[codex] simplify provider-neutral negative knowledge module

.github/workflows/ci.yml

@@ -0,0 +1,30 @@
+name: CI
+
+on:
+  push:
+  pull_request:
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.9", "3.12", "3.14"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - run: python -m pip install .
+      - run: python -m unittest discover -s tests
+      - run: python examples/quickstart.py
+
+  package:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - run: python -m pip install build
+      - run: python -m build

.gitignore

@@ -6,6 +6,9 @@ __pycache__/
 .Python
 *.egg-info/
 .pytest_cache/
+.coverage
+build/
+dist/
 
 # Virtual env (do NOT commit; reproduce-time only)
 .venv/
@@ -22,12 +25,5 @@ env/
 *.swp
 *.swo
 
-# Local outputs that get regenerated
-# (keep canonical artifacts but ignore replay scratch)
-# (Actually we WANT to keep pred_results/*.npy for inspection; comment if you want exclude)
-
-# Symlinks
-.venv
-
 # Logs that exceed reasonable size
 *.gigantic.log

CITATION.cff

@@ -0,0 +1,22 @@
+cff-version: 1.2.0
+message: "If you use this work, please cite the paper below."
+title: "Negative Knowledge as Failure-aware Shared Memory for AutoResearch"
+version: "0.2.0"
+license: MIT
+authors:
+  - family-names: Wang
+    given-names: Hanchun
+    affiliation: "Department of Applied Mathematics and Theoretical Physics, University of Cambridge"
+    email: hw660@cam.ac.uk
+url: "https://github.com/hch-wang/Negative_Knowledge"
+repository-code: "https://github.com/hch-wang/Negative_Knowledge"
+preferred-citation:
+  type: conference-paper
+  title: "Negative Knowledge as Failure-aware Shared Memory for AutoResearch"
+  authors:
+    - family-names: Wang
+      given-names: Hanchun
+      affiliation: "Department of Applied Mathematics and Theoretical Physics, University of Cambridge"
+  collection-title: "ICML 2026 AI4Research Workshop on AI as a Tool for Mathematics, Computer Science, and Machine Learning"
+  year: 2026
+  url: "https://github.com/hch-wang/Negative_Knowledge"

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Hanchun Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,82 +1,47 @@
 # Negative Knowledge
 
-Code and reproducibility artifacts for the ICML 2026 AI4Research Workshop
-paper *Negative Knowledge as Failure-aware shared memory for AutoResearch*.
+A single-file Python module that turns failed attempts into reusable memory.
+It has no dependencies and works with any model backend.
 
-A *negative-knowledge record* (NK) is a structured, machine-readable
-summary of one or more failed attempts on a task.
+## Install
 
----
-
-## The bounded failure schema
-
-The NK record is a JSON object with a fixed shape. Every field draws
-from a controlled vocabulary; no field is free text in a way that
-would defeat downstream machine consumption. The base form has 6
-required fields:
-
-```
-task_id, attempted_route, observation,
-failure { layer, scope, degree, recommended_action, risk },
-rationale, recommended_alternative
-```
-
-A **depth-N** extension adds three cross-round fields
-(`rounds_summary`, `ruled_out_routes`, `synthesised_diagnosis`)
-when the curator reads more than one failed round.
-
-The schema is implemented in
-[`section3_reproduce/nk_curator.py`](section3_reproduce/nk_curator.py)
-(constants `SCHEMA_BASE_FIELDS`, `SCHEMA_DEEP_EXTRA`, `LAYERS`, …),
-with `validate_nk(nk, depth)` returning a list of violations.
-
----
-
-## Running the §3 experiment
-
-### Verify mode (no API key, ~5 seconds)
+Copy [`negative_knowledge.py`](negative_knowledge.py) into your project, or:
 
 ```bash
-cd section3_reproduce
-python analyze_results.py
+python3 -m pip install git+https://github.com/hch-wang/Negative_Knowledge.git
 ```
 
-Re-derives every §3 paper claim from the bundled archive in `logs/`.
-Expected output: `31/31 claims match`. Full report at
-`section3_reproduce/results/claim_report.md`.
+## Use
 
-### End-to-end mode (Anthropic API, ~$1, ~10 min)
+```python
+from negative_knowledge import append, curate, load
 
-```bash
-cd section3_reproduce
-pip install anthropic
-export ANTHROPIC_API_KEY=sk-ant-...
-git clone https://github.com/OSU-NLP-Group/ScienceAgentBench
-export SAB_BENCH=$(pwd)/ScienceAgentBench/benchmark
+def backend(prompt: str) -> str:
+    return your_agent.generate_json(prompt)
 
-python run_pipeline.py --task 072 --use-saved-trace
-```
+record = curate(
+    backend,
+    task_id="072",
+    task="Map Sub01 EEG signals to Sub03",
+    evidence={
+        "code": failed_code,
+        "error": stderr,
+        "reasoning": reasoning,
+    },
+)
 
-Re-runs the depth-3 deepNKR pipeline on task_072 (the §3 breakthrough
-task): dispatches the canonical curator, builds a fresh Sonnet solver
-sandbox, runs `candidate.py` + the task evaluator, prints the result.
-A `score=1` reproduces the §3 task_072 PASS.
-
-Full guide:
-[`section3_reproduce/README.md`](section3_reproduce/README.md).
+append("negative_knowledge.jsonl", record)
+memory = load("negative_knowledge.jsonl")
+```
 
----
+That is the whole interface:
 
-## Citation
+- `curate(...)` creates and validates one record.
+- `validate(record)` returns schema problems.
+- `append(path, record)` stores one record as JSONL.
+- `load(path)` reads the memory back.
 
-```bibtex
-@inproceedings{NegativeKnowledge2026,
-  title  = {Negative Knowledge as Failure-aware shared memory for AutoResearch},
-  author = {...},
-  booktitle = {ICML 2026 AI4Research Workshop},
-  year   = {2026},
-}
-```
+See [`examples/quickstart.py`](examples/quickstart.py) for an offline example.
+The paper artifacts remain under [`reproduction/`](reproduction/).
 
-ScienceAgentBench: Chen et al., 2024
-(https://github.com/OSU-NLP-Group/ScienceAgentBench).
+MIT licensed.