| # Paper artifact note — PR-M1 / v0.1.0 |
|
|
| Date: 2026-11-05. Written because `ARTIFACT-SPEC.zh-CN.md` requires this note |
| from M1 onward, and because a reader must not be able to mistake this package |
| for the experiments behind the sealed research report. |
|
|
| ## Identity and availability |
|
|
| | Field | Value | |
| |---|---| |
| | Artifact | `cognitive-runtime` 0.1.0 (PR-M1) | |
| | Wheel | `dist/cognitive_runtime-0.1.0-py3-none-any.whl`, 46,448 B, SHA-256 `64e02c1ab36023f4ae66c6da0d0ba6fba959bd88664d6b37590b3905c71f28e1` | |
| | Source package | `dist/cognitive_runtime-0.1.0.tar.gz`, 51,237 B, SHA-256 `d69ad3ce98d3a63fdf2519724675f599920b4173cf3a4a381bee00b6fc9b0213` | |
| | Public URL | `https://huggingface.co/WayneRR/cognitive-learning-lab` (model repo, published 2026-11-05; initial publication commit `4225068d`; live commit hashes are tracked in `RELEASE-MANIFEST.json` → `publication`). Artifacts are also attached to GitHub Release v0.1.0 of the source repository | |
| | Source-of-truth repository | the private working repository, branch `public-runtime-m1`; this bundle is the public-safe subset. The public source copy lives at `https://github.com/kialajin-l/cognitive-learning-lab` | |
| | Code license | `Apache-2.0` (`LICENSE`), owner decision 2026-11-05; covers the package's code only, not the upstream checkpoint. Package is thin — no weights bundled | |
|
|
| ## How to obtain and run it |
|
|
| ```bash |
| python -m pip install ./dist/cognitive_runtime-0.1.0-py3-none-any.whl # or: pip install . |
| python -m cognitive_runtime doctor --store ./memory.sqlite |
| python -m cognitive_runtime infer --input "$(python -c \ |
| 'import cognitive_runtime.config as c;print(c.EXAMPLES_DIR/"request.json")')" \ |
| --store ./memory.sqlite --namespace demo |
| python -m pytest tests -q --basetemp=<tmp> -p no:cacheprovider # 28 tests, from the source tree |
| ``` |
|
|
| The `--basetemp` and `-p no:cacheprovider` flags were needed on this audit |
| machine (`PermissionError [WinError 5]` on a `pytest-of-<user>` temp directory |
| it cannot write). They are harness workarounds, not part of the package; |
| whether a plain `python -m pytest tests -q` works elsewhere is not verified |
| here. |
|
|
| The default provider is `mock`. The real provider needs the `model` extra and |
| one upstream download (`convaiinnovations/laya@55cf4c4e…`) — or a local |
| directory, which is what the smoke in `acceptance/PR-M1.md` used. Measured |
| correction to earlier text: the `answerdotai/ModernBERT-large` weights are not |
| read at inference on `laya 0.3.20`; see `MODEL_CARD.md`. |
|
|
| ## What this package is a contribution of |
|
|
| An engineering artifact: the runtime boundary the research line kept needing — |
| input sanitising before the gate, integrity-checked memory, retrieval, a |
| fail-closed typed rule judgment, a three-layer boundary that separates a model |
| proposal from an authorised answer, receipt/lineage verification, controlled |
| writeback, and cross-process replay — packaged so it installs and runs without |
| the private repository, and without any answer keys or task sets. |
|
|
| ## Correspondence with the sealed experiments |
|
|
| The frozen research record for this mechanism is |
| `docs/LSG-SG4-R3-D708-FINAL-VERDICT-v1.zh-CN.md` (2026-10-29) inside |
| `docs/LSG-SG4-R3-PROJECT-BOUNDARY-CLOSEOUT-SCOPE-v1.zh-CN.md`, status |
| `PROJECT_SCOPE_CLOSED`. Mapping of ideas, not of numbers: |
|
|
| | Sealed experiment line | What it studied | What this package does with that | |
| |---|---|---| |
| | R23–R25, R27, R35–R37 memory participating in decisions | retrieval + evidence into the decision | reimplemented as `retrieval.py` + `judgment.py`, on a new synthetic task domain | |
| | R16–R19, R29–R30 correct writeback and reuse | verified writeback surviving replay | reimplemented as `pipeline._writeback` + `lineage-consistency-v1` | |
| | R26/R28, R40 rejecting contaminated memory and conservative abstention | pollution rejection, refuse when undecidable | reimplemented as fail-closed `select_rule` reasons and `sanitize.py` refusal of caller trust fields | |
| | R39/R44/R46/R47 conflict handling | multi-source conflict, partially unresolved | only the narrow single-assignment conflict is implemented: two promoted rules that disagree → `conflicting_verified_rules`, no answer, no writeback | |
| | R31–R34, R37, R48 open-web acquisition | autonomous search, `3/8` admission on open web | **not implemented** — that is PR-M3 scope, and the sealed evidence does not support it being reliable | |
| | long-horizon positive gain | explicitly unverified | explicitly not claimed; nothing here measures learning over time | |
|
|
| ## Differences a reader must not overlook |
|
|
| 1. **Different code path.** The sealed numbers were produced by archived |
| `runners/` scripts against sealed task sets and local-state model assets. |
| This package re-implements the boundary in `cognitive_runtime` and has never |
| been run on those task sets. Equal semantics are not assumed, and no |
| R-number result is re-stated as a property of this package. |
| 2. **Different evidence class.** The acceptance evidence in |
| `acceptance/PR-M1.md` comes from the mock provider plus synthetic |
| `Directory record: X is assigned to Y` fixtures. The single real-model |
| receipt (2026-11-05, offline CPU, `provider_class="real"`) is a load-and-run |
| smoke on one of those same synthetic examples — it proves the pinned |
| checkpoint drives the chain end-to-end; it is **not** an accuracy result and |
| adds no evidence toward any paper claim. This package has never been run on |
| the sealed task sets. |
| 3. **No answer keys, no trust root.** The paper experiments used held-out |
| answer keys in a separate evaluation path. This package refuses fields that |
| look like answer keys or trust assertions (`gold`, `expected`, `verified`, |
| `trusted`, `ground truth`) at the input boundary, and ships no signing key — |
| verification means hash + cited receipt + lineage, which is provenance |
| consistency, not truth. |
| 4. **Scope of "memory growth".** The closed project's question was whether |
| memory can participate in decisions and show observable gain, plateau or |
| degradation under a controlled loop. This package answers the narrower |
| engineering question of whether that boundary can be installed and |
| reproduced at all. Gain is not measured here. |
|
|
| ## Reading order for a reviewer |
|
|
| `README.md` → `acceptance/PR-M1.md` (what actually passed, what did not) → |
| `SOURCE-INVENTORY.md` (licenses, revisions, hashes, audit gaps) → |
| `MODEL_CARD.md` (model identity and limitations) → the sealed verdict document |
| above for the research claims, which this note neither repeats nor extends. |
|
|
| M2 and M3 will update this note with their own version, run method and |
| correspondence changes. This note does not modify any sealed experiment, |
| figure, or research conclusion, and adds no new D/R numbering. |
|
|
