# Collective Intelligence Index maintenance The index is a static, bilingual editorial map of this site's research and OSS. It operates offline and does not execute or install any indexed software. The initial review is dated 2026-09-21; regeneration never advances review dates. ## Source authority and reviewed scope `works.html` JSON-LD remains the publication authority, including SoftwareSourceCode records. `scripts/generate_agent_catalogs.py` derives `research-catalog.json`, `oss-catalog.json`, `oss.html` and `agent-index.json`. Do not maintain a second title/author/DOI list. The bibliography is generated from selected scholarly records only. OASG, CCR, PIC and CPCF software DOI records remain software. The initial screening covers 233 research records and 54 public repository metadata records. Selection contains 19 papers (8 core, 11 supporting) and initially 12 repositories (6 core, 6 supporting). The JSON registry's `corpus_audit` records every screened identity, depth, decision and reason. Keyword matches outside the selected full-text review remain unresolved rather than rejected on age. This is not a comprehensive survey of world research, or a full-text review of all 233 records. Adjacent public repositories were considered through their catalogue descriptions/topics and release observations; only the selected 12 received the stated README/contract/interface/license review. Included paper titles and DOI kinds were checked against DataCite registration metadata, and scope was inspected in the public `paper-tex-backup` archive. The registry stores immutable archive paths, hashes, locators and observation dates, not copies of manuscripts. Hashes bind source bytes, not mathematical truth. CGT's concept/series DOI has multiple versions and different catalogue and current registration dates; its exact TeX/version binding remains explicit as unresolved. Distinct CGT supplements must not be invented as separate DOI records or collapsed into a single claimed implementation. No exact Boundary Exchange title/DOI was resolved in the scanned catalogue. The original website license remains CC BY 4.0. The 18 inspected OSS licenses are Apache-2.0. The Audit-Closed repository separately licenses its bundled protocol paper CC BY 4.0. Repository metadata for other projects is not overwritten with a blanket license assertion. ## Rebuild and validate ```text python -m pip install -r requirements-index.txt python scripts/generate_agent_catalogs.py --offline python scripts/generate_paper_pages.py python scripts/generate_collective_intelligence_index.py python scripts/generate_collective_intelligence_index.py --check python scripts/validate_collective_intelligence_index.py python scripts/validate_paper_pages.py python scripts/audit_collective_intelligence_coverage.py --check python -m unittest discover -s tests -v ``` The new generator and `--check` never use the network. Paper generation reuses the checked-in PDF metadata cache. New paper records without cached metadata may need the existing explicit PDF refresh workflow. `--check` compares bytes without writing. JSON Schema Draft 2020-12 and semantic reference validation are complementary. The registry digest covers the entire JSON value except `content_digest`, serialized with sorted keys, compact separators, unescaped Unicode and no NaN. The schema is separately served and versioned at https://kadubon.github.io/github.io/schemas/collective-intelligence-index.schema.json. All representations use one normalized model. Both HTML pages contain complete cards, relationships, limits and links without JavaScript. Stable ASCII IDs are independent of titles and versions. Do not reuse a retired ID. ## Editorial changes 1. Add or correct a paper in the existing `works.html` source, regenerate its catalogue and landing page, then add its existing source ID to `data/collective-intelligence-curation.json`. Verify exact DOI identity, type, title, author and dates; distinguish concept and version DOIs. 2. Read the actual scope/limitations in the paper. Record a bounded evidence entry with URL, immutable revision, locator, hash and observation date in `data/collective-intelligence-evidence.json`. Add bilingual question, contribution, relevance and limitations. Mark any unavailable source or ambiguous version explicitly; do not infer peer review from a DOI. 3. Classify software by its actual repository catalogue ID. Read README, license, public interface, schemas, examples, tests and release documents. Record source revision/version separately from release observation. An unknown package location is not an invitation to invent a pip command. 4. Add a problem ID with symptoms, first reads, inputs, outputs, unsupported conditions and stop/handoff conditions. Roles are local editorial classifications, not an upstream standard or privileged agent instruction. 5. Record relationships with direction, exact version pair, checked fields, evidence origin and unsupported obligations. `imports_artifact_from` goes consumer → producer; `exports_proposal_to` goes producer → host. `based_on` is a citation, not complete implementation. 6. `tested_upstream` requires a cited native check report and scope. It is not a test performed by this website task. `proposed_integration_with` must remain `proposed`. Source inspection cannot become native test execution. Compatibility never propagates transitively or to a new version pair. 7. Advance review dates only after an actual review. Regenerate, run tests, inspect both rendered pages, review the diff and use a normal pull request. ## Factual refresh versus scientific review ```text python scripts/generate_agent_catalogs.py python scripts/refresh_oss_release_snapshots.py python scripts/generate_agent_catalogs.py --offline python scripts/generate_collective_intelligence_index.py ``` The explicit refresh uses bounded `gh api` requests (four workers, three attempts, timeout and backoff). It stores latest published GitHub release metadata and current default-branch commits for all catalogue repositories. A 404 on the latest-release endpoint means no published GitHub release was observed, not no installable source. Temporary failures preserve the previous record and mark it stale. GitHub release status does not verify PyPI, npm, installed bytes, CI, or production readiness. Reviewed source versions remain fixed in the evidence layer; new source revisions get a not-editorially-reviewed status and do not renew relationships. Periodic factual refresh does not auto-select new papers or perform scientific editorial review. `works.html` software entries display latest observed release versions while retaining original publication dates and historical archive DOI links. Updating those entries is an editorial change: do not relabel an old immutable archive DOI as the new release. The catalogue generator preserves software version, license and source-repository metadata from that source. ## Publication, concurrency and crawler scope The existing main-branch GitHub Pages publication is retained. `.nojekyll` makes this already static site serve Markdown as files, avoiding automatic Markdown conversion. There are no Jekyll layouts, includes or configuration in the inspected baseline. The explicit existing HTML pages remain the reader destinations. No new hosting service, release, version bump or endpoint exists. Both trusted derived-content writers share `derived-content-writers` and check out the current `main` when starting, rather than a queued event's old commit, so a preceding writer's generated updates are retained. They rebuild after rebasing before pushing; read-only pull-request CI has only contents-read permission and does not use `pull_request_target`. Existing paper generation preserves non-paper sitemap metadata. The feed still represents publications, not an index page presented as a new paper. The existing IndexNow workflow sees changed HTML/sitemap URLs; submission is not evidence of indexing. After normal-policy merge, check the exact Pages build/deployment commit and unauthenticated public status, MIME type, UTF-8 bytes and content digest for HTML, JSON, schema, Markdown and bibliography. Check homepage, OSS page, agent-index and sitemap discovery. The public verification report is retained as an execution artifact, not fabricated in advance by this document. Origin-wide robots policy belongs at https://kadubon.github.io/robots.txt. The file at https://kadubon.github.io/github.io/robots.txt is project content; it cannot set origin-wide crawler rules. The account-root repository is out of scope. Search indexing, retrieval, training and actual adoption are separate outcomes. Optional llms files do not grant access policy or guarantee use. Engineering references: [Schema.org SoftwareSourceCode](https://schema.org/SoftwareSourceCode), [Google AI feature guidance](https://developers.google.com/search/docs/appearance/ai-features), [robots location](https://developers.google.com/crawling/docs/robots-txt/create-robots-txt), and the optional [llms.txt proposal](https://llmstxt.org/). ## Symptom routing (schema 1.1, 2026-09-24) The current selection has 19 papers and 18 OSS resources. Six additional supporting repositories received a bounded source review on 2026-09-24; [the generated coverage report](collective-intelligence-coverage.md) records exact revisions, declared source versions, selection reasons and limitations. The earlier evidence dates remain unchanged. `last_reviewed_at` is derived per resource from its cited evidence; a routing edit does not renew source review. License observation dates likewise come from the cited license evidence. No additional paper or manually maintained bibliography was introduced. The existing 12 problem IDs remain stable. Schema 1.1 adds bilingual `query_aliases`, `symptom_group`, `related_problem_ids`, derived `relevant_resource_ids`, root `symptom_groups` and `unresolved_intents`. Consumers pinned to the 1.0 schema must update their schema before accepting 1.1 records. There are currently 27 routes in eight symptom groups, each with 1–3 first reads, 2–4 related problems, inputs, outputs, unsupported conditions and stop/handoff conditions. Related links can form cycles; they are not an execution graph and the router never recursively expands them. `route(model, query)` normalizes NFKC, case, curly quotes, whitespace and trailing question/exclamation marks, then matches an exact ID, legacy query, bilingual question, symptom or declared alias. It performs no fuzzy or substring inference. Unmatched queries return `None` (JSON null). Ambiguous cross-route aliases are errors. Neither matches nor related links confer authority, establish truth, settlement, interoperability, production readiness or AGI/ASI detection. Operational vocabulary belongs in `data/collective-intelligence-curation.json`. The A–O acceptance fixture in `data/collective-intelligence-coverage-seeds.json` contains 168 bilingual seed pairs and frozen measurements against the recorded baseline commit. It is a diagnostic, not a second operational routing table. After intentional vocabulary/fixture changes, regenerate the report with: ```text python scripts/audit_collective_intelligence_coverage.py --write python scripts/audit_collective_intelligence_coverage.py --check ``` Both commands are offline and deterministic. The current result is 163 matches and five explicit gaps per language: OAuth for agents, MCP token passthrough, package hallucination, slopsquatting, and installation of the wrong package. Do not fill these gaps by implying that an adjacent evidence ledger implements OAuth or package-supply-chain defenses. Search intent coverage does not measure ranking, general language understanding, crawler adoption or empirical acceleration. For UI review, inspect both languages at desktop and narrow widths, including symptom links, related links, native disclosures, long source revisions and downloads. Essential content is HTML/Markdown; no JavaScript is required.