feat(dtpr-ai): v2 schema reshape — Element.context, subchains, PII, AIAAIC research#274
Merged
feat(dtpr-ai): v2 schema reshape — Element.context, subchains, PII, AIAAIC research#274
Conversation
Adds four research corpus entries informing the next AI schema beta: MIT AI Risk Repository v4 (April 2026, 1,725 risks across 7 domains), Atlas of AI Risks (citizen-facing AI city register prior art), AIAAIC Collaborative Harms Taxonomy (the chosen organizing cut for v2's risks_mitigation, victim-oriented over cause-oriented), and AIR 2024 (policy-grounded reference, surveyed and rejected for citizen audience). Adds dtpr-ai/content/9.attribution.md documenting AIAAIC's CC BY-SA 4.0 license, full citation, and the mixed-license boundary between DTPR (CC BY 4.0) and AIAAIC-derived risk content. Establishes the per-element citation pattern via the existing Element.citation field. Stress-test of the v2 structural shape lives in .context/dtpr-ai-v2-structural-proposal.md (gitignored). The schema:new run for ai@2026-04-27-beta follows this commit. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lays the structural seams that the new ai@2026-04-27-beta needs. - Element.context (optional override; fully replaces Category.context) - Nullable color on context values (null = renderer treats value as a tag instead of a colored dot; baked icons skip null variants) - DatachainType.subchains[] (named, ordered groups of categories) - DatachainInstance.subchain_instances[], elements[].subchain_instance_id, elements[].actions[], instance_version, updated_at, sources[], linked_instance_ids[] - Validator rules extended to cover element-level context, null colors, subchain refs/uniqueness, subchain locale strings - json-emitter filters null-color variants from icon_variants and uses effective context for variant resolution Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the same `^[a-zA-Z0-9_-]+$` whitelist Element.id has used since the port. The new convention drops the legacy `<datachain_type>__<slug>` prefix in favor of bare slugs (`accountable`, `functional_modes`, …); the description is updated accordingly. The forbid-`__` refinement from the v2 proposal is intentionally NOT enforced at the structural layer so the existing ai@2026-04-16-beta (which still uses `ai__` ids) continues to validate. The new ai@2026-04-27-beta uses bare slugs by content; the convention can be hardened once the older beta is retired. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lands the schema-level edits captured in the v2 structural proposal. Element-level authoring (the six functional-mode elements, the nine AIAAIC risk elements, the consolidated accountable element with logo support, and the family-typed processing catalog) follows in subsequent dtpr-element-design / dtpr-category-audit passes. - Drop the `ai__` prefix on category directories and IDs (eleven categories renamed to bare slugs). - Rename `ai__decision` → `functional_modes`; empty its element catalog (the six retired decision elements are deleted in beta). - Reorder categories by citizen priority: purpose, accountable, functional_modes, risks_mitigation, rights, input_dataset, processing, output_dataset, access, retention, storage. - accountable: rounded-square → hexagon (port-bug correction); add Role context (vendor/deployer, color: null = tag rendering). - processing: hexagon → circle (matches data-in-motion shape contract); description rewritten to clarify the technical-detail framing. - input_dataset: description rewritten to clarify runtime input (not training data); add PII context (none color: null, anonymized #4A90D9, identifiable #FFD700, biometric #D9342F). - output_dataset: same PII context. - datachain-type.yaml: subchains[data_flow = input_dataset → processing → output_dataset]; pt added to localized name. Validates and builds cleanly via schema:validate / schema:build. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Glosses the two new constructs introduced in ai@2026-04-27-beta so authors and reviewers reach for the same vocabulary. - 5.shape-contract.md: advisory mapping (hexagon = who/why/what, circle = data in motion, rounded-square = data at rest, octagon = context/guardrails). Documents the convention without making the schema enforce it. - 6.subchains.md: glosses datachain-type-level `subchains` and the per-instance `subchain_instances` realization, with the smart- intersection example from the proposal. - 0.index.md: links the two new pages. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Too many files changed for review. ( |
verify.mjs's CORPUS_SLUG_RE expects the slug to end in `.md` (matching the on-disk filenames); the four AIAAIC/AI-risk rows in INDEX.md were missing the suffix, failing the offline plugin conformance check. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
5 tasks
pichot
added a commit
that referenced
this pull request
May 5, 2026
Three fixes to make the previous commit deploy. 1. Drop content.config.ts + app/pages/index.vue c12 loads our user-layer content.config.ts via jiti, which transitively requires @nuxt/content -> @nuxt/kit. jiti evals @nuxt/kit/dist/index.mjs:182's `import.meta.dev` via new vm.Script (CJS context), throwing SyntaxError. Docus's own content.config.ts has the same imports but loads inside a Nuxt context where jiti is configured differently. The homepage UContainer override is a polish item; restoring it needs a docus-friendly extension that doesn't trigger this load chain. Following up separately. 2. Update @dtpr/ui test fixtures for v2 schema reshape The v2 reshape (#274) made `actions` and `subchains` required defaults via z.ZodDefault, which Zod 4 surfaces as required in the inferred output type. vite-plugin-dts type-checks tests during declaration generation, so missing fields blocked dts emission. Add `actions: []` to InstanceElement fixtures, plus `subchains: []`, `subchain_instances: []`, `sources: []`, `linked_instance_ids: []` to DatachainType / DatachainInstance. 3. DtprIcon: reset failed state when isDark changes Watch already reset `failed.value` when src/darkSrc changed. isDark wasn't in the deps, so a dark-mode toggle could leave a stale 404 marker, blocking retry against the new url. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
pichot
added a commit
that referenced
this pull request
May 5, 2026
* feat: all-versions schema deploy + @dtpr/ui dark mode
Three independent changes bundled per request.
1. Schema deploy enumerates every version in the source tree
- .github/workflows/api-deploy.yaml: drop the hard-coded
ai@2026-04-16-beta default, walk api/schemas/<type>/*/, and
run schema:build + r2-upload for each version. Validation gates
the worker deploy. workflow_dispatch.version remains as an
optional single-version override.
- api/scripts/r2-prune.ts: new step that drops index entries
whose source dir was deleted (e.g. after schema:promote
rename). Worker reads gate through schemas/index.json
(api/src/rest/version-resolver.ts), so index-only prune is
sufficient — no R2 object deletion required.
2. @dtpr/ui dark mode
- packages/ui/src/vue/styles.css: dark-mode CSS tokens, two
triggers (html.dark for @nuxt/color-mode hosts, prefers-color
-scheme for standalone SSR consumers).
- packages/ui/src/vue/use-dark-mode.ts: reactive dark-mode flag
mirroring the host page's color mode.
- packages/ui/src/core/types.ts: ElementDisplayIcon.urlDark so
callers can supply a dark-variant icon source.
- DtprIcon swaps src based on the resolved mode; tests cover
the resolution order and SSR fallback.
3. dtpr-ai docs polish
- dtpr-ai/app/pages/index.vue + content.config.ts: override
docus's default landing template so the homepage is wrapped
in the same UContainer max-width as the rest of the site.
- Taxonomy pages and Playground pick up dark-mode-aware icon
URLs from the new @dtpr/ui surface.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix: unblock dtpr-ai Cloudflare build
Three fixes to make the previous commit deploy.
1. Drop content.config.ts + app/pages/index.vue
c12 loads our user-layer content.config.ts via jiti, which
transitively requires @nuxt/content -> @nuxt/kit. jiti evals
@nuxt/kit/dist/index.mjs:182's `import.meta.dev` via new
vm.Script (CJS context), throwing SyntaxError. Docus's own
content.config.ts has the same imports but loads inside a Nuxt
context where jiti is configured differently.
The homepage UContainer override is a polish item; restoring it
needs a docus-friendly extension that doesn't trigger this load
chain. Following up separately.
2. Update @dtpr/ui test fixtures for v2 schema reshape
The v2 reshape (#274) made `actions` and `subchains` required
defaults via z.ZodDefault, which Zod 4 surfaces as required in
the inferred output type. vite-plugin-dts type-checks tests
during declaration generation, so missing fields blocked dts
emission. Add `actions: []` to InstanceElement fixtures, plus
`subchains: []`, `subchain_instances: []`, `sources: []`,
`linked_instance_ids: []` to DatachainType / DatachainInstance.
3. DtprIcon: reset failed state when isDark changes
Watch already reset `failed.value` when src/darkSrc changed.
isDark wasn't in the deps, so a dark-mode toggle could leave a
stale 404 marker, blocking retry against the new url.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
pichot
added a commit
that referenced
this pull request
May 6, 2026
…families, semantic input/output + pii ramp (#278) * feat(dtpr-ai): functional_modes — six modes (Focus 1) Authors the six functional-mode elements that PR #274 (7c9b930) deliberately left empty after renaming ai__decision → functional_modes. Each element carries a verb-forward title (Analytical — decides), a plain-language description with examples, and a citation to Narain Jashanmal's *AI Taxonomy* (v1.1, January 2026) — the source DTPR adopted by name for the six-mode partition. Elements added (category_id: functional_modes, all en/es/fr/pt; km + tl deferred to a translator pass): - analytical_mode — decides - semantic_mode — understands and remembers - generative_mode — creates - agentic_mode — acts - perceptive_mode — senses - physical_mode — moves Each element references a first-draft symbol stub at api/schemas/ai/2026-04-27-beta/symbols/mode_<verb>.svg — minimal verb-glyph geometry pending a designer pass for sign-scale legibility. Source captured durably at plugin/dtpr/research/2026-05-06T1443-narain-jashanmal-ai-taxonomy.md, with boundary cues (perceptive↔analytical input shape, semantic↔generative new-content test, agentic↔physical action target, analytical↔semantic input modality) recorded so future dtpr-element-design / dtpr-comprehension-audit runs over functional_modes pick up the same partition. A multi-PR plan tracker lives at docs/plans/2026-05-05-002-feat-ai-2026-04-27-beta-element-authoring-plan.md and covers the four other categories PR #274 deferred (risks_mitigation AIAAIC expansion, accountable consolidation + logo, processing family typing, output_dataset first-pass authoring). Focus 1 is marked drafted in the tracker; Focus 2 (AIAAIC harms) is in flight in a parallel session and its files will land in a separate commit. schema:validate and schema:build pass; composed icons emit hexagon × symbol correctly under dist/.../icons/<mode>_mode/{default,dark}.svg. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(dtpr-ai): risks_mitigation — replace 6 mechanism elements with AIAAIC's 9 victim-oriented harms (Focus 2) Retire the cause-oriented elements (compromise_of_privacy, function_creep, opaque_decision_making, overreliance_automation_bias, system_drift, unequal_performance) and their symbols. Replace with the 9 victim-oriented harm types from the AIAAIC Collaborative Harms Taxonomy: autonomy_loss, physical_harm, psychological_harm, reputational_harm, financial_harm, civil_liberties_harm, societal_cultural_harm, political_economic_harm, environmental_harm. Each carries the AIAAIC citation per locale; CC BY-SA 4.0 flows from the citation field per dtpr-ai/content/9.attribution.md. Mechanism and outcome framings overlapped awkwardly (e.g. compromise_of_privacy collided with autonomy_loss + civil_liberties_harm), forcing authors into mechanism-vs-outcome coin flips for the same scenario. Beta is mid-flight — remap downstream datachain instances; 2026-04-16-beta keeps the old IDs. Symbols are placeholder geometry pending designer pass. Non-en locale strings drafted by implementer; translator review pending. Keeping functional_modes and risks_mitigation orthogonal — the mode↔harm cross-product belongs in the renderer, not in element context. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(dtpr-ai): processing — replace 7 technique elements with 12 family-typed (Focus 4) Retire the technique-flavored elements (llm, text-to-speech, sentiment-analysis, time-series-forecasting, recommendation-systems, optimization, privacy-preserving) and replace with a family-typed catalog where each element id IS a family: language_models, computer_vision, biometric_recognition, speech_audio, classification_prediction, affect_emotion_analysis, anomaly_detection, optimization, recommendation_ranking, search_retrieval, clustering_segmentation, privacy_transformation. Family encoding: bare slugs on Element.id — not Element.context override and not a category-level context axis. Context fits per-instance selection (PII on input_dataset); families are catalog-level facts that should not require the author to re-pick at instance time. Naming-prefix conventions were declined as fragile. Coverage rationale: 12 families span the public-space AI surface DTPR targets — chatbots/kiosks, occupancy and crowd vision, accessibility audio, demand and traffic forecasting, equipment and fraud monitoring, signal/transit optimization, wayfinding ranking, document QA, audience analytics, de-identified camera pipelines. Biometric Recognition and Affect & Emotion Analysis are split out from Computer Vision and Classification & Prediction respectively because both carry distinct regulatory weight (EU AI Act prohibitions, ICO guidance) and high public-trust signal that disclosure should surface independently. Symbols: ten reuse purpose-designed processing/decision-making/data-shape icons from the existing symbol set. Two (search_retrieval → connectivity, clustering_segmentation → social) borrow closest-fit existing icons and need designer follow-up before stable promotion. Comprehension check (rubric 2026-04-20): - Audience fit: pass — commuter-grade verbs throughout. - Plain-language: partial — RAG, k-anonymization, differential privacy un-glossed in search_retrieval and privacy_transformation; copy pass before stable. - Symbol legibility: partial — two reused symbols flagged for design. - Ambiguity flags: pass — explicit boundary callouts (computer_vision ↔ biometric_recognition, speech_audio ↔ biometric_recognition, clustering_segmentation flags inferred-vs-authored cluster labels). - Locale coverage: partial — all six locales drafted, technical anglicisms in Khmer and Tagalog need translator review. - Variable-substitution clarity: n/a — no variables. - Overlap and distinctness: pass. - Overall: partial — ship-blockers are the two symbol stand-ins and the copy pass for un-glossed jargon. `pnpm --filter ./api schema:validate ai@2026-04-27-beta` passes (11 categories, 88 elements). Earlier 2026-04-16-beta is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(dtpr-ai): input_dataset + output_dataset — replace 7 format elements with 11 semantic categories × 2 sides (Focus 5) Reframes both dataset categories around what the data is *about*, not how it's *encoded*. Eight independently-developed citizen-facing transparency frameworks (Apple Privacy Nutrition Labels, Google Play Data Safety, W3C DPV PD v2.3, GDPR Art. 9, EU AI Act Art. 3(1) + Annex III, TILT, DaPIS, original DTPR's own semantic three) all categorize this way; "binary" / "tabular" / "boolean" read as programmer documentation to a non-technical commuter. Captured in corpus at plugin/dtpr/research/2026-05-06T1515-semantic-data-categories-public-disclosure.md. 11 bidirectional categories with 22 element files (input_* + output_* sharing one symbol_id each): About a person / body / place / behaviour / measurement Sensitive personal information Operational data (school occupancy, routes, budgets, public records) A decision about you A recommendation or prediction Generated content A physical action PII context dropped from both categories — once "About a person", "About a body", and "Sensitive personal" are explicit elements, the context is double-bookkeeping. output_dataset description and prompt broadened beyond "data products" to cover decisions, content, and physical actions per EU AI Act Art. 3(1). Six new symbol stubs (about_a_body, about_behaviour, sensitive_personal, operational_data, generated_content, physical_action) are minimal geometric placeholders flagged for designer pass. Existing five symbol_ids reused (personal, spatial, values_time, dm_accept-or-deny, dm_priority-ranking). Authored across 4 locales (en/es/fr/pt) matching Focus 1's precedent; km/tl translation deferred to native-speaker pass. schema:validate ✓ 11 categories, 103 elements schema:build ✓ 400 dist files test ✓ 388 tests (340 workers + 48 cli) Beta-stage breaking change for any datachain instance pinned to ai@2026-04-27-beta and using old element ids (personal, tabular, binary, etc.); migration table in the corpus entry. In-policy per the plan's Out-of-scope section. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(dtpr-ai): restore pii context on input_dataset + output_dataset (Focus 5 amendment) The first Focus 5 cut dropped the `pii` context dimension on a "double-bookkeeping" argument that only held for one of four old values (`biometric` ↔ `*_about_a_body`). Restore the dimension as a pure identifiability ramp — orthogonal to the semantic element catalog — to bring back original DTPR's at-a-glance colour-band signal. PII context (both categories, six locales): - `de_identified` (#4A90D9 blue) — about people, link to who is broken - `pseudonymous` (#9575CD purple, new) — linked via opaque token, name not revealed - `identifiable` (#FFD700 yellow) — directly identified - absence of value carries the "no PII claim" meaning (matches Role on `accountable`) Old `none` and `biometric` retired. Schema validates: 11 categories, 103 elements, 466 dist files (up from 400 for new per-value icon variants). 388 tests pass. Drive-by typo fixes (longstanding): - `dm_anomay-detection` → `dm_anomaly-detection` (symbol + element symbol_id) - `processing_privacy-preservoing-transformation` → `…-preserving-transformation` Plan (`docs/plans/2026-05-05-002-…`): - Reflect refined PII context. - Drop Focus 3 (`accountable` consolidation + logo support) — not this version. - Drop km/tl translator-pass follow-ups (out of scope this version). - Correct Focus 4 status line (was `ai@2026-05-06-beta`; actually `ai@2026-04-27-beta`). Corpus (`plugin/dtpr/research/2026-05-06T1515-…`): - Replace "Why the PII context becomes redundant" with "Why the PII context stays — refined to a pure identifiability ramp," documenting per-value rationale. - Refresh `content_hash` to current schema build. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Element.contextoverride (full replace, not merge), nullablecoloron context values (null = tag rendering),DatachainType.subchains[], and per-instancesubchain_instances/subchain_instance_id/actions/sources/linked_instance_ids/instance_version/updated_at.ai@2026-04-27-betalands the structural reshape: bare-slug category ids (drop theai__prefix),decision→functional_modes(catalog emptied for the element-design pass), accountable shape fix (rounded-square→hexagon) + Role context, processing shape (hexagon→circle) + technical-detail framing, PII context on input_dataset/output_dataset (none/anonymized/identifiable/biometric), citizen-priority display order, andsubchains: [data_flow]on the datachain-type.Test plan
🤖 Generated with Claude Code