All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

• ScientificReport schema 1.3.0 — immutable SourceRecord, claim–evidence graph, write barrier, reconstruct path (current public artifact baseline)
• ScientificReport schema 1.2.0 — lifecycle fields, evidence_content_hash, validation_run_hash (still accepted; unchanged by decision-time companions)
• ScientificReport schema 1.1.0 — Ed25519 signature, review chain hash, audit metadata
• ScientificReport schema 1.0.0 — initial canonical artifact model
• Companion schemas 1.0.0 — integration_envelope, decision_time_evidence_bundle, akta_action_context, scientific_outcome_link (additive; do not mutate ScientificReport)

• ADR 0001: VSA / AKTA / Scientific Memory boundaries and decision-time semantics
• SECURITY.md and docs/threat-model.md
• Integration envelope builder (src/vsa/identity/)
• Decision-time evidence bundle with cutoff filter, verify, and compare (src/vsa/decision_time/)
• AKTA action-context export with non-authorization semantic validation (src/vsa/export/akta.py)
• PCS-shaped portable evidence refs and signed bundle verification
• Pilot modes (retrospective / shadow / read_only), privacy-safe metrics, outcome append links
• CLI: snapshot-evidence, verify-decision-time-bundle, compare-cutoffs, export-akta-context, link-outcome
• Operator guide, decision-time quick start, and independent reconstruction commands: docs/operator-decision-time.md, docs/decision-time-quickstart.md
• Public companion schema mirrors under schemas/ (byte-identical to src/vsa/schemas/)
• Adversarial decision-time tests (tamper, forged signature, digest mismatch, missing portable refs)
• Dev quality tooling: Ruff (warnings as errors) and MyPy scoped to companion modules

• Contradiction/uncertainty hardening for decision-time bundles (post-cutoff, high-quality conflicts, metadata-only, entity ambiguity, computational→experimental extrapolation)
• ScientificReport advisory checks now emit warn status (stale evidence, entity ambiguity, computational→experimental language)
• Bundle verify fail-closed on missing portable refs, corrupted non-claims; signatures verified over recomputed hash
• privacy_safe_metrics included in bundle_hash; AKTA export rejects evidence digest disagreement with report
• Missing timestamps fail closed in all snapshot modes (not only retrospective/read_only)
• Verify fail-closed on included/excluded evidence_id overlap and duplicate included IDs
• Privacy-safe metrics revalidated on verify (reject digest/secret-shaped keys)
• Optional out-of-band --trusted-key pin for decision-time signature verify
• compare-cutoffs --fail-on-verify and clearer CLI load errors (missing/empty/invalid JSON)
• AKTA rejects grant-like tool_or_execution_class tokens

• README decision-time CLI section, claim boundaries, upstream PCS/AKTA/signing pins
• Docs index versions table includes companion schemas 1.0.0
• Schema docs: public mirrors (not symlinks); ClinVar ambiguity → low reliability

• Scientific credibility module (src/vsa/scientific/credibility.py) with ambiguity caps and report warnings
• Validation checks: ambiguous reliability cap, AlphaFold predicted labeling, ClinVar ambiguity visibility
• Benchmark expanded to 50 tasks with category minimums (10 adversarial, 5 each for ambiguity/contradiction/metadata-only/no-evidence)
• Core benchmark metrics: source recall/precision, citation integrity, evidence-ID validity, review-boundary accuracy, contradiction detection, bundle reproducibility
• Tests for credibility policies and Materials Project missing-key degradation

• ClinVar ambiguous queries always return reliability: low
• UniProt gene search sets retrieval_ambiguity and caps reliability to low
• Materials Project skipped cleanly with explicit warning when API key missing
• Scientific credibility warnings in markdown and HTML render output
• Retrieval pipeline applies credibility hardening before evidence is returned
• Acceptance bar and CI assert benchmark category minimums (50 tasks)
• Streamlit UI surfaces scientific credibility warnings and evidence reliability/ambiguity flags
• vsa benchmark prints category coverage summary on success

• Optional API key auth (VSA_API_KEY, X-API-Key header)
• vsa verify-review top-level command
• scripts/acceptance.sh and make acceptance (CI parity)
• Mocked connector tests: Semantic Scholar, Europe PMC
• Release workflow attaches full artifact bundle zip

• Semantic Scholar connector sets content_level metadata
• Build pipeline adds metadata-only CONTENT WARNING to limitations
• Acceptance CI job uses shared acceptance script
• Documentation consolidated and updated across README, docs/, and RELEASE_STATUS

• Review subcommands: start, approve-claim, request-corrections, reject, verify
• API review endpoints: /v1/review/start, /v1/review/approve-claim, /v1/review/verify
• Typed Pydantic request/response models for core API routes
• Structured API error codes (MISSING_INPUT, VALIDATION_ERROR, RATE_LIMITED)
• Optional rate limiting (VSA_API_RATE_LIMIT) and deterministic mode (VSA_API_DETERMINISTIC)
• Review chain verification via vsa review verify and verify_review_chain()

• docs/api.md expanded with env vars, error format, review parity
• CI runs review workflow smoke after bundle verification

• vsa verify-bundle — manifest hash verification and attestation check
• Export bundle v1.1: report.md, sources/, manifest created_at, schema/validation versions
• Benchmark regression gate (100% pass rate), expanded metrics, 27 tasks (fake evidence ID, stale source)
• CI acceptance job: clean-clone make demo && make test && vsa benchmark
• docs/benchmark.md; RELEASE_STATUS verified-commit table

• Deeper variant/paper/protein/material claim templates
• ClinVar HGVS-only lookup; UniProt gene disambiguation warnings
• macOS CI smoke installs [signing]; Makefile install includes [api]
• Metadata-only paper claims use explicit CONTENT WARNING

• Publication content levels (metadata, abstract, fulltext) on OpenAlex, Crossref, and Europe PMC connectors
• Paper claim templates: bibliographic identity, abstract-derived observation, full-text availability flag
• Export bundle manifest.json with SHA-256 hashes for all artifacts; attestation included by default
• vsa compare-audit for diffing audit artifacts across runs
• Validation check: warn when all publication evidence is metadata-only
• Streamlit: export bundle download and attestation generation in Provenance tab
• E2E pipeline test (build → validate → audit → attest → export)
• Expanded benchmark gold labels (paper_doi, tp53, egfr)

• REST API (vsa serve, [api] extra): /v1/build, /validate, /audit, /attest, /retrieve, /health
• SLSA Provenance v1 / in-toto Statement attestation (vsa attest, vsa verify-attestation)
• OpenTelemetry hooks (VSA_OTEL_ENABLED=1, [otel] extra) with httpx instrumentation
• Schema migration CLI (vsa migrate-schema) for 1.0.0/1.1.0 → 1.2.0
• Benchmark gold label scoring (gold_labels in tasks.json)
• Live connector tests (pytest -m live) and weekly GitHub Actions workflow
• macOS CI smoke job

• Schema 1.2.0 lifecycle fields: input_question, retrieval_plan, retrieval_warnings, limitations, domain, etc.
• Provenance: evidence_content_hash, validation_run_hash
• ClinVar candidate ranking (up to 10 search hits, ambiguity flags, reliability scoring)
• Paper deduplication across OpenAlex/Crossref/PubMed/Europe PMC/Semantic Scholar
• UniProt Swiss-Prot vs TrEMBL labeling; AlphaFold predicted-structure warnings
• 25-task offline benchmark with adversarial cases and expanded scoring
• Audit artifacts: vsa audit --out, vsa export --out-dir
• Docs: docs/architecture.md, docs/schema.md, docs/connectors.md, docs/release_checklist.md
• RELEASE_STATUS.md, README badges, known limitations
• Connector tests under tests/connectors/
• Expanded variant claim templates (identity, classification, ambiguity)

• CI uploads report/audit/benchmark artifacts
• make demo uses --audit-mode rule for deterministic output

• LLM-backed scientific verifier (vsa audit --audit-mode auto|rule|llm)
• Hybrid audit merge: rule structural checks plus LLM semantic review (conservative)
• Audit output fields: missing_evidence, confidence_concerns, verifier_method
• Packaged verifier prompt at src/vsa/prompts/verifier_prompt_v1.md
• Streamlit audit mode selector (auto / rule / llm)

• Ed25519 report signing (vsa sign, vsa verify-signature) via [signing] extra
• Rule-based scientific audit layer (vsa audit) implementing verifier prompt logic
• Schema 1.1.0: signature, review_chain_hash, review_event_hash, PDF output format
• Legacy claim-ledger migration (vsa migrate)
• Benchmark CLI (vsa benchmark) with offline and --live modes
• Materials domain example input and pipeline test
• Streamlit: audit, sign, and verify-signature actions in Provenance tab

• human_review included in provenance report_hash core
• human_review.required is conditional (clinical/speculative/contradictions), not always true
• Provenance preserves signature, generated_by, and review_chain_hash across re-stamping

• NCBI PubMed connector (src/vsa/connectors/ncbi_pubmed.py)
• PDF report renderer (pip install verified-science-agent[pdf])
• Evidence graph (Mermaid) in UI and src/vsa/render/graph.py
• Streamlit v3: build pipeline, LLM controls, human review tab, PDF download
• Material/chemical subject parsing (LiFePO4, cathode keywords)
• Retrieval warnings surfaced in CLI and UI
• Release workflow on git tags (.github/workflows/release.yml)

• Connector cache now propagates through --cache-dir
• Materials Project routing for material entity types
• CLI: inspect --json, compare --strict, review --out --notes
• Legacy scripts delegate to vsa CLI

• LLM claim extraction via OpenAI and Anthropic (--claim-mode auto|rule|llm)
• vsa extract command for inspectable claim JSON
• vsa review command for claim-by-claim human review workflow
• Materials Project connector (requires MATERIALS_PROJECT_API_KEY)
• Environment config via .env (python-dotenv)
• .env.example template

• vsa build auto-selects LLM when API keys are present; falls back to rule-based
• Provenance records review events and extraction method stack

• Canonical ScientificReport JSON Schema (v1.0.0) with support for genomics, proteins, papers, chemicals, materials, and experiments
• Installable Python package with vsa CLI: validate, render, hash, inspect, compare, retrieve, build
• Validation engine with schema conformance, evidence linkage, confidence bounds, review boundaries, contradiction detection, and provenance hash verification
• Read-only database connectors: OpenAlex, Europe PMC, UniProt, ClinVar, AlphaFold DB, Crossref, Semantic Scholar
• Evidence retrieval pipeline with file-based caching
• Rule-based claim extraction (no LLM source field invention)
• Explainable evidence quality scoring
• Markdown and HTML report renderers
• Streamlit UI v2 with claim explorer, evidence view, provenance, and validation failures
• Benchmark suite with 6 domain tasks and offline fixtures
• CI workflow: schema validation, pytest, CLI smoke tests, render/hash tests
• Example reports and bad examples for regression testing