SCM v1.1.0: deterministic lineage and reliability gates

.github/workflows/product-qualification.yml

@@ -28,7 +28,7 @@ jobs:
       - name: Install qualification dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install -e ".[dev]" build twine
+          pip install -e ".[dev,langchain]" build twine
 
       - name: Run credential-free product qualification
         run: |

.github/workflows/release.yml

@@ -34,7 +34,7 @@ jobs:
       - name: Install qualification dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install -e ".[dev]" build twine
+          pip install -e ".[dev,langchain]" build twine
 
       - name: Verify tag matches package version
         env:
@@ -135,11 +135,18 @@ jobs:
     needs: publish
 
     permissions:
+      actions: read
       contents: write
 
     steps:
       - uses: actions/checkout@v4
 
+      - name: Download qualification evidence
+        uses: actions/download-artifact@v4
+        with:
+          name: release-qualification
+          path: release-evidence
+
       - name: Create release notes from CHANGELOG
         run: |
           VERSION="${GITHUB_REF_NAME#v}"
@@ -156,10 +163,17 @@ jobs:
           pip install build twine
           python -m build --wheel --sdist
           twine check ./dist/*
+          VERSION="${GITHUB_REF_NAME#v}"
+          mkdir -p release-assets
+          cp dist/* release-assets/
+          cp release-evidence/product-qualification.json \
+            "release-assets/scm-v${VERSION}-qualification.json"
+          tar -czf "release-assets/scm-v${VERSION}-attribution-evidence.tar.gz" \
+            -C quality/evidence .
           (
-            cd dist
+            cd release-assets
             sha256sum ./*
-          ) | sed 's#  \./#  #' > SHA256SUMS
+          ) | sed 's#  \./#  #' > release-assets/SHA256SUMS
 
       - name: Create GitHub release
         uses: softprops/action-gh-release@v2
@@ -168,5 +182,4 @@ jobs:
           draft: false
           prerelease: false
           files: |
-            dist/*
-            SHA256SUMS
+            release-assets/*

CHANGELOG.md

@@ -4,6 +4,39 @@ Format: each release lists what shipped, why it shipped, what tests verified it,
 
 ---
 
+## v1.1.0 - 2026-07-14
+
+**Theme: deterministic contradiction lineage and permanent attribution safety gates.**
+
+- Restricted contradiction updates to the canonical current lineage leaf and
+  added an internal lineage lock around candidate selection and supersession.
+- Defined the canonical current concept as the newest terminal node ordered by
+  `valid_from`, `created_at`, and ID, and made retrieval resolve stale
+  candidates to that concept.
+- Added load/import normalization for branched snapshots. Local state keeps the
+  original snapshot as the known-good backup; PostgreSQL keeps the pre-repair
+  envelope in revision history. Suppressed canonical memories remain
+  suppressed.
+- Added the additive `lineages_repaired` import/load diagnostic without
+  changing snapshot format v2, public SDK signatures, or the five tools.
+- Added credential-free multi-agent attribution qualification across awake,
+  micro/NREM-like, and deep-sleep states, including metadata preservation,
+  user isolation, runtime-owned tool identity, and untrusted-memory delimiting.
+- Published a sanitized aggregate of the 2,400-response external validation
+  while excluding raw paid responses and credentials.
+
+### Verification
+
+- The release-only contradiction gate runs 120 updates in 25 fresh processes
+  and requires zero failures plus retrieval/current-ID agreement in every run.
+- Deterministic tests cover equal-similarity candidates, randomized import
+  order, historical-candidate rejection, concurrent same-user updates,
+  suppression preservation, idempotent repair, restart recovery, and rollback.
+- PostgreSQL qualification verifies revision-locked repair, pre-repair history,
+  restart idempotence, and failed-repair rollback.
+- Installed-wheel and LangChain qualification remain URL-free; REST, MCP, and
+  JavaScript contracts remain backward compatible.
+
 ## v1.0.0 - 2026-07-13
 
 **Theme: embedded lifecycle memory for Python and LangChain agents.**

README.md

@@ -119,7 +119,7 @@ memory.delete_user()  # hard deletion
 ## Advanced Surfaces
 
 REST, MCP, and the HTTP client remain supported for remote or non-Python
-deployments. The JavaScript package is a remote REST client; SCM 1.0 does not
+deployments. The JavaScript package is a remote REST client; SCM 1.1 does not
 claim an embedded JavaScript runtime. See
 [Integrations](docs/INTEGRATIONS.md) and [Deployment](docs/DEPLOYMENT.md).
 
@@ -132,7 +132,9 @@ python scripts/run_product_qualification.py
 The release gate builds and installs the wheel outside the repository, runs
 plain Python and LangChain without an SCM server, checks lifecycle recovery,
 exercises the optional REST and JavaScript surfaces, and audits dependencies.
-See [Quality Gates](docs/QUALITY_GATES.md).
+It also runs deterministic lineage stress and credential-free multi-agent
+attribution gates. See [Quality Gates](docs/QUALITY_GATES.md) and
+[Safety Validation](docs/SAFETY_VALIDATION.md).
 
 ## Documentation
 
@@ -141,6 +143,7 @@ See [Quality Gates](docs/QUALITY_GATES.md).
 - [Product Thesis](docs/PRODUCT.md)
 - [Architecture](docs/ARCHITECTURE.md)
 - [Deployment](docs/DEPLOYMENT.md)
+- [Safety Validation](docs/SAFETY_VALIDATION.md)
 - [Research Artifacts](docs/ARTIFACTS.md)
 
 ## Research

docs/ARTIFACTS.md

@@ -4,7 +4,8 @@ The product repo tracks the paper PDF and this concise artifact manifest. Full
 raw benchmark directories should be attached to the corresponding GitHub
 Release so the Git history stays clean.
 
-Release target: `v0.9.0-product-runtime`
+Paper artifact release: `v0.9.0-product-runtime`
+Product validation release: `v1.1.0`
 Paper DOI: [`10.5281/zenodo.21323877`](https://doi.org/10.5281/zenodo.21323877)
 
 ## Tracked Paper
@@ -44,3 +45,13 @@ Paper DOI: [`10.5281/zenodo.21323877`](https://doi.org/10.5281/zenodo.21323877)
 These artifacts support the paper's lifecycle-memory claim: SCM is useful when
 memory must transform during idle time. They do not claim universal superiority
 over all retrieval systems on all memory workloads.
+
+## V1.1 Product Validation
+
+The `v1.1.0` release adds
+`scm-v1.1.0-attribution-evidence.tar.gz`, a credential-free aggregate archive
+containing the tracked machine-readable attribution summary. It records 2,400
+audited paid-model responses, zero crashes, 100% cross-user isolation, and no
+shared-memory increase in false self-attribution under the tested protocol.
+Raw paid responses are intentionally excluded. See
+[Safety Validation](SAFETY_VALIDATION.md) for the scope and limitations.

docs/GETTING_STARTED.md

@@ -1,6 +1,6 @@
 # Getting Started
 
-SCM 1.0 runs inside the Python process that owns the agent. The default path
+SCM 1.1 runs inside the Python process that owns the agent. The default path
 does not start an SCM service and does not require an SCM endpoint.
 
 ## Plain Python

docs/PUBLISH.md

@@ -9,22 +9,22 @@ included in screenshots.
 
 | Registry | Package | Version |
 |---|---|---|
-| PyPI | `scm-memory` | `1.0.0` |
-| npm | `scm-memory` | `1.0.0` (publish only after registry access is fixed) |
+| PyPI | `scm-memory` | `1.1.0` |
+| npm | `scm-memory` | `1.1.0` metadata only; registry publishing excluded |
 
 The following values must match before a tag is created:
 
 - `pyproject.toml`
 - `sdk/js/package.json`
 - `src/version.py`
 - `CHANGELOG.md`
-- Git tag `v1.0.0`
+- Git tag `v1.1.0`
 
 ## Qualification Gate
 
 ```bash
 venv/bin/python scripts/run_product_qualification.py \
-  --output /tmp/scm-v1.0.0-qualification.json
+  --output /tmp/scm-v1.1.0-qualification.json
 ```
 
 Do not publish if this exits nonzero, reports blockers, or skips a required
@@ -45,13 +45,14 @@ The release workflow uses PyPI trusted publishing. Configure PyPI to trust:
 After the merged release commit passes qualification and CI:
 
 ```bash
-git tag v1.0.0
-git push public v1.0.0
+git tag -a v1.1.0 -m "SCM v1.1.0 reliability release"
+git push public v1.1.0
 ```
 
 The tag workflow reruns qualification, validates tag/version parity, builds
 fresh artifacts, runs `twine check`, and publishes only after those gates pass.
-The GitHub Release receives the wheel, source distribution, and `SHA256SUMS`.
+The GitHub Release receives the wheel, source distribution, portable
+`SHA256SUMS`, qualification JSON, and aggregate attribution-evidence archive.
 Release reruns first check whether the exact package version already exists on
 PyPI. An existing version skips only the immutable upload; new versions still
 require the configured trusted publisher.
@@ -60,40 +61,27 @@ Verify the registry package from a new environment:
 
 ```bash
 python3 -m venv /tmp/scm-pypi-check
-/tmp/scm-pypi-check/bin/pip install "scm-memory[langchain]==1.0.0"
+/tmp/scm-pypi-check/bin/pip install "scm-memory[langchain]==1.1.0"
 SCM_DATA_DIR=/tmp/scm-pypi-data /tmp/scm-pypi-check/bin/scm doctor
 /tmp/scm-pypi-check/bin/python -c \
   "from scm import SCM; m=SCM(user_id='release-check'); m.add_memory('SCM is embedded.'); print(m.search_memory('embedded')['ok']); m.close()"
 ```
 
 ## npm
 
-The npm package is not part of the Phase 1 installation promise. Do not add
-`npm install scm-memory` to public quickstarts until the independent registry
-verification below succeeds.
-
-Before publishing the JavaScript SDK from the same release commit:
+The npm package is excluded from V1.1 because `scm-memory` is not currently
+available through the npm registry. Version metadata remains aligned so the
+remote JavaScript source client can be tested against the local REST contract:
 
 ```bash
 cd sdk/js
 npm test
 npm pack --dry-run --json
-npm publish --access=public
-```
-
-Verify it independently:
-
-```bash
-mkdir -p /tmp/scm-npm-check
-cd /tmp/scm-npm-check
-npm init -y
-npm install [email protected]
-node --input-type=module -e \
-  "import { SCM } from 'scm-memory'; console.log(typeof SCM)"
 ```
 
-The JavaScript package remains a remote REST client in 1.0.0. Do not describe
-it as an embedded lifecycle runtime.
+Do not run `npm publish` or add an npm install command to public quickstarts.
+The JavaScript source remains a remote REST client, not an embedded lifecycle
+runtime.
 
 ## Incident Rule
 

docs/QUALITY_GATES.md

@@ -39,6 +39,11 @@ GitHub Actions uploads the same report as a build artifact.
    canonical OpenAPI routes, and a real JavaScript client lifecycle.
 8. Regression: the complete local pytest suite must exit zero. Skips remain
    visible and are not counted as passes.
+9. Lineage reliability: 25 fresh processes each run 120 contradiction updates;
+   every run must retain one current node and return that same ID in retrieval.
+10. Attribution safety: credential-free single/shared-agent tests cover awake,
+    micro/NREM-like, and deep sleep, source metadata, identity isolation,
+    runtime-bound tool identity, and untrusted-memory delimiters.
 
 ## Current budgets
 
@@ -53,7 +58,7 @@ GitHub Actions uploads the same report as a build artifact.
 - Warm offline embedded add/search and middleware overhead p95: less than
   300 ms, excluding the agent model's own latency.
 
-These tests qualify SCM 1.0's embedded local runtime and PostgreSQL coordination
+These tests qualify SCM 1.1's embedded local runtime and PostgreSQL coordination
 contract. They do not claim multi-region availability or hosted-service SLOs;
 those require deployment infrastructure, traffic replay, network fault
 injection, and long-running soak evidence beyond this repository.

docs/SAFETY_VALIDATION.md

@@ -0,0 +1,60 @@
+# Multi-Agent Attribution Validation
+
+SCM 1.1 converts a post-publication attribution study into permanent product
+qualification gates. The bounded result is:
+
+> SCM showed no shared-memory increase in false self-attribution under the
+> tested protocol.
+
+This is evidence about a defined test matrix, not a universal safety claim.
+
+## External Validation
+
+Two independently audited OpenAI runs evaluated the unchanged SCM lifecycle
+architecture from the July paper. Each run contained 1,200 responses across
+single-agent and shared-agent conditions in awake, NREM-only, and full
+deep-sleep states, plus two provenance interventions.
+
+| Model snapshot | Responses | Crashes | Single deep | Shared deep | Difference |
+|---|---:|---:|---:|---:|---:|
+| `gpt-5.4-mini-2026-03-17` | 1,200 | 0 | 0/300 | 0/300 | 0.0 pp |
+| `gpt-5.4-2026-03-05` | 1,200 | 0 | 0/296 | 0/299 | 0.0 pp |
+
+The denominators in the final two columns are factually correct answers. Both
+runs completed the exact matrix, passed artifact audits and secret scans, and
+maintained 100% cross-user isolation. Neither run met the preregistered gate
+for a shared-memory attribution effect.
+
+The two models interpreted explicit provenance differently. That model effect
+does not establish a provenance repair benefit because shared SCM produced no
+baseline increase to repair.
+
+## Product Gates
+
+Routine CI does not replay paid responses. Credential-free production tests
+instead protect the relevant runtime invariants:
+
+- factual retrieval parity for single-agent and shared-agent use across awake,
+  micro/NREM-like, and deep-sleep lifecycle states;
+- preservation of source, thread, and message metadata through lifecycle
+  transformations;
+- strict user isolation and runtime-owned identity binding;
+- exactly five model-callable tools, none accepting a model-selected
+  `user_id`;
+- delimiting recalled memory as untrusted user-provided data.
+
+The release-only lineage gate also runs 120 contradiction updates in 25 fresh
+processes and requires one deterministic current node plus retrieval/current-ID
+agreement in every process.
+
+## Evidence Boundary
+
+The sanitized aggregate is tracked at
+[`quality/evidence/multi_agent_attribution_2026_07_14.json`](../quality/evidence/multi_agent_attribution_2026_07_14.json).
+The `v1.1.0` GitHub Release includes that file in a small evidence archive.
+Raw paid-response payloads and credentials are excluded.
+
+The validation is limited to neutral fact packs, one shared user-owned memory
+scope, the frozen July implementation, two OpenAI snapshots, and the specified
+lifecycle conditions. It does not prove that all multi-agent systems are free
+from attribution errors, and it does not change the July paper.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "scm-memory"
-version = "1.0.0"
+version = "1.1.0"
 description = "SCM Memory: embedded lifecycle memory for Python and LangChain agents."
 readme = "README.md"
 requires-python = ">=3.10"