoff; a clock that begins with route preparation and ends
at accepted delivery; a planned budget beside the observed resource receipt;
a phase receipt that separates cold start, discovery, planning, tool/result
transfer, context update, and synthesis;
a two-receipt accepted-cost card that keeps every attempt and failure in the
route numerator and only contract-passing artifacts in the denominator;
a five-receipt incident chain from report through independent recheck; and
the four-state handoff ladder—accepted change, repeat contribution, maintenance
authority, independent recovery; and worked examples where assistance both
helped and failed. Lead with the job, not
model names.

**Where they gather.** The Data Visualization Society describes its community
as a place for practitioners and enthusiasts to connect and offers Slack,
events, and interest or local groups through membership. Its 2024 State of the
Industry survey was disseminated through email, Nightingale, Slack, LinkedIn,
X, and local meetups, showing a real multi-route field network rather than a
single feed. The survey had 980 starts and 763 completions; the existing
practitioner report uses the 825 respondents who answered its AI question and
keeps that denominator visible. [DVS membership](https://www.datavisualizationsociety.org/membership),
[DVS State of the Industry 2024](https://www.datavisualizationsociety.org/soti-report-2024).

**Access strategy.** Earn editorial treatment first. Nightingale accepts
original member submissions in a roughly 500–3,000 word range and describes
promotion through the DVS newsletter, Slack, and LinkedIn. Pitch a
self-contained finding or practical protocol, not a summary of everything.
Respect its exclusivity terms and do not duplicate the submitted essay during
the stated window. [Nightingale submission guidance](https://nightingaledvs.com/article-submission/).

### 2. BI and analytics leaders

**Who this includes.** Heads of analytics, BI platform owners, analytics
engineering leaders, data governance leads, report developers, and technically
engaged business sponsors. Authors, platform owners, and budget owners are not
the same persona; the asset should make their distinct responsibilities visible.

**Decision moments.** A vendor adds AI to an existing license; leadership asks
for a conversational interface; a pilot reaches production data; an assistant
returns one wrong executive number; or an agent appears capable of rebuilding a
dashboard end to end.

**Jobs.** Separate demo capability from acceptable organizational behavior;
determine whether semantic models and metadata are ready; choose pilot scope;
define refusal, review, logging, escalation, and actor-separated recovery; and
estimate total human and machine cost.

**Questions in their language.** “Does it respect the semantic model?” “Can it
use certified metrics?” “What can it read beyond the visible report?” “Which
permission and capacity state is actually in force?” “Can we reproduce and
correct the wrong answer?” “Who owns it?” “What is fixed, periodic, and per
attempt—and over what volume?” “Is this conversational analytics or report
authoring?” “Who verified the corrected release?” “Do planned and observed
resources reconcile?” “Do we still need dashboards?” “How do we govern agents
using MCP?” “Which host, client, model, cache, streaming, concurrency, and retry
state produced this total?”

**What to give them.** A two-page decision brief, a pilot scorecard, a semantic-
readiness checklist, an incident scenario, and an eleven-lane total-cost ledger.
Keep preparation, ownership, marginal attempts, failures, and downstream
delivery separate; declare volume, useful life, and amortization; and recompute
every claimed denominator before procurement. Put the planned cap beside
observed calls and retries, tokens or image units, evaluator and tool work,
accelerator use, latency and concurrency, charges, and human minutes. Attach a
topology card, phase receipt, and accepted-artifact denominator; keep hidden
internal work explicitly missing rather than inferring it from the user-visible
response. The report's central distinction
between possible output and accepted work is more useful than a tool ranking.

**Context-envelope and correction receipt.** Before a pilot result counts,
declare and retain five controls:

| Control | Before the attempt | After the attempt |
| --- | --- | --- |
| Surface scope | Current visual/page, all pages, bookmarks, report data, semantic model, or another connected surface | What the assistant actually consulted, or explicit unknown |
| Identity and permission | User/service identity, workspace and capacity, report role, Build permission, RLS and other policy | Effective context observed; do not infer it from the UI label |
| Semantic inclusion | Certified metric, visibility, Q&A inclusion, naming/synonyms, relationships, filters, expected lineage | Object or query used, or unresolved binding failure |
| Acceptance case | Prompt, expected answer/evidence or refusal, product surface, version/date, slicer/filter state | Raw response, rendered result, exposed query/explanation, comparison and disposition |
| Correction and incident | Owner, allowed repair, stop rule, escalation and regression set | Report, repair, corrected release, maintainer recheck, independent recheck, collateral regressions, elapsed cost, and incident owner |

Two captured Fabric Community sources motivate these fields. One unresolved
practitioner issue reports model-grounded responses under a specific shared-
capacity, View-only/no-Build setup; it is not a confirmed bug or security
finding. One Microsoft-authored customer account links incorrect or inconsistent
answers to model structure, filter context, inclusion, naming and synonyms, then
reports repairs without disclosing its test denominator. These are source-
wording and preparation-pattern evidence, not a platform verdict or efficacy
comparison. [Fabric permission-scope issue](https://community.fabric.microsoft.com/t5/Issues/Copilot-Returning-Model-Grounded-Answers-for-Reports-in-Shared/idi-p/5142029),
[Microsoft-authored implementation account](https://community.fabric.microsoft.com/t5/Power-BI-Community-Blog/Power-BI-Powered-by-Copilot/ba-p/4378937).

**Where they gather.** The official Tableau community exposes user groups,
forums, Tableau Public, and Slack; its Slack page names channels including
AI-specific discussion and visualization feedback. An official AI and Tableau
user group runs recurring virtual sessions around best practices and governance.
The dbt community offers Slack and local or virtual meetups for analytics
practitioners. These are separate communities with different norms, so a useful
BI-specific artifact should be discussed in one appropriate venue at a time,
not sprayed across all of them. [Tableau community Slack](https://www.tableau.com/community/slack),
[AI and Tableau user group](https://usergroups.tableau.com/ai-tableau-user-group/),
[dbt community](https://www.getdbt.com/community).

**Access strategy.** Recruit pilot reviewers from author, platform-owner, and
governance roles before public launch. After revision, offer a concrete session:
“How to tell whether an AI dashboard pilot is ready for governed work.” Use one
worked scenario and let the scorecard be the handout.

### 3. Visualization, HCI, and AI researchers

**Who this includes.** Visualization and visual-analytics researchers, HCI
researchers, NLP and agent researchers working on visualization, benchmark and
evaluation authors, and graduate students entering the literature.

**Decision moments.** Scoping a literature review; positioning a paper; choosing
a benchmark; deciding whether a product demonstration supports a research
claim; or selecting the next study.

**Jobs.** Find the primary literature, compare unlike evaluation regimes
without flattening them, identify contradictions and nulls, reuse taxonomies,
and formulate a study whose outcome would change what the field believes.

**Questions in their language.** “What is the task definition?” “What is the
baseline?” “Is this text-to-vis, chart-to-code, visual analytics, or agentic
visualization?” “What does the benchmark actually measure?” “Was there a human
study?” “Does author success transfer to reader outcome?” “What is the
ecological validity?” “Are resources actually equal, or did the arms merely
match a round or candidate count?” “Which claim is direct, which is a bridge,
and which later evidence lane is still missing?”

**What to give them.** Stable claim IDs or section anchors, a benchmark
crosswalk, machine-readable study cards, a negative-findings table, an explicit
gap ledger, methods, an equal-resource receipt, event-level phase traces with
effective run manifests and observability state, and a frozen citable release.
Researchers should be able to reuse a table or dispute a claim without
reverse-engineering a designed page.

**Where they gather.** The official IEEE VIS 2026 workshop program includes
VISxGenAI (“GenAI, Agents, and Future of VIS”), BELIV, AccessViz, EduVis,
VISxAI, and related venues. That program is direct evidence that the topic
crosses generation, evaluation, accessibility, education, and visual analytics.
It does not prove that every attendee wants this report. The 2026 submission
deadlines have passed; the responsible current route is discussion with
organizers and attendees where invited, followed by a 2027 submission plan.
[IEEE VIS 2026 workshops](https://ieeevis.org/year/2026/info/program/workshops/).

**Access strategy.** Release a frozen research edition with citation metadata.
Deposit an approved snapshot in a repository that can assign a persistent DOI;
Zenodo describes a DOI as a persistent identifier that supports citation and
discovery. Add a root CITATION.cff if the public repository is meant to be
cited; GitHub then exposes a “Cite this repository” surface and can point to the
preferred report citation. These are publication choices and require owner
approval. [Zenodo DOI guidance](https://help.zenodo.org/docs/deposit/describe-records/reserve-doi/),
[GitHub citation files](https://docs.github.com/en/repositories/managing-your-repositorys-settings-and-features/customizing-your-repository/about-citation-files).

### 4. Data journalists, graphics editors, and newsroom developers

**Who this includes.** Reporters who analyze data, graphics and visual editors,
news application developers, data editors, investigative teams, and the people
who set newsroom policy or training.

**Decision moments.** An analysis deadline; a coding agent used on source data;
a generated graphic proposed for publication; a new newsroom policy; or a need
to explain how a result was produced.

**Jobs.** Accelerate exploration or implementation without surrendering source
custody, reproducibility, editorial judgment, disclosure, accessibility, or the
reader test. Know which evidence belongs in the published explanation.

**Questions in their language.** “Can we reproduce it?” “Can an editor inspect
the transformation?” “What did the agent invent?” “What needs disclosure?” “Is
the source material allowed to leave our environment?” “Will a reader understand
the claim on a phone?”

**What to give them.** A newsroom use protocol, a provenance checklist, a short
benchmarking guide, and a worked example that begins with source custody and
ends with the delivered reader surface. Keep product promotion out of it.

**Where they gather.** NICAR 2026 included sessions framed as “Using coding
agents for data analysis,” “Build your own AI benchmark,” and practical AI in
the newsroom. Those titles are useful language evidence: the audience is
looking for bounded methods and tests, not only “AI visualization.” NICAR 2026
has passed; the next formal opportunity should be checked through IRE rather
than implied to be open. OpenNews/Source explicitly connects developers,
designers, and data analysts working in journalism through code. [NICAR 2026](https://www.ire.org/training/conferences/nicar-2026/),
[OpenNews/Source community](https://source.opennews.org/people/).

**Access strategy.** Seek two newsroom reviews before packaging the protocol.
Then propose a training or conference session whose promise is operational:
“A reproducible acceptance test for AI-assisted data analysis and graphics.”
The complete report is follow-through evidence, not the session abstract.

### 5. Product, engineering, and tool teams

**Decision moments.** Adding an agent to a product, choosing an intermediate
representation, deciding between a specialist model and a general agent,
interpreting a benchmark win, or responding to a failure that the current eval
did not catch.

**Jobs.** Choose where model reasoning adds value; bind it to authoritative
context; expose inspectable state; create recovery paths with named verifier
identity; and test the complete delivered artifact rather than a screenshot,
successful execution, closed issue, or maintainer-only recheck.

**Language.** “Agentic workflow,” “MCP,” “tool calling,” “semantic layer,”
“structured representation,” “critic,” “browser evaluation,” “repair loop,”
“ablation,” “phase trace,” “effective run manifest,” “human-in-the-loop,” and
“end-to-end dashboard.”

**Assets and route.** Give this audience the system anatomy, technique matrix,
evaluation ladder, and reproducible ablation protocol. A repository-ready
methods page should include topology-bound event identity, cold/warm and cache
state, retries, phase timing and usage, charges or accelerator receipts, stop
reasons, and explicit missing internal phases. Compact diagrams can travel
through GitHub, Observable, and engineering blogs; only original results should
be submitted as “news.”

### 6. Educators, literacy, and accessibility specialists

**Decision moments.** Revising a curriculum, deciding what assistance is
permitted, assessing whether students learned or merely completed a task, or
reviewing a generated interface for disabled readers.

**Jobs.** Separate assisted performance from durable learning; preserve critique
and data/domain reasoning; evaluate transfer without the tool; begin access with
co-creation rather than automated remediation; let users inspect, challenge, or
take an alternative path around generated descriptions; and test the delivered
interaction with representative readers and assistive technology.

**Language.** “Visualization literacy,” “learning outcome,” “unassisted
transfer,” “retention,” “cognitive offloading,” “accessibility from the start,”
“co-creation,” “verification,” “user agency,” “alternative pathways,” “screen
reader,” “low vision,” “mobile,” and “reader comprehension.”

**Assets and route.** Give this audience the human-skills model, a classroom or
training protocol, and an accessibility acceptance sheet that covers
participation, multimodality, disclosure, verification, agency, maintenance,
and sustained use. VIS workshops such as EduVis and AccessViz provide obvious
intellectual homes; DVS education and accessibility groups may provide
practice-facing routes. The AccessViz 2025 outcome report treats adoption as a
socio-technical process and AI-mediated access as a trust and power question,
not merely an alt-text feature. Its participant denominator is not stated, so
it supplies design requirements rather than prevalence evidence.
[AccessViz 2025 outcome report](https://accessviz.github.io/docs/report_2025.pdf).

## Jobs and decisions: the product matrix

The same chapter can support several readers, but the decision artifact should
be explicit about who holds authority and what counts as an answer.

| Decision product | Primary user | Trigger | Decision supported | Required inputs | Output | Acceptance evidence |
| --- | --- | --- | --- | --- | --- | --- |
| AI visualization workflow evaluator | Practitioner | Starting or repairing a piece of work | Delegate, assist, or keep manual for each stage | Purpose, data, stakes, delivery surface, edit needs | Stage-by-stage workflow and review plan | User can choose a path and name the retained human responsibilities |
| BI pilot and procurement scorecard | BI platform and governance leads | New license, pilot, renewal, or incident | Pilot, buy, constrain, or stop | Semantic model, permissions, expected questions, incident path, planned and observed resources, topology card, and phase receipt | Readiness score with blocking conditions and pilot design | Named owner accepts scope, volume, useful life, and amortization; report, repair, corrected release, maintainer recheck, and independent recheck remain distinct; planned and observed resources reconcile by phase and effective topology |
| Benchmark and research-gap crosswalk | Researcher | Study or evaluation design | Reuse, extend, or reject an evaluation | Task, data, model, scaffold, renderer, judge, human sample, outcome | Comparable study cards and unanswered question | Reader can trace every comparison and state what result would change the synthesis |
| Newsroom use and publication protocol | Editor or newsroom developer | AI-assisted analysis or visualization proposed | Permit, revise, disclose, or prohibit | Source custody, transformations, generated code, editorial claim, delivery surface | Reproducibility and publication checklist | A second person reproduces the result; editor and representative reader checks pass |
| Agent or product evaluation protocol | Product or engineering team | Feature, skill, model, or architecture decision | Ship, change mechanism, or continue testing | Current-model baseline, tool/harness, task set, cost, failures, user context | Equal-budget test and defect ledger | Results survive deterministic and rendered checks; scope and regressions are visible |
| Learning and access protocol | Educator or accessibility lead | Curriculum or delivered-interface review | Permit assistance, revise teaching, or reject delivery | Learner profile, learning goal, assistive context, unassisted task | Assisted and unassisted outcome record | Retention/transfer or representative accessibility evidence, not completion alone |
| Executive evidence brief | Sponsor or commissioner | Investment or policy decision | Fund, pilot, require controls, or decline | Decision, stakes, current workflow, acceptable evidence | Five-page brief with unresolved risks | Sponsor can state the next gated decision without overclaiming the research |

### A reusable decision-product specification

Every derived asset should carry these fields:

1. **For whom:** role, environment, existing practice, and authority.
2. **At what moment:** the observable trigger that makes the material timely.
3. **Decision:** the actual fork, including a legitimate “do not use AI” branch.
4. **Inputs:** what local context the reader must supply.
5. **Evidence:** which report claims and limitations support the guidance.
6. **Output:** the record the reader leaves with, not merely information read.
7. **Acceptance:** how to know the decision path worked.
8. **Reopen condition:** what product, model, benchmark, or field evidence would
   require revisiting it.

This is the useful inheritance from the adjacent automated-consulting work: one
maintained corpus can produce several moment-shaped artifacts, but each artifact
must still have a named human user, decision, and acceptance test.

## The language map

### Use one canonical term and several audience routes

**Canonical umbrella:** AI-assisted data visualization.

It accurately covers human-led work that uses AI without claiming autonomy. It
also keeps research, practice, products, and reader outcomes in one field.

**Public subtitle:** What AI chart and dashboard systems can do in 2026, where
they fail, and how to evaluate them for real work.

“State of” signals comprehensiveness but not the reader's job. The subtitle and
route titles should answer that deficiency.

| Context | Words readers encounter or use | How to use them |
| --- | --- | --- |
| Broad tool discovery | AI data visualization; AI visualization tools; AI chart generator; AI dashboard generator; ChatGPT data visualization; best AI for data analysis | Use in plain-language route copy and FAQs. Do not turn the work into a “best tools” ranking the evidence cannot support. |
| Enterprise BI | Power BI Copilot; Tableau Agent; conversational analytics; generative BI; agentic analytics; semantic model; certified metrics; governance | Use in the BI route, scorecard, worked scenarios, and metadata. Keep vendor feature terms distinct from independent evidence. |
| Practitioner workflow | first draft; live data; precise edits; accurate numbers; show its work; prompts; DAX or SQL; dashboard; handoff; maintenance | Use in headings and examples. These describe lived friction more clearly than “capability frontier.” |
| Agent and engineering | MCP; tool calling; agent skills; end-to-end dashboard; browser validation; critic; repair loop; observability | Use in the builder route. Define acronyms and keep architecture terms out of the executive entry. |
| Research | AI-assisted visualization; agentic visualization; LLM-based visualization generation; text-to-vis; chart-to-code; visual analytics; mixed-initiative; human-AI collaboration; benchmark; ecological validity | Use in abstracts, methods, literature pages, and citation metadata. Reserve “agentic visualization” for systems with meaningful planning or action, not all AI help. |
| Journalism | coding agents for data analysis; practical AI in the newsroom; build your own AI benchmark; reproducible workflow; disclosure; reader trust | Use in the newsroom route and session pitch. “AI data visualization” alone is too detached from editorial work. |
| Education and access | visualization literacy; learning outcome; transfer; retention; cognitive offloading; accessibility; screen reader; low vision; mobile comprehension | Use in the human-skills route and in outcome labels, not a generic ethics appendix. |
| Anxiety and professional identity | Does this replace dashboards? Are we cooked? What am I charging for? Which numbers should I distrust? | Address the underlying concern directly. Do not use fear language as clickbait or treat bounded public posts as prevalence evidence. |

A dated search and community-language sweep found that exact “AI-assisted data
visualization” phrasing is less common in broad tool discovery than “AI data
visualization” or product-specific terms. “Agentic visualization” is already
useful in research and system-design contexts, while vendors and BI communities
use “conversational analytics” and “agentic analytics.” The correct response is
an intentional vocabulary bridge, not keyword substitution throughout the
report. [Agentic Visualization research paper](https://arxiv.org/abs/2505.19101),
[Tableau's agentic-analytics framing](https://www.tableau.com/agentic-analytics).

### Language evidence is not demand

The 15 August source pass added four useful boundaries to the language map:

- an independent visualization podcast asks how people actually use dashboards,
  where generative AI enters visualization work, whether it is effective, and
  what would make it safer; it is an interview and research preview, not a
  prevalence estimate ([PolicyViz with Melanie Tory](https://policyviz.com/podcast/dashboards-ai-data-visualization-melanie-tory/));
- a Spanish public-data exercise frames the work as converting natural-language
  questions into analysis through explicit dataset context, executable code,
  retries, corrections, and visible results; it is one implementation account,
  not evidence for a general Spanish-speaking market
  ([datos.gob.es implementation exercise](https://datos.gob.es/es/conocimiento/chateando-con-datos-publicos-una-aplicacion-practica-de-inteligencia-artificial));
- the AccessViz workshop makes co-creation, durable deployment, verification,
  alternative pathways, and user agency native language for AI-mediated
  accessibility; the report does not state a participant denominator; and
- official search documentation shows how the site's own query evidence could
  later be interpreted, while also documenting anonymized queries, row
  truncation, normalized Trends scores, filtering, and low-volume noise
  ([Search Console query dimensions](https://support.google.com/webmasters/answer/17011259?hl=en),
  [Google Trends methodology](https://support.google.com/trends/answer/4365533?hl=en)).
- a practitioner-authored Fabric issue uses the concrete language of semantic-
  model-grounded responses, shared versus Fabric-backed capacity, Viewer versus
  Build permission, and control over access to existing models; the issue is a
  single unresolved PoC report, not confirmed product behavior
  ([Fabric permission-scope issue](https://community.fabric.microsoft.com/t5/Issues/Copilot-Returning-Model-Grounded-Answers-for-Reports-in-Shared/idi-p/5142029)); and
- a Microsoft-authored community account describes incorrect or inconsistent
  answers, context and binding diagnosis, semantic-model preparation, and
  before/after validation in one customer case; its test denominator is absent,
  so the reported success and ROI claims are not used as measured evidence
  ([implementation account](https://community.fabric.microsoft.com/t5/Power-BI-Community-Blog/Power-BI-Powered-by-Copilot/ba-p/4378937)).

Use a four-rung evidence ladder:

| Rung | What is observed | What it earns | What it does not prove |
| --- | --- | --- | --- |
| 1. Source wording | A term or question in a captured community, editorial, implementation, or research source | A candidate vocabulary bridge or design requirement | That the term is prevalent or that this package is wanted |
| 2. Qualified discovery | A site-specific query or referral reaches the relevant route | Evidence that some readers find the package through that language | That they used it; anonymized and truncated queries remain missing |
| 3. Decision-path use | A reader chooses a route and completes its scorecard, protocol, or crosswalk | Evidence that the translation supported its intended job | That the resulting decision was good |
| 4. Decision consequence | A pilot, workflow, study, publication, curriculum, or access decision changes, is confirmed, or stops | Evidence of a bounded outcome | Population-wide impact |

These passes strengthen rung 1 and define how to measure rung 2. Rungs 2–4 are
still unknown. Broad Google Trends interest sits outside the ladder: it is
context, not site demand. The formerly record-only Fabric issue is now in
private custody through the platform's first-party item RSS, so its bounded
wording can shape the BI scorecard. The remaining audience-language custody
gate is privacy-approved, site-specific query/referral evidence—not more public
search snippets.

### What should AI not optimize away?

The language map also needs a retained-human-value check. Independent editorial
and practitioner sources do not just ask whether AI makes visualization faster.
They ask whether it preserves the work through which people decide what a chart
is for, learn with other people, notice and repair errors, and carry a result
through feedback into use. The May 2026
[Visualising Data Newsletter](https://visualisingdata.kit.com/posts/the-visualising-data-newsletter-issue-24-may-2026)
frames current AI discussion through tensions around text interfaces, creative
friction, and the questions an assistant should ask before proposing a chart. A
single-practitioner
[Viz Responsibly reflection](https://vizresponsibly.substack.com/p/why-humans-ai-for-positive-disruptions)
adds collaboration, whiteboards, teaching, audience research, joy, and
inspectable correction. A 2026 Portuguese-language
[University of Sao Paulo course](https://uspdigital.usp.br/apolo/apoObterAtividade?cod_oferecimentoatv=138120)
puts purpose, audience, co-design, requirements, usability testing, feedback,
and delivery alongside Python, generative AI, and Power BI.

These are bounded signals: one curated newsletter issue, one named reflection
with an informal show of hands and no denominator, and one official 21-hour
course record with 55 places. They do not establish prevalence, effectiveness,
or demand. They do identify four values that every audience translation should
make visible:

| Retained value | Question for an audience layer |
| --- | --- |
| Purpose and audience | Does the route help the reader state why this visualization exists, who must understand or act, and what context the system cannot infer? |
| Collaboration and creative friction | Does assistance leave room for co-design, critique, exploratory detours, teaching, and the interpersonal work through which knowledge is made? |
| Inspectable correction | Can the reader trace extraction and encoding choices, find an error, repair it, and see whether the repair introduced another problem? |
| Feedback, testing, and delivery | Does the route carry a draft through user feedback, usability and accessibility testing, handoff, deployment, maintenance, and actual use? |
| Distributed responsibility | Does the route distinguish accepted contribution, repeated contribution, named maintenance authority, and independent recovery instead of calling all four “handoff”? |

For an executive audience, this can be compressed to one decision question:
**what valuable work are we paying to remove, and what must remain
human-owned?** The full evidence and its limits remain in research custody;
shorter audience versions should translate that answer without silently
discarding it.

### Questions worth answering verbatim

These are stronger discovery doors than generic chapter labels:

- What can AI actually do in a data-visualization workflow in 2026?
- Can AI-generated charts and dashboards be trusted?
- How should a team evaluate Power BI Copilot, Tableau Agent, or conversational
  analytics?
- Which parts of data visualization should remain human-controlled?
- Does an AI visualization tool save time after checking and repair?
- What is agentic visualization, and how is it evaluated?
- Which benchmarks test chart generation, dashboards, interaction, or reader
  understanding?
- How should journalists disclose and validate AI-assisted analysis and graphics?
- What data-visualization skills still matter when AI can produce a first draft?
- How do accessibility and mobile testing change the verdict?

Each should have a direct two- or three-sentence answer, a visible evidence date,
and a route into the deeper report.

## Public artifact architecture

### One canonical hub, not sibling islands

The current rendered artifacts already have substantial strengths: crawlable
text, literal titles and descriptions, responsive viewport declarations, real
heading structure, section anchors, skip links, and visible evidence dates. The
research and practitioner sites are usable long-form reading surfaces.

The content map adds a local prototype of the routing layer: it shows which
draft owns which question, emphasizes four reader paths, compares coverage
against the comprehensive-report promise, and keeps planned bridges visually
separate from existing work. It does not yet supply durable public URLs,
canonical links, source navigation, or publication state.

The current draft package is still a set of sibling artifacts. Inspection on 14
August found no canonical-link declaration, Open Graph or social-card metadata,
author metadata, article structured data, alternate feed, or shared citation
surface in the AI-visualization artifact heads. The artifacts also do not yet
provide one audience router, public change log, or shared source and download
index. These are packaging gaps, not research gaps.

The public package should have this shape:

1. **Canonical hub.** One durable URL, one-sentence thesis, evidence date, scope,
   audience router, major findings, and links to every rendering.
2. **Audience routes.** “I make visualizations,” “I run analytics,” “I research
   visualization,” “I publish data journalism,” “I build AI systems,” and “I
   teach or evaluate access.” Each begins with decisions, not biography.
3. **Executive brief.** HTML first; a well-made PDF as a printable companion.
4. **Comprehensive report.** The connective argument and complete evidence
   accounting.
5. **Standalone essays.** Self-contained capability, workflow, reader-outcome,
   methods, and contested-claim entries with canonical links back to the hub.
6. **Decision tools.** The evaluator, BI scorecard, research crosswalk, newsroom
   protocol, and learning/access protocol.
7. **Reusable evidence.** Public-safe figures, tables, CSV or JSON exports,
   definitions, and claim/source index. Never publish the private source corpus.
8. **Methods and limits.** Search dates, inclusion rules, evidence ladder,
   known absences, conflicts, corrections, and disclosure of assisted methods.
9. **Citation and versions.** Preferred citation, authors or accountable owners,
   date published and modified, frozen release, and change log.

### The route page contract

Every audience route should fit this sequence above the fold:

- **Moment:** “You are deciding whether to…”
- **Short answer:** the strongest conclusion the evidence supports.
- **Decision tool:** a checklist, scorecard, or crosswalk.
- **Three findings:** each linked to its evidence and limitation.
- **Worked case:** a realistic path through the decision.
- **Go deeper:** the relevant report sections and methods.
- **Recheck:** evidence date and next update trigger.

Do not begin with an abstract history of AI or a decorative manifesto. The
reader should know within one screen whether this page can help with the current
decision.

### Search and AI visibility

Google's official 2026 guidance says the same foundations support ordinary
Search and generative AI features: useful original content, crawlable text,
clear structure, semantic HTML, indexability, and a good page experience. It
also says no special AI markup or tiny “AI-sized” chunks are required and that
Google does not use llms.txt. The rollout should therefore invest in the
substance and technical legibility of the canonical site, not an AEO ritual.
[Google's guide to optimization for generative AI](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide).

At publication:

- use stable, descriptive URLs rather than dates alone;
- add canonical links, sitemap, robots policy, RSS or Atom for material updates,
  and internal links between every rendering;
- supply unique title and description metadata, Open Graph and social cards,
  accountable author or organization, date published, and date modified;
- use Article or TechArticle structured data only when the visible page supports
  every field, and Dataset metadata for genuinely downloadable public datasets;
- keep substantive findings in HTML text rather than images, canvas, or a PDF;
- give each major table and figure a stable anchor, caption, alt text, source,
  and downloadable form;
- expose a preferred citation and a frozen release only after publication is
  approved; and
- verify indexation and actual referrals separately from merely shipping tags.

The website is the primary reading and linking surface. The PDF is for printing,
offline reading, and institutional circulation. It should not be the only
indexable version.

After publication, use Search Console's query dimension only as qualified
discovery evidence. Preserve impressions, clicks, page, country, device, date
window, and known omitted-query share where available; do not turn the displayed
row list into a complete denominator. Google Trends may help compare broad
relative interest in candidate terms, but its sampled, normalized 0–100 values
do not describe this site's readers. Neither source measures whether a reader
made a better decision.

## Rollout strategy

The sequence is **reader proof, canonical release, earned interpretation,
sustained evidence**. The public report should launch once, but the audience
routes should travel over several weeks.

### Phase 0 — prove decision usefulness before publication

Timing: approximately two weeks, or until the acceptance threshold is met.

Recruit 8–12 named reviewers, with at least two from each of the first four
audiences. Include different authority positions: practitioner and reviewer; BI
author and platform/governance owner; early-career and senior researcher;
reporter/developer and editor.

Give each reviewer one scenario, not the whole site and a satisfaction survey.
Observe whether they can:

1. identify their route without explanation;
2. find the evidence relevant to a named decision;
3. distinguish provider claims, benchmark results, human evidence, and unknowns;
4. produce the intended decision record;
5. name what remains unresolved; and
6. cite or send the right unit to a colleague.

Record time to decision, wrong turns, unanswered questions, language mismatches,
and whether the artifact changed, confirmed, or failed to affect the decision.

**Gate:** do not launch until at least one reader in each primary audience can
complete its decision path without live guidance and no unresolved defect would
misstate evidence, expose private source material, or obscure ownership.

### Phase 1 — publish the complete canonical package

Publish together:

- the hub and audience router;
- executive brief;
- comprehensive report;
- research and practitioner companions;
- at least three decision tools: practitioner evaluator, BI scorecard, and
  research crosswalk;
- methods, public-safe source index, preferred citation, and corrections policy;
- public change log and evidence/recheck date; and
- social cards and a compact set of reusable figures.

The launch note should make one claim, show one original figure, state who the
work is for, and link to the router. Avoid “we are thrilled to announce,” an
exhaustive feature list, or a thread that reproduces the report without giving a
reader a reason to use it.

**Gate:** human owner approves public claims, source rights, authorship,
disclosure, repository/site destination, DOI choice, and publication. HTTP,
mobile, accessibility, metadata, links, downloads, and analytics are separate
acceptance checks.

### Phase 2 — earned distribution by audience

Timing: three to four weeks after the canonical release. The exact dates depend
on editorial and community permission.

**Week 1: visualization practice.** Pitch one original essay or protocol to
Nightingale. Share in DVS only under the applicable community rules, framed as a
substantive question and resource. Offer the practitioner evaluator, not just
the report link. Give newsletters and podcasts one figure plus the evidence
behind it.

**Week 2: BI and analytics.** Run or propose a concrete scorecard session for an
AI and Tableau, Fabric/Power BI, or analytics-engineering group. Use product
names in the scenario but keep the evaluation vendor-neutral. Ask participants
to test the scorecard against one real pilot and report missing inputs.

**Week 3: research.** Circulate the research edition and benchmark crosswalk to
relevant labs and workshop participants. Offer it as a citable synthesis and
request corrections, conflicting evidence, and missing studies. The 2026 IEEE
VIS submission cycle is already past; plan a 2027 paper, workshop contribution,
or dataset release only if the resulting work is genuinely original.

**Week 4: journalism and education.** Offer the newsroom protocol to IRE/NICAR
and OpenNews-adjacent practitioners for review or training. Offer the learning
and access protocol separately to EduVis, AccessViz, and teaching communities.
Do not collapse editorial accountability and classroom learning into one ethics
essay.

Good secondary carriers include established visualization newsletters and
podcasts, but outreach should wait until there is a distinctive finding, figure,
or protocol they can discuss. Personal LinkedIn posts can make the work visible
to existing relationships; they are exhaust from the release, not the strategy.

### Phase 3 — keep the evidence alive without creating a content factory

- Publish a change note only when a material claim, source, tool contract,
  benchmark, or forecast status changes.
- Review high-volatility product and model statements monthly internally.
- Publish a quarterly evidence update or gap-ledger delta if it contains real
  changes.
- Preserve dated releases; never silently rewrite the historical snapshot.
- Invite corrections and missing primary sources through a bounded form.
- Turn repeated reader questions into better route pages or decision tools.
- Publish new experiments when they add evidence, not to fill a cadence.

## Channel map and participation rule

| Route | Audience reached | Native contribution | Evidence of fit | Participation rule | First ask |
| --- | --- | --- | --- | --- | --- |
| Nightingale | Visualization practitioners and adjacent readers | Original essay, case, or method | Publisher explicitly accepts member work and promotes through its network | Join/submit under editorial and exclusivity terms | Editorial fit for one standalone finding |
| DVS Slack, groups, and events | Practitioners, researchers, educators | Discussion, peer review, event, research recruitment | Official membership and research-recruitment surfaces | Follow channel rules; no repetitive link dropping | Review a workflow evaluator or answer one field question |
| Tableau AI user group and Slack | BI authors, admins, consultants, product users | Live demo, governance case, office-hours discussion | Official AI group and named Slack topics | Use product-specific scenario; disclose independence | Test the BI pilot scorecard |
| Microsoft Fabric or Power BI user groups | BI practitioners and platform owners | User-group session or technical discussion | Candidate route; the configured capture ladder could not preserve the current events page | Recapture and verify the current group and organizer route before use; do not cross-post indiscriminately | Critique a semantic-model and incident scenario |
| dbt community | Analytics engineers and data leaders | Governance or semantic-layer discussion | Official Slack and meetup network | Keep visualization claim tied to analytics-engineering work | Review the context/readiness section |
| IEEE VIS workshops and labs | Researchers, builders, students | Paper, position, dataset, poster, or reading-group discussion | 2026 program has several directly relevant workshops | Respect closed deadlines and archival standards | Correct the benchmark crosswalk and gap ledger |
| NICAR/IRE | Reporters, editors, newsroom developers | Training, session, reproducible protocol | 2026 program used practical agent and benchmark language | Verify next call; serve newsroom work rather than promote a product | Test the newsroom protocol |
| OpenNews/Source network | Journalism developers, designers, analysts | Case study, method, community discussion | Official community identity matches the intersection | Follow community norms and contribute operational detail | Review a reproducible worked case |
| GitHub and archival repository | Researchers and builders | Versioned report, data, methods, issue-based corrections | Native citation and release mechanics | Public-safe material only; clear license and custody boundary | Reuse or challenge a stable artifact |
| Trusted newsletters and podcasts | Mixed but already opted-in audiences | Interview, figure, case, or debate | Editorial relationship and thematic fit | Pitch one audience-specific story; do not send a generic blast | Discuss one defensible finding and its consequence |

The DVS research-recruitment program is a particularly appropriate route for
future original studies: it describes a monthly opt-in email to members and
states that member data is not shared with researchers. That is a research
recruitment mechanism, not a report-promotion loophole. Use it only when there
is a reviewed protocol and human-subjects obligations are clear.
[DVS research recruitment](https://www.datavisualizationsociety.org/research-recruitment/).

## Visibility and usefulness metrics

Use a funnel that preserves the difference between being available and changing
work:

| Level | Question | Useful measures | What it does not prove |
| --- | --- | --- | --- |
| Eligibility | Can the right reader or system retrieve it? | Successful crawl, indexed canonical pages, valid metadata, working feeds, HTTP and mobile checks | That anyone noticed or read it |
| Qualified circulation | Did intended readers reach the relevant route? | Route views by referral and audience, engaged reading, return visits, methods/source-index use | That the material changed a decision |
| Decision use | Did someone use the artifact for its intended job? | Scorecard completions, protocol runs, decision records, report sections cited in internal memos, reviewer scenarios completed | That the decision produced a good outcome |
| Reuse | Did the evidence travel with attribution? | Backlinks, academic citations, syllabus inclusion, figures or datasets reused, methods referenced | That readers interpreted it correctly |
| Relationship | Did qualified readers consent to a continuing connection? | Update subscriptions, interview opt-ins, correction submissions, community invitations | Revenue or durable adoption |
| Outcome | Did work change? | Pilot changed or stopped, evaluation run, policy revised, research gap adopted, error prevented, reader test added | Population-wide impact |

Report denominators and segment the measures by audience route. Do not combine
one highly engaged research cohort with general traffic and call the average
“engagement.” Do not optimize raw visits if the work is meant to change a small
number of consequential decisions.

### First-cycle targets

Targets should be learning thresholds, not synthetic growth forecasts:

- 8–12 prelaunch decision-path reviews across the four primary audiences;
- one corrected route or artifact from each audience's feedback;
- three or more completed uses of each launch decision tool;
- at least five attributable reuses across citations, internal protocols,
  syllabi, presentations, or linked technical work within the first quarter;
- at least two concrete outcome accounts, including “we did not adopt” or “we
  narrowed the pilot” as legitimate positive evidence; and
- a public ledger of substantive corrections and missing evidence.

If the package receives broad traffic but no recorded decision use, the problem
is not solved. If a small number of credible teams use it to improve or stop a
decision, that is meaningful early success.

## Ownership and gates

| Work | Canonical owner | Handoff or gate |
| --- | --- | --- |
| Private source custody, captures, grades, and question ledger | Private evidence owner | Never expose private captures merely because the synthesis is publishable |
| Cross-source synthesis, relevance, recommendations, and experiment agenda | Research synthesis owner | Every material public claim traces to an eligible source and states its limit |
| Public report design and route implementation | Public presentation owner, not the private source notebook | Preserve content parity, accessibility, stable URLs, and report authority |
| Publication, authorship, licensing, DOI, and outreach | Human owner | Explicit approval per surface; one approval does not authorize every channel |
| Community participation | Named human participant or approved representative | Respect community rules and answer questions; no unsupervised promotion |
| Audience and outcome evidence | Named audience/research owner | Keep observed use distinct from traffic, citations, and inferred value |

No automated posting, synthetic persona outreach, private-community scraping,
or posting quota belongs in this strategy. Automation can prepare views,
maintain evidence, check links, assemble release assets, and surface approved
opportunities. A human should own relationships, claims, and participation.

## Risks and countermeasures

| Risk | Likely symptom | Countermeasure |
| --- | --- | --- |
| The comprehensive report becomes the only product | Praise without use; high abandonment; vague sharing | Route by decision and ship tools that create a decision record |
| A tool-ranking frame overwhelms the evidence | Search traffic but rapid staleness and vendor comparison demands | Explain tool families and evaluation; refuse unsupported “best” rankings |
| The audience language fragments the field | Separate pages make inconsistent claims | One canonical vocabulary and evidence base; audience terms are routes, not rival truths |
| Community distribution reads as promotion | Deleted posts, low trust, no substantive replies | Earn access, contribute native value, ask one real question, and follow host rules |
| Private source custody leaks into public work | Copyright, confidentiality, or provenance failure | Publish only public-safe synthesis, short attributed quotations where allowed, and source metadata |
| A flashy figure loses its denominator | Attention followed by credibility damage | Put denominator, evidence class, date, and limitation in the figure itself |
| Updating destroys historical citability | Readers cannot reproduce what they cited | Frozen versions plus visible change log and current living view |
| Metrics collapse into page views | Distribution is mistaken for usefulness | Track eligibility, circulation, decision use, reuse, relationship, and outcome separately |
| “Automated consulting” becomes the headline | Readers debate replacement rather than use the work | Lead with the human decision product; describe assisted maintenance in methods |
| Cadence produces thin derivative content | Many posts, fewer reasons to care | Publish on evidence changes and recurring reader questions, not a quota |

## The first six artifacts to build

1. **Canonical hub and audience router.** This is the missing connective public
   surface and should precede broad distribution.
2. **Should AI help with this visualization job?** A practitioner evaluator
   using purpose, stakes, data authority, correction, delivery, and reader
   evidence.
3. **Is this BI agent ready for a real pilot?** A scorecard covering the
   context envelope, semantics, permissions, refusal, reproducible correction,
   incidents, total cost, and maintenance.
4. **What does each benchmark establish?** A research crosswalk with task,
   model, scaffold, renderer, judge, sample, measured outcome, and limits.
5. **A newsroom acceptance protocol for AI-assisted analysis and graphics.** A
   reproducibility path from source custody through editorial and reader review.
6. **A public-safe evidence bundle.** Figures, tables, method, citation, version,
   and change log that make the work reusable without releasing the private
   corpus.

The executive brief, comprehensive report, and current long-form companions are
already substantive. These six pieces make the corpus navigable and actionable.

## Ninety-day sequence

| Window | Work | Exit evidence |
| --- | --- | --- |
| Days 1–14 | Build router prototype and three decision tools; recruit 8–12 scenario reviewers | Completed decision-path records across four primary audiences; blocking defects resolved or explicit |
| Days 15–30 | Complete canonical metadata, public-safe source/index exports, version and citation mechanics, accessibility and mobile acceptance | Human publication approval plus technical and content gates; frozen release candidate |
| Days 31–45 | Canonical launch and visualization-practitioner editorial route | One launched package; one earned editorial conversation or substantive DVS review; evaluator use recorded |
| Days 46–60 | BI-specific scorecard sessions and analytics-community reviews | At least three real or reconstructed pilot uses; missing inputs and governance defects logged |
| Days 61–75 | Research crosswalk circulation and corrections; archive approved release | Corrections or missing studies adjudicated; citable frozen version; 2027 research route chosen or declined |
| Days 76–90 | Newsroom and education/access reviews; publish first material change note | Protocol uses recorded; first outcome accounts; prioritized next evidence work rather than a content calendar |

## What would change this strategy

Revisit audience order if decision-path tests show another audience completing
more consequential jobs with less explanation. Revisit channel choices if
community policies, editorial terms, or event calendars change. Revisit the
artifact family if readers consistently use the full report and reject the
decision tools, or if the tools are used without consulting their evidence and
limits. Revisit the positioning if search and referral evidence shows a more
precise term consistently brings the intended readers without distorting scope.

Most importantly, replace this modeled audience map with observed use as soon
as the first cohort exists. The strategy has succeeded when the report becomes
an evidence spine that real people use to make named decisions—not when it has
been reformatted into the largest number of posts.

## Update log

- **2026-08-16 — Translated B91's access stop without changing its content
  state.** Every route now carries the unchanged `6 / 20 / 1 / 0` content
  ledger beside one access-blocked and unassessed row, zero active generic B91
  targets, and four exact reopen conditions. Research effort moves only when
  new custody appears; paused never becomes negative.

- **2026-08-16 — Translated InfoViP's actor map and closed opportunity across
  seven routes.** Readers can now route research questions to three named
  mentors while preserving zero named service-operation, maintenance,
  QA-execution, release-approval, or authorization roles. Closed remains
  separate from selected, started, built, approved, and released.

- **2026-08-16 — Repaired one InfoViP date and role chain across seven
  routes.** Every route now treats the apparent 2026 future-tense page as
  content-current to 2024, the 2026 project-lead biography as a precise role
  rather than operations authority, and the Elsa 4.0/HALO launch as separate
  from an unobserved InfoViP-specific integration.

- **2026-08-16 — InfoViP's production-component and QA answer translated seven
  ways.** Every route now carries the final report's approval, AWS/AERS
  integration, >30-million-history and ~8,000-daily receipts beside its absent
  QA plan, planned audits, incomplete routine roles, and uninvestigated
  downstream effect. The current no-ATO inventory remains visible, and a 2026
  Elsa/API/UI opportunity is treated only as a future recheck trigger.

- **2026-08-16 — InfoViP field-use, cost, and governance evidence translated
  seven ways.** Every route starts from 20 unique real-work evaluators, a later
  29-million-history plus daily operating pipeline, and a $2.40 million
  obligated / $2.86 million estimated three-award envelope. Every route keeps
  current routine use, acceptance/ATO, outcome, return, whole cost, and ROI
  unresolved; 15 named cases still yield zero complete episodes.

- **2026-08-16 — InfoViP operational-governance result translated seven
  ways.** Added one canonical line from practitioner co-design through an
  operational pilot, at-scale incoming processing, performance and human
  control, and current development/no-ATO counterevidence. The audit is now 15
  named cases and zero complete episodes.

- **2026-08-16 — One publisher-abstract answer translated seven ways.** Every
  route starts from B92's exact multicriteria, Monte Carlo, rank-stability
  abstract and the current `6 / 20 / 1 / 0` ledger. B91 remains unassessed; no
  route promotes a non-AI method into human use, delivery, outcome, or impact.

- **2026-08-16 — One partial-lineage answer translated seven ways.** Every
  route starts from one live page, zero exact public builds, eight of twelve
  matching dependency names, two of three canonical asset matches, and zero
  audience outcomes. No route promotes a source link, demonstration,
  instrumentation, or intended audience into deployment or use.

- **2026-08-16 — One public-delivery afterlife translated seven ways.** Every
  route starts from B110's official abstract, pinned source artifacts, and one
  current same-named live interface while preserving B92 and B91 as
  content-unassessed at that pass. No route promotes reachability into continuous use,
  source custody into the deployed build, or intended audience into measured
  outcome.

- **2026-08-16 — One residual-recovery answer translated seven ways.** Every
  route starts from 14 investigated rows, 11 substantive surfaces, one full
  chapter, three metadata-only rows, one evaluation-to-redesign mechanism, two
  practice-context near misses, and zero complete AI lifecycles. No route
  promotes expert evaluation, stakeholder application, or multi-unit
  evaluation into deployment or outcome.

- **2026-08-16 — One prototype-evaluation answer translated seven ways.**
  Every route starts from nine investigated rows, eight substantive surfaces,
  seven explicit human evaluations, InfoViP's prospective production intent,
  and 14 remaining content-unassessed DOI rows. No route turns future tense
  into deployment, conventional NLP into GenAI, or evaluation into outcome.

- **2026-08-16 — One field-use result translated seven ways.** Every route
  starts from the same eight-key authority repair and non-AI cancer-diary
  comparator while keeping one gated primary text, missing lifecycle receipts,
  and the unknown complete-review denominator visible.

- **2026-08-16 — One primary-evidence ladder translated seven ways.** Five
  recovered full texts add controlled audience measurement, public-sector co-
  design/demo, and enterprise production-intent lanes, but no accepted field
  release plus later affected-audience recheck. Every route preserves 22 DOI
  rows and eight non-DOI stable keys as unassessed.

- **2026-08-16 — One mature platform result translated without borrowing an
  AI role.** Seven routes now keep AIDSVu's ten-year delivery, aggregate use,
  governance, and later data release while retaining its missing AI role,
  version-bound audience outcome, and affected-audience recheck. Eighty-six
  abstract cue nonmatches remain unexcluded.

- **2026-08-16 — One custody answer translated without inventing audience
  evidence.** Seven routes now begin from the same bounded result: 122 titles
  triaged, one explicit GPT signal, one pinned supplemental-output bundle, and
  zero audience-delivery or later-recheck chains. The 121 title nonmatches
  remain not surfaced rather than excluded.

- **2026-08-16 — Longitudinal evaluation boundary translated seven ways.**
  Practice gets an evaluation-to-acceptance card; BI gets a release-linked
  decision ledger; research gets the twelve-receipt matrix; newsrooms get a
  publication gate; builders get a drift-aware release export; learning and
  access keep outcomes separate; sponsors get one near-miss and zero complete
  episodes across 14 named cases. No route promotes development selection or
  reported confidence into audience impact.

- **2026-08-16 — Full 122-of-127 review frame translated across seven
  routes.** Every route gets the same five unresolved authority joins and one
  quarantined identifier conflict. The useful layer changes by audience;
  candidate identity, review provenance, unstarted screening, and the null
  lifecycle count do not.

- **2026-08-16 — One 12-of-13 Education result translated across seven
  routes.** Every route now distinguishes an authoritative review key from a
  plausible external candidate. `M. Lu (2020)` stays unresolved until A208-
  controlled evidence joins it; no audience may promote confidence into
  provenance or a lifecycle count.

- **2026-08-16 — One repair overlay translated seven ways.** Every route now
  receives the same source-preserving result: the supplement still prints `?`,
  while A208's article and DOI metadata uniquely reconstruct Hernández-
  Calderón et al. (2023). The decision language changes by audience; the absent
  publisher correction, unfinished stable-key crosswalk, and unknown lifecycle
  count do not.

- **2026-08-16 — Review inventory integrity translated seven ways.** Every
  route now receives the same three-state answer: 127 studies reported, 126
  appendix citations identified, and the complete GenAI lifecycle count
  unknown. Audience language changes the decision gate without guessing the
  missing Education row or turning it into a negative result.

- **2026-08-16 — Review-lineage boundary translated seven ways.** Every route
  begins with the same 127-study coverage map and unknown qualifying lifecycle
  count, then asks for its own artifact, pilot, study, publication,
  instrumentation, access, or funding receipt. No translation reports zero of
  127 or treats review breadth as field efficacy.

- **2026-08-16 — Conditional retry burden translated seven ways.** Every route
  begins with the same 732-of-888 endpoint and 584-versus-714 revision record,
  then asks for its own stopping, telemetry, acceptance, or reader receipt. No
  translation turns fewer generations into lower total cost or audience use.

- **2026-08-16 — Production recovery translated seven ways.** Every audience
  route begins with the same 3-afterlife, 2-restoration, 0-affected-actor-
  recovery, 0-authority-transfer result, then receives its own decision
  question without turning issue closure or contribution into recovery.

- **2026-08-15 — Reader-outcome boundary translated seven ways.** Each audience
  route keeps the same one-direct, five-adjacent, zero-delivery/recheck result
  while changing the decision question; no route upgrades participation,
  creator interpretation, or model QA into delivered audience evidence.

- **2026-08-15 — Governance boundary translated seven ways.** Practice, BI,
  research, journalism, product, learning/access, and sponsor routes now begin
  with the same 3-control-contract, 2-organizational-use, 2-adjacent-process,
  0-complete result and ask for the first missing same-deployment receipt.

- **2026-08-15 — Public acquisition translated seven ways.** Practice, BI,
  research, journalism, product, learning/access, and sponsor routes now begin
  with one provenance-aware result: 12/12 named listings have install signals,
  while 0/12 exposes invocation, retention, organizational acceptance, or
  outcome. No route turns registry or repository attention into adoption.
- **2026-08-15 — System anatomy translated seven ways.** Practice, BI,
  research, journalism, product, learning/access, and sponsor routes now enter
  one eleven-stage information-and-authority map. The canonical result remains
  6/6 generation-execution-critique, 4/6 meaningful human control, and 0/6
  accepted-delivery, intended-reader, or maintenance chains; no route turns it
  into an architecture winner or observed audience demand.
- **2026-08-15 — Capability-to-practice bridge translated seven ways.** Ten
  primary cases yield three controlled partial bridges and zero exact-version
  accepted-delivery or complete episodes. Practice, BI, research, journalism,
  product, learning/access, and sponsor layers now ask for the first missing
  receipt without changing the canonical result or implying audience use.
- **2026-08-15 — Fixed-budget result translated seven ways.** Selective TTS
  adds one matched declared partial call/output-token budget without supplying
  observed route-wide use or contract acceptance. Practice, BI, research,
  journalism, product, learning/access, and sponsor routes now keep the three
  receipt levels separate. The accepted-cost result advances to 0/11 and no
  route winner is implied.
- **2026-08-15 — Six-receipt null translated seven ways.** Practice, BI,
  research, journalism, product, learning/access, and sponsor routes now use one
  exact release-validation card. The answer stays zero of seven in named
  surfaces; every layer retains the first missing receipt and the return-user
  reopen condition.
- **2026-08-15 — Longitudinal co-design result translated seven ways.** Graphy
  adds three returning blind co-designers, 12 sessions, four workshops, eight
  months, and implemented changes. Every audience layer keeps formal efficacy,
  workshop-version binding, release, representative recheck, adoption, and
  whole cost visibly missing.
- **2026-08-15 — Accepted-cost result translated seven ways.** Practice, BI,
  research, journalism, product, learning/access, and sponsor layers now ask
  for one route-wide numerator and one frozen accepted-output denominator.
  At that evidence cut, zero of nine held fragments supplied both; no route
  winner was implied.
- **2026-08-15 — Version-floor boundary translated seven ways.** Added the
  distinction among study surface, first containing release, participant-tested
  build, later maintenance, representative recheck, and rewrite. MAIDR narrows
  the target but still contributes zero complete lifecycle episodes.
- **2026-08-15 — One lifecycle result translated seven ways.** Added the exact
  artifact-version join, four rejected near-misses, and a decision layer for
  practice, BI, research, journalism, product, learning/access, and sponsors.
  Every route keeps the same result: zero new episodes, a strong pre-AI
  accessibility baseline, and no cross-version human-evidence transfer.
- **2026-08-15 — One validity result translated seven ways.** Added a shared
  five-threat configuration boundary and audience-specific questions for
  practice, BI, research, journalism, product, learning/access, and sponsors.
  Every route keeps the same result: benchmark success is lane-bound and no
  current Vizier configuration winner exists.
- **2026-08-15 — Phase and topology language added.** Practitioner, BI,
  research, and builder layers now translate the same receipt: cold start,
  discovery, execution phase, effective topology and runtime state stay beside
  planned and observed totals; hidden work remains missing.
- **2026-08-15 — Handoff now has four audience-facing states.** The
  practitioner, BI, research, newsroom, product, education/access, and executive
  layers now distinguish accepted change, repeat contribution, maintenance
  authority, and independent recovery. Prism reaches the first two; OpenClaw
  reaches the first and has an open repair proposal. Neither reaches the last
  two.
- **2026-08-15 — Recovery now has five actor-separated receipts.** Prism's
  multi-attempt Home Assistant chain turns report, repair, corrected release,
  maintainer recheck, and independent recheck into a practitioner, BI, research,
  newsroom, product, and access decision rule. The last receipt is missing in
  the case, so the route is not called independently recovered.
- **2026-08-15 — Planned and observed resource receipts.** The practitioner,
  BI, research, and product routes now ask whether each arm stayed inside a
  frozen multidimensional cap and retain what each task actually used. Matching
  rounds, candidates, rates, nominal caps, or model labels no longer counts as
  equal budget; no comparison ran and no cost winner exists.
- **2026-08-15 — Fabric context envelope and correction receipt.** Recovered
  the formerly gated practitioner issue through a first-party RSS representation
  and added a Microsoft-authored implementation account. The BI route now asks
  what the assistant can traverse, under which permissions, which semantic
  objects it can use, and whether a wrong answer can be reproduced, corrected,
  regression-tested, and owned. Both sources are C-grade self-report; they do
  not establish a bug, breach, efficacy, prevalence, demand, or reader use.
- **2026-08-15 — Total-cost decision products expanded.** The practitioner
  clock now begins at route preparation, while the BI procurement route keeps
  five cost classes separate, declares volume and amortization, and rejects
  source telemetry whose denominator does not reconcile. This is a decision
  aid, not evidence that either route is cheaper.
- **2026-08-15 — Retained-human-value check.** Added independent newsletter,
  practitioner, and Portuguese teaching evidence for purpose, collaboration,
  creative friction, audience context, inspectable correction, feedback, and
  delivery. These are candidate evaluation values, not demand or effect
  estimates.
- **2026-08-15 — Audience-language evidence ladder.** Added primary custody for
  an independent podcast, a Spanish public-data implementation, AccessViz
  community outcomes, and official search-measurement limits. Separated source
  wording, qualified discovery, decision-path use, and changed work; expanded
  the accessibility route around co-creation, verification, agency, and
  sustained use. At that pass, Power BI/Fabric community wording remained a
  capture gap; the later first-party RSS pass above closes only that custody
  condition.
- **2026-08-14 — Initial public edition.** Defined audience routes, decision
  products, evidence boundaries, publication gates, and an outcome-oriented
  rollout sequence.