Add reproducible prior art evidence matrix

.github/workflows/ci.yml

@@ -606,3 +606,40 @@ jobs:
           echo "::error::any dependency, run 'uv lock' and commit the result in the SAME commit."
           echo "::error::A stale lock means the committed pins are not what this project resolves."
           exit 1
+
+  prior-art:
+    name: Prior art evidence corpus
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7
+        with:
+          fetch-depth: 0
+
+      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97  # v7
+        with:
+          python-version: "3.11"
+
+      - name: Validate canonical prior art records
+        run: python -m tools.prior_art validate
+
+      - name: Verify generated prior art reports
+        run: python -m tools.prior_art render --check
+
+      - name: Check prior work declarations in changed experiments
+        shell: bash
+        run: |
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            exit 0
+          fi
+          if [ "${{ github.event_name }}" = "pull_request" ]; then
+            base="${{ github.event.pull_request.base.sha }}"
+          else
+            base="${{ github.event.before }}"
+          fi
+          if [ -z "$base" ] || [[ "$base" =~ ^0+$ ]]; then
+            exit 0
+          fi
+          mapfile -t files < <(git diff --name-only --diff-filter=AM "$base" "${{ github.sha }}" -- 'benchmarks/**/*.py' 'results/**/*.py')
+          if ((${#files[@]})); then
+            python -m tools.prior_art check-experiments "${files[@]}"
+          fi

benchmarks/EXPERIMENT-CONVENTION.md

@@ -12,6 +12,12 @@ or, when nothing was found:
 Prior work: none found (docs_search "<query>", source_type=memory)
 ```
 
+For changed or newly added modules, run the repository check with the affected paths:
+
+```text
+python -m tools.prior_art check-experiments benchmarks/my_probe.py
+```
+
 ## Why this exists rather than a rule someone remembers
 
 On 2026-07-28 an abstention investigation was run that had already been run on 2026-07-24, and

docs/PRIOR_ART.md

@@ -1,5 +1,26 @@
 # Prior art, and where this comes from
 
+## Structured prior art evidence system
+
+This page remains the public narrative, while the canonical comparison data now lives in the
+[RE-call prior art evidence system](prior_art/README.md). The structured corpus records sources,
+systems, capability definitions, atomic claims, evidence locations, and human review decisions.
+
+The generated [capability matrix](prior_art/generated_matrix.md) and [gap report](prior_art/generated_gap_report.md)
+are reproducible from the checked in JSON and JSON Lines records. They preserve `verified`,
+`partial`, `not_evidenced`, `contradicted`, and `unknown` as separate states. Missing evidence is
+not treated as proof that a system lacks a capability.
+
+The current RE-call research hypothesis combines evidence backed claims, explicit validity and
+supersession, reversible provenance lineage, authority and scope enforcement, deletion propagation
+through derived artifacts, support based abstention, and action outcome feedback. The matrix may
+identify this as an unverified research gap, but it does not automatically claim that RE-call is
+first, unique, or unprecedented.
+
+The collection scope covers agent memory systems, agent runtimes, benchmarks, privacy, governance,
+and continual learning. The search cutoff and evidence rules are recorded in
+[report_config.json](prior_art/report_config.json).
+
 ## Prior art — and where this genuinely differs
 
 Agent memory is a crowded field. Everything below is Apache-2.0 and further along than this

docs/prior_art/README.md

@@ -0,0 +1,43 @@
+# RE-call prior art evidence system
+
+This directory is the canonical, reviewable evidence corpus behind the public prior art matrix.
+
+The corpus separates five things:
+
+1. Sources, which are external papers, repositories, documentation, benchmarks, protocols, or standards.
+2. Systems, which are the products or research systems being compared.
+3. Capabilities, which are stable, defined matrix dimensions.
+4. Claims, which are atomic statements about one system and one capability.
+5. Reviews, which record the human decision that allows a claim into the generated report.
+
+The accepted capability values are `verified`, `partial`, `not_evidenced`, `contradicted`, and
+`unknown`. `not_evidenced` means that reviewed sources did not establish a capability. `unknown`
+means that the investigation is incomplete. Neither value proves that a system lacks a capability.
+
+## Workflow
+
+Run validation and rendering from the repository root:
+
+```text
+python -m tools.prior_art validate
+python -m tools.prior_art render
+python -m tools.prior_art render --check
+python -m tools.prior_art check-experiments benchmarks/my_probe.py
+python -m tools.prior_art check-links
+```
+
+The link check is intentionally manual because external availability is unstable. It does not
+rewrite source records.
+
+## Evidence policy
+
+Secondary sources may discover candidates, but accepted `verified` claims require primary evidence.
+Each accepted claim has a source, an evidence locator, a human review record, and a short evidence
+note. The renderer never creates a claim from silence and never emits an automatic claim of novelty.
+
+## Generated reports
+
+* [Capability matrix](generated_matrix.md)
+* [Gap report](generated_gap_report.md)
+* [Machine readable summary](generated_summary.json)
+* [Public prior art position](../PRIOR_ART.md)