@agentproto/corpus-presets 0.1.0 → 0.2.0

package/dist/research/index.d.ts

@@ -0,0 +1,31 @@
+import { CorpusPreset } from '@agentproto/corpus';
+
+/**
+ * AUTO-GENERATED build artifact — DO NOT COMMIT, DO NOT EDIT.
+ *
+ * Regenerated from @agentproto/corpus/test/fixtures/research/
+ * by the prebuild / precheck-types / pretest / predev hooks on
+ * this package. To change the preset content, edit the fixtures.
+ */
+declare const RESEARCH_PRESET_FILES: Readonly<Record<string, string>>;
+
+/**
+ * Research corpus preset.
+ *
+ * Pure TS data — `files` is inlined from the M0 fixture workspace at
+ * build time via `scripts/gen-research.mjs`. The M0 conformance test
+ * proves every file validates against the actual AgentProto JSON
+ * Schemas, so a host that writes these to disk gets a known-good
+ * starter workspace.
+ *
+ * Usage:
+ *
+ *     import { ResearchCorpusPreset } from "@agentproto/corpus-presets/research"
+ *     for (const [rel, content] of Object.entries(ResearchCorpusPreset.files)) {
+ *       await fs.writeFile(path.join(workspaceRoot, rel), content)
+ *     }
+ */
+
+declare const ResearchCorpusPreset: CorpusPreset;
+
+export { RESEARCH_PRESET_FILES, ResearchCorpusPreset };

package/dist/research/index.mjs

@@ -0,0 +1,85 @@
+/**
+ * @agentproto/corpus-presets v0.1.0-alpha
+ * Pure-TS starter workspaces for the AIP-10 corpus runtime.
+ */
+
+// src/research/files.ts
+var RESEARCH_PRESET_FILES = Object.freeze({
+  "collections/corpus-candidate/COLLECTION.md": '---\nschema: collection.schema/v1\nname: corpus-candidate\ntitle: Corpus Candidate\ndescription: |\n  Candidate corpus item awaiting analysis and promotion to an AIP-10 entry.\n  Discovered candidates live in `_candidates.yaml` sidecar; materialized\n  ITEM.md files appear only after transition out of `discovered`.\nversion: 1.0.0\ninitialStatus: discovered\nstatuses:\n  - { id: discovered, label: Discovered, transitionsTo: [analyzed, rejected] }\n  - { id: analyzed, label: Analyzed, transitionsTo: [approved, rejected, needs-work] }\n  - { id: needs-work, label: Needs Work, transitionsTo: [analyzed, rejected] }\n  - { id: approved, label: Approved, terminal: true }\n  - { id: rejected, label: Rejected, terminal: true }\nidentity:\n  slugSource: id\n  filingPath: "collections/corpus-candidate/{slug}/ITEM.md"\nfields:\n  - { name: sourcePath, type: string, required: true }\n  - { name: sourceUrl, type: url }\n  - { name: targetEntryPath, type: string }\n  - { name: corpusKind, type: enum, enum: [principle, example, pattern, critique, summary, timeline], required: true }\n  - { name: qualityScore, type: number }\n  - { name: riskScore, type: number }\n  - { name: promotionMode, type: enum, enum: [auto, human] }\n  - { name: contentHash, type: string }\n  - { name: reviewerNotes, type: text }\nlints:\n  - { id: required-field-source-path, kind: required-field, appliesTo: "*", severity: error, params: { fields: [sourcePath] } }\n  - { id: required-field-corpus-kind, kind: required-field, appliesTo: "*", severity: error, params: { fields: [corpusKind] } }\nmetadata:\n  corpus:\n    description: Used by source-scout (writes), research-analyst (writes), quality-reviewer (scores), corpus-curator (promotes).\n---\n\n# Corpus Candidate collection\n\nHolds candidates discovered by the scout, analyzed by the analyst, scored by the reviewer, and promoted by the curator. State machine enforced by `statuses[].transitionsTo`. Domain-agnostic \u2014 the `corpusKind` field drives which entry type the analyst targets.\n',
+  "collections/corpus-candidate/nature-review-2026-06/ITEM.md": '---\nschema: collection.item/v1\ncollection: corpus-candidate\nid: nature-review-2026-06\ntitle: Replication crisis systematic review \u2014 pattern extraction\nstatus: analyzed\nsourcePath: sources/fresh/nature-review-2026-06.md\nsourceUrl: https://example.com/nature/reviews/replication-crisis-2026\ntargetEntryPath: entries/patterns/2026/systematic-review-pattern.md\ncorpusKind: pattern\ncontentHash: sha256:b2e4f8a1c3d5e7f9a1b3c5d7e9f1a3b5c7d9e1f3a5b7c9d1e3f5a7b9c1d3e5f7\ntags: [methodology, replication, systematic-review]\ncreatedAt: "2026-06-10T09:00:00Z"\nupdatedAt: "2026-06-10T10:30:00Z"\nmetadata:\n  corpus:\n    qualityScore: 4.5\n    riskScore: 0.8\n    promotionMode: auto\n---\n\n# Candidate: Replication crisis systematic review\n\nCandidate extracted from a peer-reviewed Nature Reviews article. The pattern (hierarchical Bayesian meta-analysis for heterogeneous replication studies) generalizes across empirical research domains. Analyst recommends auto-promotion.\n\n## Analysis\n- Primary-authority source (peer-reviewed journal, pre-registered)\n- Methodology (hierarchical Bayesian meta-analysis) is domain-portable\n- No legal or intellectual-property risk identified\n- 847-study sample provides robust effect-size estimates\n\n## Decision\n- promotionMode: auto\n- gates pass: qualityScore 4.5 \u2265 4.2, riskScore 0.8 \u2264 1.5, archive hash present, all required fields populated\n',
+  "entries/critiques/2026/selection-bias-failure.md": '---\nschema: knowledge.entry/v1\nslug: selection-bias-failure\nkind: critique\ntitle: Selection bias as a driver of non-replication\nupdated_at: "2026-06-10T10:30:00Z"\nsources:\n  - nature-review-2026-06\nconfidence: 0.90\ntags: [methodology, bias, replication, failure-mode]\nmetadata:\n  corpus:\n    status: active\n    qualityScore: 4.4\n    riskScore: 0.5\n    domain: research\n    temporal:\n      firstSeen: "2011-01-01T00:00:00Z"\n      lastSeen: "2026-06-10T10:30:00Z"\n      mentions:\n        - { at: "2011-01-01T00:00:00Z", sourceId: nature-review-2026-06, weight: 1.0 }\n        - { at: "2026-06-10T10:30:00Z", sourceId: nature-review-2026-06, weight: 1.0 }\n    promotionMode: human\n    promotedAt: "2026-06-10T10:30:00Z"\n    promotedBy: corpus-curator\n---\n\n## Summary\nStudies that select their sample on the outcome variable or restrict inclusion to high-effect subgroups produce inflated effect sizes that fail to replicate in unselected populations.\n\n## Why it fails\nWhen inclusion criteria correlate with the outcome, the sample is not representative of the target population. The estimated effect is an artifact of the selection process, not a property of the population. Replication studies that sample more broadly will find smaller or null effects.\n\n## Failure modes\n- **Outcome-dependent sampling**: recruiting only people who experienced the effect being studied.\n- **Subgroup cherry-picking**: reporting the subgroup with the largest effect after observing the data.\n- **Survivor bias in longitudinal studies**: only measuring participants who completed the study.\n- **Publication bias at meta-analysis level**: including only published (positive) studies.\n\n## Use when\n- Evaluating a new source for corpus inclusion: check whether sample selection is described.\n- Scoring riskScore for an entry: unexplained heterogeneity across replications raises riskScore.\n- Writing critique entries that expose methodological limitations in existing literature.\n\n## Avoid when\n- The selection IS the design (e.g., case studies of rare diseases where selection is inherent).\n',
+  "entries/patterns/2026/systematic-review-pattern.md": '---\nschema: knowledge.entry/v1\nslug: systematic-review-pattern\nkind: pattern\ntitle: Systematic review with hierarchical Bayesian meta-analysis\nupdated_at: "2026-06-10T10:30:00Z"\nsources:\n  - nature-review-2026-06\nconfidence: 0.88\ntags: [methodology, meta-analysis, bayesian, replication]\nmetadata:\n  corpus:\n    status: active\n    qualityScore: 4.5\n    riskScore: 0.6\n    domain: research\n    temporal:\n      firstSeen: "2018-01-01T00:00:00Z"\n      lastSeen: "2026-06-10T10:30:00Z"\n      mentions:\n        - { at: "2018-01-01T00:00:00Z", sourceId: nature-review-2026-06, weight: 1.0 }\n        - { at: "2026-06-10T10:30:00Z", sourceId: nature-review-2026-06, weight: 1.0 }\n    promotionMode: auto\n    promotedAt: "2026-06-10T10:30:00Z"\n    promotedBy: corpus-curator\n---\n\n## Summary\nSystematic reviews achieve reliable effect-size estimates when they: (1) pre-register inclusion criteria, (2) apply hierarchical Bayesian pooling to account for between-study heterogeneity, and (3) report credible intervals rather than p-values.\n\n## Why it works\nHierarchical (random-effects) Bayesian models shrink extreme estimates toward the population mean \u2014 the replication crisis showed that underpowered studies with extreme point estimates drive non-replication. Pre-registration eliminates post-hoc analysis flexibility.\n\n## Transferable pattern\n1. Pre-register: hypothesis, inclusion/exclusion criteria, analysis plan.\n2. Screen independently: two reviewers screen abstracts; resolve conflicts by consensus.\n3. Pool with hierarchical Bayes: \u03C4\xB2 estimates between-study variance; shrink estimates accordingly.\n4. Report 95 % credible intervals, not just posterior means.\n5. Conduct sensitivity analysis: exclude low-quality studies, check robustness.\n\n## Use when\n- Synthesizing a body of empirical literature with heterogeneous study designs.\n- The corpus needs a high-confidence summary of a contested empirical question.\n- Effect-size estimation matters, not just directional conclusions.\n\n## Avoid when\n- Fewer than 10 primary studies exist (insufficient for pooling).\n- The research question is qualitative or interpretive rather than quantitative.\n',
+  "entries/principles/2026/primary-sources-beat-synthesis.md": '---\nschema: knowledge.entry/v1\nslug: primary-sources-beat-synthesis\nkind: principle\ntitle: Primary sources beat synthesis\nupdated_at: "2026-06-10T10:30:00Z"\nsources:\n  - nature-review-2026-06\nconfidence: 0.92\ntags: [methodology, evidence, authority, sourcing]\nmetadata:\n  corpus:\n    status: active\n    qualityScore: 4.7\n    riskScore: 0.4\n    domain: research\n    temporal:\n      firstSeen: "2015-01-01T00:00:00Z"\n      lastSeen: "2026-06-10T10:30:00Z"\n      mentions:\n        - { at: "2015-01-01T00:00:00Z", sourceId: nature-review-2026-06, weight: 1.0 }\n        - { at: "2026-06-10T10:30:00Z", sourceId: nature-review-2026-06, weight: 1.0 }\n    promotionMode: human\n    promotedAt: "2026-06-10T10:30:00Z"\n    promotedBy: corpus-curator\n---\n\n## Summary\nWhen a synthesis contradicts its primary sources, trust the primary sources. Systematic reviews compress information; compression loses detail that sometimes reverses the conclusion.\n\n## Why it works\nSynthesis introduces two layers of error: sampling bias in study selection and aggregation error in effect-size combination. Primary data carries neither of these artefacts by construction.\n\n## Transferable pattern\n- Cite primary studies, not meta-analyses, when a claim is contested.\n- When a meta-analysis conflicts with a landmark primary study, read the primary study.\n- Record confidence lower when your only source is a secondary synthesis.\n\n## Use when\n- Evaluating evidentiary strength of a claim for corpus promotion.\n- Deciding authority tier for a new source (primary vs secondary vs rumour).\n- Writing confidence scores on entries derived from secondary sources.\n\n## Avoid when\n- The domain has no accessible primary literature (policy documents, grey literature).\n- Speed matters more than precision and a trusted synthesis is available.\n',
+  "KNOWLEDGE.md": '---\nschema: knowledge.workspace/v1\nname: research-corpus\ntitle: Research Expert Corpus\ndescription: |\n  Domain-agnostic research knowledge for autonomous agents. Composes AIP-10 (knowledge),\n  AIP-18 (curation collections), AIP-9 (operators), AIP-15 (workflows), and AIP-41\n  (routines) into an autonomous discovery-and-improvement loop.\nversion: 1.0.0\nentityTypes:\n  - { name: Principle, description: "Durable finding or research rule" }\n  - { name: Example, description: "Concrete source-derived example or case study" }\n  - { name: Pattern, description: "Reusable transferable methodology or mechanic" }\n  - { name: Critique, description: "Negative example, failure mode, or methodological flaw" }\n  - { name: Summary, description: "Compressed synthesis of a source or topic cluster" }\n  - { name: Timeline, description: "Time-sensitive trend or field evolution" }\nsources:\n  retention: forever\n  signing: optional\n  hashAlgo: sha256\n  authorityDefault: secondary\nlints:\n  - { id: require-source-on-examples, kind: require-source, appliesTo: Example, severity: error }\n  - { id: min-confidence-principles, kind: min-confidence, appliesTo: Principle, severity: warn, params: { min: 0.6 } }\n  - { id: max-age-timelines, kind: max-age, appliesTo: Timeline, severity: warn, params: { days: 730 } }\n  - { id: max-age-summaries, kind: max-age, appliesTo: Summary, severity: warn, params: { days: 180 } }\n  - { id: broken-ref-all, kind: broken-ref, appliesTo: "*", severity: error }\n  - { id: orphan-all, kind: orphan, appliesTo: "*", severity: info }\ncuration:\n  tone: "neutral, expert, source-backed"\n  depth: medium\n  autoLink: byName\n  conflictResolution: authority\nqueryHints:\n  preferRecent: true\n  preferAuthoritative: true\nmetadata:\n  corpus:\n    version: 1\n    domain: research\n    entryLayout:\n      kind: hierarchical\n      groupBy: entityType\n    autoPromote:\n      enabled: true\n      requires:\n        qualityScore: { min: 4.2 }\n        riskScore: { max: 1.5 }\n        hasArchiveHash: true\n        requiredFields: [why_it_works, transferable_pattern, use_when, avoid_when]\n        notRestricted: true\n    temporal:\n      defaultHalfLifeDays: 365\n      scoreFormula: max-mention-decay\n    retrievalBoosts:\n      qualityScore: { weight: 0.30 }\n      temporalScore: { weight: 0.20 }\n      mentionCount: { weight: 0.10 }\n      confidence: { weight: 0.15 }\n    retrievalDefaults:\n      status: active\n      minQualityScore: 3.5\n      minTemporalScore: 0.1\n    accessModes:\n      read: { allowedRoles: ["*"] }\n      cite: { allowedRoles: ["*"] }\n      flag-learning: { allowedRoles: ["*"], rateLimit: { perOperator: 20, window: "24h" } }\n      curate: { allowedRoles: [corpus-curator, admin] }\n      promote: { allowedRoles: [corpus-curator, admin] }\n      admin-reindex: { allowedRoles: [admin] }\n      bypass-default-filters: { allowedRoles: [corpus-curator, admin], audit: true }\n---\n\n# Research Corpus\n\nSource-of-truth for autonomous research agents. Composes AgentProto AIPs into a closed-loop knowledge-improvement system for any domain.\n\n## Three loops\n- **Knowledge curation** \u2014 source-scout \u2192 analyst \u2192 entries\n- **Quality gating** \u2014 quality-reviewer scores + routes (auto-promote vs human review)\n- **Retrieval-quality feedback** \u2014 every query logged; `utility` + `lift` written back to entries\n\nAll corpus-specific policy lives under `metadata.corpus.*`. The six entity types (Principle, Example, Pattern, Critique, Summary, Timeline) are domain-agnostic \u2014 topics come from sources, not from the scaffold.\n',
+  "operators/corpus-curator/OPERATOR.md": `---
+name: Corpus Curator
+id: corpus-curator
+persona_summary: Promotes approved candidates to AIP-10 entries, maintains _index.md, handles deprecation + supersession. The corpus's editor-in-chief.
+version: "1.0.0"
+profile:
+  role: |
+    Promote candidates that pass the auto-promote gate. Review the
+    human-review queue, decide approve/reject/needs-work. Keep the
+    corpus tidy: dedupe, deprecate stale entries, resolve
+    contradictions.
+  voice: |
+    Direct, decisive, opinionated about quality bar. First-person
+    plural for institutional voice. Cites specific entry slugs when
+    discussing deprecation or supersession.
+  boundaries:
+    - Never promote an entry that fails the auto-promote gate without explicit bypass.
+    - Never deprecate without a written reason in metadata.corpus.archiveReason.
+    - Always preserve the audit chain \u2014 _log.md events are append-only.
+governance:
+  audit_log: "audit:corpus/research/_log.md"
+  autonomy: autonomous
+  policies:
+    - "policy:corpus/research/curation"
+skills:
+  - source-analysis
+  - taxonomy
+  - editorial-judgment
+tools:
+  - corpus-promote-candidate
+  - corpus-log-event
+  - knowledge-query
+participation:
+  mode: proactive
+memory:
+  kind: operator-context
+runtime:
+  kind: in-process
+tags: [research, curator, corpus]
+metadata:
+  corpus:
+    domain: research
+    overlays:
+      policy: open
+      maxActiveCount: 10
+---
+
+# Corpus Curator
+
+Runs after the reviewer scores a candidate. Promotes auto-approved entries; queues human-review-required candidates for the team. Drives entry lifecycle (active \u2192 archived).
+`,
+  "operators/quality-reviewer/OPERATOR.md": '---\nname: Quality Reviewer\nid: quality-reviewer\npersona_summary: Scores analyzed candidates against the corpus\'s quality + risk rubric. Decides promotionMode (auto vs human). Flags concerns the analyst missed.\nversion: "1.0.0"\nprofile:\n  role: |\n    Independent second pair of eyes on every analyzed candidate.\n    Adversarial-by-design \u2014 assumes the analyst was too generous\n    and looks for reasons NOT to promote.\n  voice: |\n    Critical, concrete, fair. Cites specific clauses of the\n    candidate body, not vibes. Always proposes a concrete score\n    (qualityScore 0-5, riskScore 0-5).\n  boundaries:\n    - Never promote on the analyst\'s authority \u2014 score and route.\n    - Never auto-promote a candidate flagged as restricted / personal / legal.\n    - Always justify a sub-2.0 quality score in reviewerNotes.\ngovernance:\n  audit_log: "audit:corpus/research/_log.md"\n  autonomy: supervised\n  policies:\n    - "policy:corpus/research/review-policy"\nskills:\n  - source-analysis\n  - critical-reading\ntools:\n  - corpus-flag-learning\n  - knowledge-query\nparticipation:\n  mode: proactive\nmemory:\n  kind: operator-context\nruntime:\n  kind: in-process\ntags: [research, reviewer, corpus]\nmetadata:\n  corpus:\n    domain: research\n---\n\n# Quality Reviewer\n\nRuns after `research-analyst` produces an analyzed ITEM.md. Scores against:\n\n- **qualityScore** \u2014 accuracy, transferability, evidence strength, authority tier\n- **riskScore** \u2014 legal/IP/compliance hazards, retracted studies, contested findings\n- **promotionMode** \u2014 auto (gates pass) vs human (needs curator)\n',
+  "operators/research-analyst/OPERATOR.md": '---\nname: Research Analyst\nid: research-analyst\npersona_summary: Domain-agnostic research analyst who converts raw sources into expert principles, patterns, and critiques for the research corpus.\nversion: "1.0.0"\nprofile:\n  role: |\n    Convert candidate sources discovered by the scout into corpus-grade\n    expert analysis: principles, patterns, critiques. Score quality and\n    risk to inform promotion gates. Works across any research domain \u2014\n    the topic comes from the source, not from the operator persona.\n  voice: |\n    Concrete, source-backed, never superlative. Cite specific numbers and\n    named authors or publications. First-person plural ("we") for\n    institutional voice.\n  boundaries:\n    - Never invent quantitative claims; cite a source or mark as opinion.\n    - Never promote without quality-reviewer signoff (unless auto-promote gates pass).\n    - Never edit raw sources \u2014 only entries.\ngovernance:\n  audit_log: "audit:corpus/research/_log.md"\n  autonomy: supervised\n  policies:\n    - "policy:corpus/research/curation"\nskills:\n  - source-analysis\n  - research-synthesis\ntools:\n  - corpus-flag-learning\n  - knowledge-query\nparticipation:\n  mode: proactive\nmemory:\n  kind: operator-context\nruntime:\n  kind: in-process\ntags: [research, analyst, corpus]\nmetadata:\n  corpus:\n    domain: research\n    knowledgeViews:\n      - corpus: research\n        filter:\n          domain: [research]\n          minQualityScore: 3.5\n        defaultBoosts:\n          qualityScore: { weight: 0.40 }\n          temporalScore: { weight: 0.30 }\n    overlays:\n      policy: gated\n      maxActiveCount: 5\n      requireApprovalFrom: corpus-curator\n      allowedKinds: [overlay, block-replacement]\n---\n\n# Research Analyst\n\nReads candidate sources from the scout, writes expert analysis as AIP-10 entries. Scores quality and risk. Signals the curator on auto-promote gate pass.\n',
+  "operators/source-scout/OPERATOR.md": '---\nname: Source Scout\nid: source-scout\npersona_summary: Curates external feeds for fresh research source material \u2014 academic papers, articles, videos, social posts, and primary documents \u2014 and archives the worth-keeping ones into sources/.\nversion: "1.0.0"\nprofile:\n  role: |\n    Discover new candidate sources for the research corpus. Filter\n    noise vs signal at the scout stage so the analyst only sees\n    promising material. Source categories: article, video, social,\n    paper, primary, news.\n  voice: |\n    Efficient, terse, evidence-first. Names URLs + numbers, not\n    adjectives. Reports findings as `{source: <id>, why: <reason>}`\n    triples.\n  boundaries:\n    - Never re-archive an already-archived source (check content_hash).\n    - Never archive content with PII unless explicitly sanitized.\n    - Never archive paywalled / DMCA-restricted material without authorization.\ngovernance:\n  audit_log: "audit:corpus/research/_log.md"\n  autonomy: autonomous\n  policies:\n    - "policy:corpus/research/source-policy"\nskills:\n  - web-research\n  - source-analysis\ntools:\n  - corpus-create-candidate\n  - knowledge-query\n  - web-search\nparticipation:\n  mode: proactive\nmemory:\n  kind: operator-context\nruntime:\n  kind: in-process\ntags: [research, scout, corpus]\nmetadata:\n  corpus:\n    domain: research\n    knowledgeViews:\n      - corpus: research\n        filter:\n          domain: [research]\n---\n\n# Source Scout\n\nRuns on-demand or on schedule (via `source-scout` routine). Fetches configured feeds and channels, hashes content, dedups against existing sources, and writes new candidates to `_candidates.yaml`. The analyst takes over from there.\n\nSupported source categories: `article` (blog/news), `video` (YouTube, conference talks), `social` (threads, posts), `paper` (peer-reviewed), `primary` (official documents, datasets), `news` (breaking coverage).\n',
+  "routines/source-scout/ROUTINE.md": '---\nschema: routine/v1\nid: source-scout\ndescription: |\n  On-demand (or scheduled) routine that runs the source-scout operator to\n  discover new candidate sources across configured external feeds and channels.\n  Discovered sources are archived under sources/ and rows appended to\n  _candidates.yaml.\nversion: "1.0.0"\nschedule:\n  kind: cron\n  cron: "0 8 * * 1"\n  timezone: "UTC"\n  catchup: skip\ntarget:\n  workflow: ingest-source\n  inputs:\n    feeds: [web-search, youtube, arxiv, news-api]\n    maxItemsPerRun: 30\nretry:\n  max_attempts: 3\n  backoff: exponential\non_failure:\n  create_work_item: true\n  fire_event: corpus.scout.failed\nfires_events:\n  - corpus.scout.completed\n  - corpus.candidate.discovered\nenabled: true\ntags: [corpus, scout, weekly]\nmetadata:\n  corpus:\n    domain: research\n---\n\n# Source Scout Routine\n\nRuns weekly (Monday 08:00 UTC) by default; can also be triggered on-demand. Scouts external feeds for new candidate sources, archives them, and appends to `_candidates.yaml`. The candidate event downstream triggers `analyze-candidate` workflow.\n\nAdjust the `schedule` and `target.inputs.feeds` list to match your research domain\'s primary discovery channels.\n',
+  "sources/fresh/nature-review-2026-06.md": '---\nschema: knowledge.source/v1\nid: nature-review-2026-06\npath: sources/fresh/nature-review-2026-06.md\ntitle: Systematic review of replication failures in social science \u2014 June 2026\ncaptured_at: "2026-06-10T09:00:00Z"\ncaptured_by: source-scout\ncontent_hash: sha256:b2e4f8a1c3d5e7f9a1b3c5d7e9f1a3b5c7d9e1f3a5b7c9d1e3f5a7b9c1d3e5f7\nauthority: primary\nlanguage: en\ntags: [systematic-review, replication, methodology, social-science]\nmetadata:\n  corpus:\n    originalUrl: https://example.com/nature/reviews/replication-crisis-2026\n    archiveTier: hot\n    archivedBy: source-scout\n    snapshotPath: sources/fresh/nature-review-2026-06.snapshot.mhtml\n---\n\n# Systematic review: replication failures in social science (2026)\n\nCaptured from a peer-reviewed Nature Reviews article in June 2026. Meta-analysis of 847 attempted replications across psychology, economics, and political science (2010\u20132026).\n\n## Key findings\n- Only 39 % of original findings replicated at p < 0.05 in pre-registered replications\n- Effect sizes in successful replications averaged 68 % of originals\n- Primary driver of non-replication: underpowered original studies, not fraud\n\n## Methodology note\nAuthors applied a hierarchical Bayesian model to account for heterogeneity across domains. Pre-registered analysis plan at OSF (osf.io/example123).\n\n## Relevance\nFoundational reference for any research corpus reasoning about evidentiary quality, source authority, and confidence thresholds.\n',
+  "workflows/analyze-candidate/WORKFLOW.md": '---\nname: Analyze Corpus Candidate\nid: analyze-candidate\ndescription: |\n  Convert a discovered corpus candidate into an analyzed ITEM.md with\n  expert analysis, quality score, and risk score. Triggered by the\n  source-scout when it appends a row to _candidates.yaml.\nversion: "1.0.0"\ninputs:\n  type: object\n  required: [candidateId, sourcePath]\n  properties:\n    candidateId: { type: string, pattern: "^[a-z0-9][a-z0-9-]*$" }\n    sourcePath: { type: string }\noutputs:\n  type: object\n  required: [itemPath, status]\n  properties:\n    itemPath: { type: string }\n    status: { type: string, enum: [analyzed, rejected, needs-work] }\n    qualityScore: { type: number }\n    riskScore: { type: number }\nsteps:\n  - id: load-source\n    kind: tool\n    tool: knowledge.read_source\n    name: Load source bytes from sources/\n    inputs: { type: object, properties: { sourceId: { type: string } } }\n    next: analyze\n  - id: analyze\n    kind: tool\n    tool: research-analyst.analyze\n    name: Run research-analyst operator over the source\n    next: score\n  - id: score\n    kind: tool\n    tool: quality-reviewer.score\n    name: Compute quality + risk scores\n    next: gate\n  - id: gate\n    kind: branch\n    name: Decide promotionMode (auto vs human)\n    description: Check auto-promote gate against KNOWLEDGE.md.metadata.corpus.autoPromote\n    branches:\n      - { when: "output.qualityScore >= 4.2 && output.riskScore <= 1.5", next: "auto-promote" }\n      - { when: "true", next: "human-review" }\n  - id: auto-promote\n    kind: tool\n    tool: corpus-curator-promote\n    name: Promote to AIP-10 entry\n  - id: human-review\n    kind: approval\n    name: Queue for human curator\n    prompt: Review and approve or reject this candidate.\n    approvers:\n      - { role: corpus-curator }\ntags: [corpus, workflow, analysis]\nmetadata:\n  corpus:\n    domain: research\n    triggeredBy: source-scout\n---\n\n# Analyze Candidate workflow\n\nConverts a discovered candidate into an analyzed ITEM.md. The research-analyst operator runs the actual analysis; this workflow orchestrates the pipeline (load \u2192 analyze \u2192 score \u2192 gate).\n',
+  "workflows/ingest-source/WORKFLOW.md": '---\nname: Ingest source\nid: ingest-source\ndescription: |\n  Fetch + hash + dedup + archive a single source. Output: a new\n  candidate row in _candidates.yaml (or no-op if duplicate).\nversion: "1.0.0"\ninputs:\n  type: object\n  required: [sourceUrl]\n  properties:\n    sourceUrl: { type: string }\n    category: { type: string, enum: [article, video, social, paper, primary, news] }\noutputs:\n  type: object\n  required: [outcome]\n  properties:\n    outcome: { type: string, enum: [archived, duplicate, rejected] }\n    sourceId: { type: string }\n    candidateId: { type: string }\nsteps:\n  - id: fetch\n    kind: tool\n    tool: web-fetch\n    name: Fetch source content\n  - id: hash\n    kind: tool\n    tool: hash-content\n    name: Compute sha256 content_hash\n  - id: dedup\n    kind: branch\n    name: Dedup against existing sources\n    branches:\n      - { when: "output.duplicate === true", next: "done-dup" }\n      - { when: "true", next: "archive" }\n  - id: archive\n    kind: tool\n    tool: corpus-archive-source\n    name: Write source frontmatter + body to sources/<category>/<slug>.md\n  - id: create-candidate\n    kind: tool\n    tool: createCorpusCandidate\n    name: Append discovered row to _candidates.yaml\n  - id: done-dup\n    kind: tool\n    tool: corpus-log-event\n    name: Log dedup hit (no-op outcome)\ntags: [corpus, ingest, scout]\nmetadata:\n  corpus:\n    domain: research\n    triggeredBy: source-scout\n---\n\n# Ingest source workflow\n\nStep 1 of the corpus loop. Single-source entry point. Called by the `source-scout` routine once per discovered URL. The `category` input (article | video | social | paper | primary | news) drives the archive path under `sources/`.\n',
+  "workflows/promote-corpus-item/WORKFLOW.md": '---\nname: Promote corpus item\nid: promote-corpus-item\ndescription: |\n  Lift an approved candidate into a published AIP-10 entry. Writes\n  the entry file, regenerates _index.md, ships chunks to the\n  backing engine, appends a promotion event. Wraps the kit\'s\n  CorpusPromoter.promote().\nversion: "1.0.0"\ninputs:\n  type: object\n  required: [candidateId, entrySlug, entryKind, entryPath, frontmatter, body]\n  properties:\n    candidateId: { type: string }\n    entrySlug: { type: string }\n    entryKind: { type: string, enum: [principle, example, pattern, critique, summary, timeline] }\n    entryPath: { type: string }\n    frontmatter: { type: object }\n    body: { type: string }\n    bypassGate: { type: boolean }\noutputs:\n  type: object\n  required: [entryPath, gatePassed]\n  properties:\n    entryPath: { type: string }\n    versionToken: { type: string }\n    chunkCount: { type: integer }\n    gatePassed: { type: boolean }\n    bypassed: { type: boolean }\nsteps:\n  - id: promote\n    kind: tool\n    tool: promoteCorpusCandidate\n    name: Atomic promote (lock + write + index + chunks + event)\ntags: [corpus, promote]\nmetadata:\n  corpus:\n    domain: research\n    triggeredBy: corpus-curator\n---\n\n# Promote corpus item workflow\n\nStep 3 of the corpus loop. Single-step wrapper for callers that want the workflow shape; the underlying `CorpusPromoter` does the multi-file transaction.\n'
+});
+
+// src/research/index.ts
+var ResearchCorpusPreset = Object.freeze({
+  slug: "research",
+  title: "Research Expert Corpus",
+  description: "Domain-agnostic research knowledge \u2014 principles, patterns, critiques. Composes AIP-10/18/9/15/41 into a closed-loop discovery-and-improvement system.",
+  files: RESEARCH_PRESET_FILES
+});
+
+export { RESEARCH_PRESET_FILES, ResearchCorpusPreset };
+//# sourceMappingURL=index.mjs.map
+//# sourceMappingURL=index.mjs.map

package/dist/research/index.mjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/research/files.ts","../../src/research/index.ts"],"names":[],"mappings":";;;;;;AAQO,IAAM,qBAAA,GAA0D,OAAO,MAAA,CAAO;AAAA,EACnF,4CAAA,EAA8C,q/DAAA;AAAA,EAC9C,4DAAA,EAA8D,u4CAAA;AAAA,EAC9D,kDAAA,EAAoD,okEAAA;AAAA,EACpD,oDAAA,EAAsD,+pEAAA;AAAA,EACtD,2DAAA,EAA6D,8zDAAA;AAAA,EAC7D,cAAA,EAAgB,ojHAAA;AAAA,EAChB,sCAAA,EAAwC,CAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;AAAA;;AAAA;AAAA,CAAA;AAAA,EACxC,wCAAA,EAA0C,imDAAA;AAAA,EAC1C,wCAAA,EAA0C,o5DAAA;AAAA,EAC1C,oCAAA,EAAsC,u4DAAA;AAAA,EACtC,kCAAA,EAAoC,ksCAAA;AAAA,EACpC,wCAAA,EAA0C,89CAAA;AAAA,EAC1C,yCAAA,EAA2C,8nEAAA;AAAA,EAC3C,qCAAA,EAAuC,yrDAAA;AAAA,EACvC,2CAAA,EAA6C;AAC/C,CAAC;;;ACJM,IAAM,oBAAA,GAAqC,OAAO,MAAA,CAAO;AAAA,EAC9D,IAAA,EAAM,UAAA;AAAA,EACN,KAAA,EAAO,wBAAA;AAAA,EACP,WAAA,EACE,4JAAA;AAAA,EACF,KAAA,EAAO;AACT,CAAC","file":"index.mjs","sourcesContent":["/**\n * AUTO-GENERATED build artifact — DO NOT COMMIT, DO NOT EDIT.\n *\n * Regenerated from @agentproto/corpus/test/fixtures/research/\n * by the prebuild / precheck-types / pretest / predev hooks on\n * this package. To change the preset content, edit the fixtures.\n */\n\nexport const RESEARCH_PRESET_FILES: Readonly<Record<string, string>> = Object.freeze({\n  \"collections/corpus-candidate/COLLECTION.md\": \"---\\nschema: collection.schema/v1\\nname: corpus-candidate\\ntitle: Corpus Candidate\\ndescription: |\\n  Candidate corpus item awaiting analysis and promotion to an AIP-10 entry.\\n  Discovered candidates live in `_candidates.yaml` sidecar; materialized\\n  ITEM.md files appear only after transition out of `discovered`.\\nversion: 1.0.0\\ninitialStatus: discovered\\nstatuses:\\n  - { id: discovered, label: Discovered, transitionsTo: [analyzed, rejected] }\\n  - { id: analyzed, label: Analyzed, transitionsTo: [approved, rejected, needs-work] }\\n  - { id: needs-work, label: Needs Work, transitionsTo: [analyzed, rejected] }\\n  - { id: approved, label: Approved, terminal: true }\\n  - { id: rejected, label: Rejected, terminal: true }\\nidentity:\\n  slugSource: id\\n  filingPath: \\\"collections/corpus-candidate/{slug}/ITEM.md\\\"\\nfields:\\n  - { name: sourcePath, type: string, required: true }\\n  - { name: sourceUrl, type: url }\\n  - { name: targetEntryPath, type: string }\\n  - { name: corpusKind, type: enum, enum: [principle, example, pattern, critique, summary, timeline], required: true }\\n  - { name: qualityScore, type: number }\\n  - { name: riskScore, type: number }\\n  - { name: promotionMode, type: enum, enum: [auto, human] }\\n  - { name: contentHash, type: string }\\n  - { name: reviewerNotes, type: text }\\nlints:\\n  - { id: required-field-source-path, kind: required-field, appliesTo: \\\"*\\\", severity: error, params: { fields: [sourcePath] } }\\n  - { id: required-field-corpus-kind, kind: required-field, appliesTo: \\\"*\\\", severity: error, params: { fields: [corpusKind] } }\\nmetadata:\\n  corpus:\\n    description: Used by source-scout (writes), research-analyst (writes), quality-reviewer (scores), corpus-curator (promotes).\\n---\\n\\n# Corpus Candidate collection\\n\\nHolds candidates discovered by the scout, analyzed by the analyst, scored by the reviewer, and promoted by the curator. State machine enforced by `statuses[].transitionsTo`. Domain-agnostic — the `corpusKind` field drives which entry type the analyst targets.\\n\",\n  \"collections/corpus-candidate/nature-review-2026-06/ITEM.md\": \"---\\nschema: collection.item/v1\\ncollection: corpus-candidate\\nid: nature-review-2026-06\\ntitle: Replication crisis systematic review — pattern extraction\\nstatus: analyzed\\nsourcePath: sources/fresh/nature-review-2026-06.md\\nsourceUrl: https://example.com/nature/reviews/replication-crisis-2026\\ntargetEntryPath: entries/patterns/2026/systematic-review-pattern.md\\ncorpusKind: pattern\\ncontentHash: sha256:b2e4f8a1c3d5e7f9a1b3c5d7e9f1a3b5c7d9e1f3a5b7c9d1e3f5a7b9c1d3e5f7\\ntags: [methodology, replication, systematic-review]\\ncreatedAt: \\\"2026-06-10T09:00:00Z\\\"\\nupdatedAt: \\\"2026-06-10T10:30:00Z\\\"\\nmetadata:\\n  corpus:\\n    qualityScore: 4.5\\n    riskScore: 0.8\\n    promotionMode: auto\\n---\\n\\n# Candidate: Replication crisis systematic review\\n\\nCandidate extracted from a peer-reviewed Nature Reviews article. The pattern (hierarchical Bayesian meta-analysis for heterogeneous replication studies) generalizes across empirical research domains. Analyst recommends auto-promotion.\\n\\n## Analysis\\n- Primary-authority source (peer-reviewed journal, pre-registered)\\n- Methodology (hierarchical Bayesian meta-analysis) is domain-portable\\n- No legal or intellectual-property risk identified\\n- 847-study sample provides robust effect-size estimates\\n\\n## Decision\\n- promotionMode: auto\\n- gates pass: qualityScore 4.5 ≥ 4.2, riskScore 0.8 ≤ 1.5, archive hash present, all required fields populated\\n\",\n  \"entries/critiques/2026/selection-bias-failure.md\": \"---\\nschema: knowledge.entry/v1\\nslug: selection-bias-failure\\nkind: critique\\ntitle: Selection bias as a driver of non-replication\\nupdated_at: \\\"2026-06-10T10:30:00Z\\\"\\nsources:\\n  - nature-review-2026-06\\nconfidence: 0.90\\ntags: [methodology, bias, replication, failure-mode]\\nmetadata:\\n  corpus:\\n    status: active\\n    qualityScore: 4.4\\n    riskScore: 0.5\\n    domain: research\\n    temporal:\\n      firstSeen: \\\"2011-01-01T00:00:00Z\\\"\\n      lastSeen: \\\"2026-06-10T10:30:00Z\\\"\\n      mentions:\\n        - { at: \\\"2011-01-01T00:00:00Z\\\", sourceId: nature-review-2026-06, weight: 1.0 }\\n        - { at: \\\"2026-06-10T10:30:00Z\\\", sourceId: nature-review-2026-06, weight: 1.0 }\\n    promotionMode: human\\n    promotedAt: \\\"2026-06-10T10:30:00Z\\\"\\n    promotedBy: corpus-curator\\n---\\n\\n## Summary\\nStudies that select their sample on the outcome variable or restrict inclusion to high-effect subgroups produce inflated effect sizes that fail to replicate in unselected populations.\\n\\n## Why it fails\\nWhen inclusion criteria correlate with the outcome, the sample is not representative of the target population. The estimated effect is an artifact of the selection process, not a property of the population. Replication studies that sample more broadly will find smaller or null effects.\\n\\n## Failure modes\\n- **Outcome-dependent sampling**: recruiting only people who experienced the effect being studied.\\n- **Subgroup cherry-picking**: reporting the subgroup with the largest effect after observing the data.\\n- **Survivor bias in longitudinal studies**: only measuring participants who completed the study.\\n- **Publication bias at meta-analysis level**: including only published (positive) studies.\\n\\n## Use when\\n- Evaluating a new source for corpus inclusion: check whether sample selection is described.\\n- Scoring riskScore for an entry: unexplained heterogeneity across replications raises riskScore.\\n- Writing critique entries that expose methodological limitations in existing literature.\\n\\n## Avoid when\\n- The selection IS the design (e.g., case studies of rare diseases where selection is inherent).\\n\",\n  \"entries/patterns/2026/systematic-review-pattern.md\": \"---\\nschema: knowledge.entry/v1\\nslug: systematic-review-pattern\\nkind: pattern\\ntitle: Systematic review with hierarchical Bayesian meta-analysis\\nupdated_at: \\\"2026-06-10T10:30:00Z\\\"\\nsources:\\n  - nature-review-2026-06\\nconfidence: 0.88\\ntags: [methodology, meta-analysis, bayesian, replication]\\nmetadata:\\n  corpus:\\n    status: active\\n    qualityScore: 4.5\\n    riskScore: 0.6\\n    domain: research\\n    temporal:\\n      firstSeen: \\\"2018-01-01T00:00:00Z\\\"\\n      lastSeen: \\\"2026-06-10T10:30:00Z\\\"\\n      mentions:\\n        - { at: \\\"2018-01-01T00:00:00Z\\\", sourceId: nature-review-2026-06, weight: 1.0 }\\n        - { at: \\\"2026-06-10T10:30:00Z\\\", sourceId: nature-review-2026-06, weight: 1.0 }\\n    promotionMode: auto\\n    promotedAt: \\\"2026-06-10T10:30:00Z\\\"\\n    promotedBy: corpus-curator\\n---\\n\\n## Summary\\nSystematic reviews achieve reliable effect-size estimates when they: (1) pre-register inclusion criteria, (2) apply hierarchical Bayesian pooling to account for between-study heterogeneity, and (3) report credible intervals rather than p-values.\\n\\n## Why it works\\nHierarchical (random-effects) Bayesian models shrink extreme estimates toward the population mean — the replication crisis showed that underpowered studies with extreme point estimates drive non-replication. Pre-registration eliminates post-hoc analysis flexibility.\\n\\n## Transferable pattern\\n1. Pre-register: hypothesis, inclusion/exclusion criteria, analysis plan.\\n2. Screen independently: two reviewers screen abstracts; resolve conflicts by consensus.\\n3. Pool with hierarchical Bayes: τ² estimates between-study variance; shrink estimates accordingly.\\n4. Report 95 % credible intervals, not just posterior means.\\n5. Conduct sensitivity analysis: exclude low-quality studies, check robustness.\\n\\n## Use when\\n- Synthesizing a body of empirical literature with heterogeneous study designs.\\n- The corpus needs a high-confidence summary of a contested empirical question.\\n- Effect-size estimation matters, not just directional conclusions.\\n\\n## Avoid when\\n- Fewer than 10 primary studies exist (insufficient for pooling).\\n- The research question is qualitative or interpretive rather than quantitative.\\n\",\n  \"entries/principles/2026/primary-sources-beat-synthesis.md\": \"---\\nschema: knowledge.entry/v1\\nslug: primary-sources-beat-synthesis\\nkind: principle\\ntitle: Primary sources beat synthesis\\nupdated_at: \\\"2026-06-10T10:30:00Z\\\"\\nsources:\\n  - nature-review-2026-06\\nconfidence: 0.92\\ntags: [methodology, evidence, authority, sourcing]\\nmetadata:\\n  corpus:\\n    status: active\\n    qualityScore: 4.7\\n    riskScore: 0.4\\n    domain: research\\n    temporal:\\n      firstSeen: \\\"2015-01-01T00:00:00Z\\\"\\n      lastSeen: \\\"2026-06-10T10:30:00Z\\\"\\n      mentions:\\n        - { at: \\\"2015-01-01T00:00:00Z\\\", sourceId: nature-review-2026-06, weight: 1.0 }\\n        - { at: \\\"2026-06-10T10:30:00Z\\\", sourceId: nature-review-2026-06, weight: 1.0 }\\n    promotionMode: human\\n    promotedAt: \\\"2026-06-10T10:30:00Z\\\"\\n    promotedBy: corpus-curator\\n---\\n\\n## Summary\\nWhen a synthesis contradicts its primary sources, trust the primary sources. Systematic reviews compress information; compression loses detail that sometimes reverses the conclusion.\\n\\n## Why it works\\nSynthesis introduces two layers of error: sampling bias in study selection and aggregation error in effect-size combination. Primary data carries neither of these artefacts by construction.\\n\\n## Transferable pattern\\n- Cite primary studies, not meta-analyses, when a claim is contested.\\n- When a meta-analysis conflicts with a landmark primary study, read the primary study.\\n- Record confidence lower when your only source is a secondary synthesis.\\n\\n## Use when\\n- Evaluating evidentiary strength of a claim for corpus promotion.\\n- Deciding authority tier for a new source (primary vs secondary vs rumour).\\n- Writing confidence scores on entries derived from secondary sources.\\n\\n## Avoid when\\n- The domain has no accessible primary literature (policy documents, grey literature).\\n- Speed matters more than precision and a trusted synthesis is available.\\n\",\n  \"KNOWLEDGE.md\": \"---\\nschema: knowledge.workspace/v1\\nname: research-corpus\\ntitle: Research Expert Corpus\\ndescription: |\\n  Domain-agnostic research knowledge for autonomous agents. Composes AIP-10 (knowledge),\\n  AIP-18 (curation collections), AIP-9 (operators), AIP-15 (workflows), and AIP-41\\n  (routines) into an autonomous discovery-and-improvement loop.\\nversion: 1.0.0\\nentityTypes:\\n  - { name: Principle, description: \\\"Durable finding or research rule\\\" }\\n  - { name: Example, description: \\\"Concrete source-derived example or case study\\\" }\\n  - { name: Pattern, description: \\\"Reusable transferable methodology or mechanic\\\" }\\n  - { name: Critique, description: \\\"Negative example, failure mode, or methodological flaw\\\" }\\n  - { name: Summary, description: \\\"Compressed synthesis of a source or topic cluster\\\" }\\n  - { name: Timeline, description: \\\"Time-sensitive trend or field evolution\\\" }\\nsources:\\n  retention: forever\\n  signing: optional\\n  hashAlgo: sha256\\n  authorityDefault: secondary\\nlints:\\n  - { id: require-source-on-examples, kind: require-source, appliesTo: Example, severity: error }\\n  - { id: min-confidence-principles, kind: min-confidence, appliesTo: Principle, severity: warn, params: { min: 0.6 } }\\n  - { id: max-age-timelines, kind: max-age, appliesTo: Timeline, severity: warn, params: { days: 730 } }\\n  - { id: max-age-summaries, kind: max-age, appliesTo: Summary, severity: warn, params: { days: 180 } }\\n  - { id: broken-ref-all, kind: broken-ref, appliesTo: \\\"*\\\", severity: error }\\n  - { id: orphan-all, kind: orphan, appliesTo: \\\"*\\\", severity: info }\\ncuration:\\n  tone: \\\"neutral, expert, source-backed\\\"\\n  depth: medium\\n  autoLink: byName\\n  conflictResolution: authority\\nqueryHints:\\n  preferRecent: true\\n  preferAuthoritative: true\\nmetadata:\\n  corpus:\\n    version: 1\\n    domain: research\\n    entryLayout:\\n      kind: hierarchical\\n      groupBy: entityType\\n    autoPromote:\\n      enabled: true\\n      requires:\\n        qualityScore: { min: 4.2 }\\n        riskScore: { max: 1.5 }\\n        hasArchiveHash: true\\n        requiredFields: [why_it_works, transferable_pattern, use_when, avoid_when]\\n        notRestricted: true\\n    temporal:\\n      defaultHalfLifeDays: 365\\n      scoreFormula: max-mention-decay\\n    retrievalBoosts:\\n      qualityScore: { weight: 0.30 }\\n      temporalScore: { weight: 0.20 }\\n      mentionCount: { weight: 0.10 }\\n      confidence: { weight: 0.15 }\\n    retrievalDefaults:\\n      status: active\\n      minQualityScore: 3.5\\n      minTemporalScore: 0.1\\n    accessModes:\\n      read: { allowedRoles: [\\\"*\\\"] }\\n      cite: { allowedRoles: [\\\"*\\\"] }\\n      flag-learning: { allowedRoles: [\\\"*\\\"], rateLimit: { perOperator: 20, window: \\\"24h\\\" } }\\n      curate: { allowedRoles: [corpus-curator, admin] }\\n      promote: { allowedRoles: [corpus-curator, admin] }\\n      admin-reindex: { allowedRoles: [admin] }\\n      bypass-default-filters: { allowedRoles: [corpus-curator, admin], audit: true }\\n---\\n\\n# Research Corpus\\n\\nSource-of-truth for autonomous research agents. Composes AgentProto AIPs into a closed-loop knowledge-improvement system for any domain.\\n\\n## Three loops\\n- **Knowledge curation** — source-scout → analyst → entries\\n- **Quality gating** — quality-reviewer scores + routes (auto-promote vs human review)\\n- **Retrieval-quality feedback** — every query logged; `utility` + `lift` written back to entries\\n\\nAll corpus-specific policy lives under `metadata.corpus.*`. The six entity types (Principle, Example, Pattern, Critique, Summary, Timeline) are domain-agnostic — topics come from sources, not from the scaffold.\\n\",\n  \"operators/corpus-curator/OPERATOR.md\": \"---\\nname: Corpus Curator\\nid: corpus-curator\\npersona_summary: Promotes approved candidates to AIP-10 entries, maintains _index.md, handles deprecation + supersession. The corpus's editor-in-chief.\\nversion: \\\"1.0.0\\\"\\nprofile:\\n  role: |\\n    Promote candidates that pass the auto-promote gate. Review the\\n    human-review queue, decide approve/reject/needs-work. Keep the\\n    corpus tidy: dedupe, deprecate stale entries, resolve\\n    contradictions.\\n  voice: |\\n    Direct, decisive, opinionated about quality bar. First-person\\n    plural for institutional voice. Cites specific entry slugs when\\n    discussing deprecation or supersession.\\n  boundaries:\\n    - Never promote an entry that fails the auto-promote gate without explicit bypass.\\n    - Never deprecate without a written reason in metadata.corpus.archiveReason.\\n    - Always preserve the audit chain — _log.md events are append-only.\\ngovernance:\\n  audit_log: \\\"audit:corpus/research/_log.md\\\"\\n  autonomy: autonomous\\n  policies:\\n    - \\\"policy:corpus/research/curation\\\"\\nskills:\\n  - source-analysis\\n  - taxonomy\\n  - editorial-judgment\\ntools:\\n  - corpus-promote-candidate\\n  - corpus-log-event\\n  - knowledge-query\\nparticipation:\\n  mode: proactive\\nmemory:\\n  kind: operator-context\\nruntime:\\n  kind: in-process\\ntags: [research, curator, corpus]\\nmetadata:\\n  corpus:\\n    domain: research\\n    overlays:\\n      policy: open\\n      maxActiveCount: 10\\n---\\n\\n# Corpus Curator\\n\\nRuns after the reviewer scores a candidate. Promotes auto-approved entries; queues human-review-required candidates for the team. Drives entry lifecycle (active → archived).\\n\",\n  \"operators/quality-reviewer/OPERATOR.md\": \"---\\nname: Quality Reviewer\\nid: quality-reviewer\\npersona_summary: Scores analyzed candidates against the corpus's quality + risk rubric. Decides promotionMode (auto vs human). Flags concerns the analyst missed.\\nversion: \\\"1.0.0\\\"\\nprofile:\\n  role: |\\n    Independent second pair of eyes on every analyzed candidate.\\n    Adversarial-by-design — assumes the analyst was too generous\\n    and looks for reasons NOT to promote.\\n  voice: |\\n    Critical, concrete, fair. Cites specific clauses of the\\n    candidate body, not vibes. Always proposes a concrete score\\n    (qualityScore 0-5, riskScore 0-5).\\n  boundaries:\\n    - Never promote on the analyst's authority — score and route.\\n    - Never auto-promote a candidate flagged as restricted / personal / legal.\\n    - Always justify a sub-2.0 quality score in reviewerNotes.\\ngovernance:\\n  audit_log: \\\"audit:corpus/research/_log.md\\\"\\n  autonomy: supervised\\n  policies:\\n    - \\\"policy:corpus/research/review-policy\\\"\\nskills:\\n  - source-analysis\\n  - critical-reading\\ntools:\\n  - corpus-flag-learning\\n  - knowledge-query\\nparticipation:\\n  mode: proactive\\nmemory:\\n  kind: operator-context\\nruntime:\\n  kind: in-process\\ntags: [research, reviewer, corpus]\\nmetadata:\\n  corpus:\\n    domain: research\\n---\\n\\n# Quality Reviewer\\n\\nRuns after `research-analyst` produces an analyzed ITEM.md. Scores against:\\n\\n- **qualityScore** — accuracy, transferability, evidence strength, authority tier\\n- **riskScore** — legal/IP/compliance hazards, retracted studies, contested findings\\n- **promotionMode** — auto (gates pass) vs human (needs curator)\\n\",\n  \"operators/research-analyst/OPERATOR.md\": \"---\\nname: Research Analyst\\nid: research-analyst\\npersona_summary: Domain-agnostic research analyst who converts raw sources into expert principles, patterns, and critiques for the research corpus.\\nversion: \\\"1.0.0\\\"\\nprofile:\\n  role: |\\n    Convert candidate sources discovered by the scout into corpus-grade\\n    expert analysis: principles, patterns, critiques. Score quality and\\n    risk to inform promotion gates. Works across any research domain —\\n    the topic comes from the source, not from the operator persona.\\n  voice: |\\n    Concrete, source-backed, never superlative. Cite specific numbers and\\n    named authors or publications. First-person plural (\\\"we\\\") for\\n    institutional voice.\\n  boundaries:\\n    - Never invent quantitative claims; cite a source or mark as opinion.\\n    - Never promote without quality-reviewer signoff (unless auto-promote gates pass).\\n    - Never edit raw sources — only entries.\\ngovernance:\\n  audit_log: \\\"audit:corpus/research/_log.md\\\"\\n  autonomy: supervised\\n  policies:\\n    - \\\"policy:corpus/research/curation\\\"\\nskills:\\n  - source-analysis\\n  - research-synthesis\\ntools:\\n  - corpus-flag-learning\\n  - knowledge-query\\nparticipation:\\n  mode: proactive\\nmemory:\\n  kind: operator-context\\nruntime:\\n  kind: in-process\\ntags: [research, analyst, corpus]\\nmetadata:\\n  corpus:\\n    domain: research\\n    knowledgeViews:\\n      - corpus: research\\n        filter:\\n          domain: [research]\\n          minQualityScore: 3.5\\n        defaultBoosts:\\n          qualityScore: { weight: 0.40 }\\n          temporalScore: { weight: 0.30 }\\n    overlays:\\n      policy: gated\\n      maxActiveCount: 5\\n      requireApprovalFrom: corpus-curator\\n      allowedKinds: [overlay, block-replacement]\\n---\\n\\n# Research Analyst\\n\\nReads candidate sources from the scout, writes expert analysis as AIP-10 entries. Scores quality and risk. Signals the curator on auto-promote gate pass.\\n\",\n  \"operators/source-scout/OPERATOR.md\": \"---\\nname: Source Scout\\nid: source-scout\\npersona_summary: Curates external feeds for fresh research source material — academic papers, articles, videos, social posts, and primary documents — and archives the worth-keeping ones into sources/.\\nversion: \\\"1.0.0\\\"\\nprofile:\\n  role: |\\n    Discover new candidate sources for the research corpus. Filter\\n    noise vs signal at the scout stage so the analyst only sees\\n    promising material. Source categories: article, video, social,\\n    paper, primary, news.\\n  voice: |\\n    Efficient, terse, evidence-first. Names URLs + numbers, not\\n    adjectives. Reports findings as `{source: <id>, why: <reason>}`\\n    triples.\\n  boundaries:\\n    - Never re-archive an already-archived source (check content_hash).\\n    - Never archive content with PII unless explicitly sanitized.\\n    - Never archive paywalled / DMCA-restricted material without authorization.\\ngovernance:\\n  audit_log: \\\"audit:corpus/research/_log.md\\\"\\n  autonomy: autonomous\\n  policies:\\n    - \\\"policy:corpus/research/source-policy\\\"\\nskills:\\n  - web-research\\n  - source-analysis\\ntools:\\n  - corpus-create-candidate\\n  - knowledge-query\\n  - web-search\\nparticipation:\\n  mode: proactive\\nmemory:\\n  kind: operator-context\\nruntime:\\n  kind: in-process\\ntags: [research, scout, corpus]\\nmetadata:\\n  corpus:\\n    domain: research\\n    knowledgeViews:\\n      - corpus: research\\n        filter:\\n          domain: [research]\\n---\\n\\n# Source Scout\\n\\nRuns on-demand or on schedule (via `source-scout` routine). Fetches configured feeds and channels, hashes content, dedups against existing sources, and writes new candidates to `_candidates.yaml`. The analyst takes over from there.\\n\\nSupported source categories: `article` (blog/news), `video` (YouTube, conference talks), `social` (threads, posts), `paper` (peer-reviewed), `primary` (official documents, datasets), `news` (breaking coverage).\\n\",\n  \"routines/source-scout/ROUTINE.md\": \"---\\nschema: routine/v1\\nid: source-scout\\ndescription: |\\n  On-demand (or scheduled) routine that runs the source-scout operator to\\n  discover new candidate sources across configured external feeds and channels.\\n  Discovered sources are archived under sources/ and rows appended to\\n  _candidates.yaml.\\nversion: \\\"1.0.0\\\"\\nschedule:\\n  kind: cron\\n  cron: \\\"0 8 * * 1\\\"\\n  timezone: \\\"UTC\\\"\\n  catchup: skip\\ntarget:\\n  workflow: ingest-source\\n  inputs:\\n    feeds: [web-search, youtube, arxiv, news-api]\\n    maxItemsPerRun: 30\\nretry:\\n  max_attempts: 3\\n  backoff: exponential\\non_failure:\\n  create_work_item: true\\n  fire_event: corpus.scout.failed\\nfires_events:\\n  - corpus.scout.completed\\n  - corpus.candidate.discovered\\nenabled: true\\ntags: [corpus, scout, weekly]\\nmetadata:\\n  corpus:\\n    domain: research\\n---\\n\\n# Source Scout Routine\\n\\nRuns weekly (Monday 08:00 UTC) by default; can also be triggered on-demand. Scouts external feeds for new candidate sources, archives them, and appends to `_candidates.yaml`. The candidate event downstream triggers `analyze-candidate` workflow.\\n\\nAdjust the `schedule` and `target.inputs.feeds` list to match your research domain's primary discovery channels.\\n\",\n  \"sources/fresh/nature-review-2026-06.md\": \"---\\nschema: knowledge.source/v1\\nid: nature-review-2026-06\\npath: sources/fresh/nature-review-2026-06.md\\ntitle: Systematic review of replication failures in social science — June 2026\\ncaptured_at: \\\"2026-06-10T09:00:00Z\\\"\\ncaptured_by: source-scout\\ncontent_hash: sha256:b2e4f8a1c3d5e7f9a1b3c5d7e9f1a3b5c7d9e1f3a5b7c9d1e3f5a7b9c1d3e5f7\\nauthority: primary\\nlanguage: en\\ntags: [systematic-review, replication, methodology, social-science]\\nmetadata:\\n  corpus:\\n    originalUrl: https://example.com/nature/reviews/replication-crisis-2026\\n    archiveTier: hot\\n    archivedBy: source-scout\\n    snapshotPath: sources/fresh/nature-review-2026-06.snapshot.mhtml\\n---\\n\\n# Systematic review: replication failures in social science (2026)\\n\\nCaptured from a peer-reviewed Nature Reviews article in June 2026. Meta-analysis of 847 attempted replications across psychology, economics, and political science (2010–2026).\\n\\n## Key findings\\n- Only 39 % of original findings replicated at p < 0.05 in pre-registered replications\\n- Effect sizes in successful replications averaged 68 % of originals\\n- Primary driver of non-replication: underpowered original studies, not fraud\\n\\n## Methodology note\\nAuthors applied a hierarchical Bayesian model to account for heterogeneity across domains. Pre-registered analysis plan at OSF (osf.io/example123).\\n\\n## Relevance\\nFoundational reference for any research corpus reasoning about evidentiary quality, source authority, and confidence thresholds.\\n\",\n  \"workflows/analyze-candidate/WORKFLOW.md\": \"---\\nname: Analyze Corpus Candidate\\nid: analyze-candidate\\ndescription: |\\n  Convert a discovered corpus candidate into an analyzed ITEM.md with\\n  expert analysis, quality score, and risk score. Triggered by the\\n  source-scout when it appends a row to _candidates.yaml.\\nversion: \\\"1.0.0\\\"\\ninputs:\\n  type: object\\n  required: [candidateId, sourcePath]\\n  properties:\\n    candidateId: { type: string, pattern: \\\"^[a-z0-9][a-z0-9-]*$\\\" }\\n    sourcePath: { type: string }\\noutputs:\\n  type: object\\n  required: [itemPath, status]\\n  properties:\\n    itemPath: { type: string }\\n    status: { type: string, enum: [analyzed, rejected, needs-work] }\\n    qualityScore: { type: number }\\n    riskScore: { type: number }\\nsteps:\\n  - id: load-source\\n    kind: tool\\n    tool: knowledge.read_source\\n    name: Load source bytes from sources/\\n    inputs: { type: object, properties: { sourceId: { type: string } } }\\n    next: analyze\\n  - id: analyze\\n    kind: tool\\n    tool: research-analyst.analyze\\n    name: Run research-analyst operator over the source\\n    next: score\\n  - id: score\\n    kind: tool\\n    tool: quality-reviewer.score\\n    name: Compute quality + risk scores\\n    next: gate\\n  - id: gate\\n    kind: branch\\n    name: Decide promotionMode (auto vs human)\\n    description: Check auto-promote gate against KNOWLEDGE.md.metadata.corpus.autoPromote\\n    branches:\\n      - { when: \\\"output.qualityScore >= 4.2 && output.riskScore <= 1.5\\\", next: \\\"auto-promote\\\" }\\n      - { when: \\\"true\\\", next: \\\"human-review\\\" }\\n  - id: auto-promote\\n    kind: tool\\n    tool: corpus-curator-promote\\n    name: Promote to AIP-10 entry\\n  - id: human-review\\n    kind: approval\\n    name: Queue for human curator\\n    prompt: Review and approve or reject this candidate.\\n    approvers:\\n      - { role: corpus-curator }\\ntags: [corpus, workflow, analysis]\\nmetadata:\\n  corpus:\\n    domain: research\\n    triggeredBy: source-scout\\n---\\n\\n# Analyze Candidate workflow\\n\\nConverts a discovered candidate into an analyzed ITEM.md. The research-analyst operator runs the actual analysis; this workflow orchestrates the pipeline (load → analyze → score → gate).\\n\",\n  \"workflows/ingest-source/WORKFLOW.md\": \"---\\nname: Ingest source\\nid: ingest-source\\ndescription: |\\n  Fetch + hash + dedup + archive a single source. Output: a new\\n  candidate row in _candidates.yaml (or no-op if duplicate).\\nversion: \\\"1.0.0\\\"\\ninputs:\\n  type: object\\n  required: [sourceUrl]\\n  properties:\\n    sourceUrl: { type: string }\\n    category: { type: string, enum: [article, video, social, paper, primary, news] }\\noutputs:\\n  type: object\\n  required: [outcome]\\n  properties:\\n    outcome: { type: string, enum: [archived, duplicate, rejected] }\\n    sourceId: { type: string }\\n    candidateId: { type: string }\\nsteps:\\n  - id: fetch\\n    kind: tool\\n    tool: web-fetch\\n    name: Fetch source content\\n  - id: hash\\n    kind: tool\\n    tool: hash-content\\n    name: Compute sha256 content_hash\\n  - id: dedup\\n    kind: branch\\n    name: Dedup against existing sources\\n    branches:\\n      - { when: \\\"output.duplicate === true\\\", next: \\\"done-dup\\\" }\\n      - { when: \\\"true\\\", next: \\\"archive\\\" }\\n  - id: archive\\n    kind: tool\\n    tool: corpus-archive-source\\n    name: Write source frontmatter + body to sources/<category>/<slug>.md\\n  - id: create-candidate\\n    kind: tool\\n    tool: createCorpusCandidate\\n    name: Append discovered row to _candidates.yaml\\n  - id: done-dup\\n    kind: tool\\n    tool: corpus-log-event\\n    name: Log dedup hit (no-op outcome)\\ntags: [corpus, ingest, scout]\\nmetadata:\\n  corpus:\\n    domain: research\\n    triggeredBy: source-scout\\n---\\n\\n# Ingest source workflow\\n\\nStep 1 of the corpus loop. Single-source entry point. Called by the `source-scout` routine once per discovered URL. The `category` input (article | video | social | paper | primary | news) drives the archive path under `sources/`.\\n\",\n  \"workflows/promote-corpus-item/WORKFLOW.md\": \"---\\nname: Promote corpus item\\nid: promote-corpus-item\\ndescription: |\\n  Lift an approved candidate into a published AIP-10 entry. Writes\\n  the entry file, regenerates _index.md, ships chunks to the\\n  backing engine, appends a promotion event. Wraps the kit's\\n  CorpusPromoter.promote().\\nversion: \\\"1.0.0\\\"\\ninputs:\\n  type: object\\n  required: [candidateId, entrySlug, entryKind, entryPath, frontmatter, body]\\n  properties:\\n    candidateId: { type: string }\\n    entrySlug: { type: string }\\n    entryKind: { type: string, enum: [principle, example, pattern, critique, summary, timeline] }\\n    entryPath: { type: string }\\n    frontmatter: { type: object }\\n    body: { type: string }\\n    bypassGate: { type: boolean }\\noutputs:\\n  type: object\\n  required: [entryPath, gatePassed]\\n  properties:\\n    entryPath: { type: string }\\n    versionToken: { type: string }\\n    chunkCount: { type: integer }\\n    gatePassed: { type: boolean }\\n    bypassed: { type: boolean }\\nsteps:\\n  - id: promote\\n    kind: tool\\n    tool: promoteCorpusCandidate\\n    name: Atomic promote (lock + write + index + chunks + event)\\ntags: [corpus, promote]\\nmetadata:\\n  corpus:\\n    domain: research\\n    triggeredBy: corpus-curator\\n---\\n\\n# Promote corpus item workflow\\n\\nStep 3 of the corpus loop. Single-step wrapper for callers that want the workflow shape; the underlying `CorpusPromoter` does the multi-file transaction.\\n\",\n})\n","/**\n * Research corpus preset.\n *\n * Pure TS data — `files` is inlined from the M0 fixture workspace at\n * build time via `scripts/gen-research.mjs`. The M0 conformance test\n * proves every file validates against the actual AgentProto JSON\n * Schemas, so a host that writes these to disk gets a known-good\n * starter workspace.\n *\n * Usage:\n *\n *     import { ResearchCorpusPreset } from \"@agentproto/corpus-presets/research\"\n *     for (const [rel, content] of Object.entries(ResearchCorpusPreset.files)) {\n *       await fs.writeFile(path.join(workspaceRoot, rel), content)\n *     }\n */\n\nimport type { CorpusPreset } from \"@agentproto/corpus\"\nimport { RESEARCH_PRESET_FILES } from \"./files.js\"\n\nexport const ResearchCorpusPreset: CorpusPreset = Object.freeze({\n  slug: \"research\",\n  title: \"Research Expert Corpus\",\n  description:\n    \"Domain-agnostic research knowledge — principles, patterns, critiques. Composes AIP-10/18/9/15/41 into a closed-loop discovery-and-improvement system.\",\n  files: RESEARCH_PRESET_FILES,\n})\n\n// Re-export the raw file map for callers that want it directly.\nexport { RESEARCH_PRESET_FILES }\n"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agentproto/corpus-presets",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "@agentproto/corpus-presets — catalog of starter workspaces for the AIP-10 corpus runtime. Subpath exports per vertical (marketing, sales, …). Each preset is pure TS data — host iterates `files` and writes via FsPort. Mirrors @agentproto/role-catalog.",
   "keywords": [
     "agentproto",
@@ -34,6 +34,11 @@
       "import": "./dist/marketing/index.mjs",
       "default": "./dist/marketing/index.mjs"
     },
+    "./research": {
+      "types": "./dist/research/index.d.ts",
+      "import": "./dist/research/index.mjs",
+      "default": "./dist/research/index.mjs"
+    },
     "./package.json": "./package.json"
   },
   "agentproto-corpus-preset": {
@@ -44,6 +49,12 @@
         "entry": "./dist/marketing/index.mjs",
         "export": "MarketingCorpusPreset",
         "description": "Marketing knowledge corpus — principles, patterns, critiques, playbooks. Composes AIP-10/12/18/9/15/41 into a closed-loop improvement system."
+      },
+      {
+        "slug": "research",
+        "entry": "./dist/research/index.mjs",
+        "export": "ResearchCorpusPreset",
+        "description": "Domain-agnostic research knowledge corpus — principles, patterns, critiques. Composes AIP-10/18/9/15/41 into a closed-loop discovery-and-improvement system."
       }
     ]
   },
@@ -57,7 +68,7 @@
   },
   "sideEffects": false,
   "dependencies": {
-    "@agentproto/corpus": "0.1.0"
+    "@agentproto/corpus": "0.2.0"
   },
   "devDependencies": {
     "@types/node": "^25.6.2",
@@ -67,15 +78,16 @@
     "@agentproto/tooling": "0.1.0-alpha.0"
   },
   "scripts": {
-    "predev": "node scripts/gen-marketing.mjs",
+    "predev": "node scripts/gen-marketing.mjs && node scripts/gen-research.mjs",
     "dev": "tsup --watch",
-    "prebuild": "node scripts/gen-marketing.mjs",
+    "prebuild": "node scripts/gen-marketing.mjs && node scripts/gen-research.mjs",
     "build": "tsup",
-    "clean": "rm -rf dist src/marketing/files.ts",
-    "precheck-types": "node scripts/gen-marketing.mjs",
+    "clean": "rm -rf dist src/marketing/files.ts src/research/files.ts",
+    "precheck-types": "node scripts/gen-marketing.mjs && node scripts/gen-research.mjs",
     "check-types": "tsc --noEmit",
-    "pretest": "node scripts/gen-marketing.mjs",
+    "pretest": "node scripts/gen-marketing.mjs && node scripts/gen-research.mjs",
     "test": "vitest run --passWithNoTests",
-    "gen:marketing": "node scripts/gen-marketing.mjs"
+    "gen:marketing": "node scripts/gen-marketing.mjs",
+    "gen:research": "node scripts/gen-research.mjs"
   }
 }