开源AI工作流：Epistemic AI技能

Install, Package, and Runtime Use

Claude Installation

1. Package the skill from this repository.
2. Open Claude.ai and go to Settings > Skills, or open Claude Skills.
3. Upload daee-epistemics.skill.
4. Enable the skill and test it with a query that should trigger epistemic diagnosis or objection handling.

Maintainer Optional Route/Check Harness

The optional script-capable route/check harness is repo/dev machinery. It can make routing executable rather than merely interpretive for maintainers, but it is not the public identity of the skill, is not packaged in the canonical user-facing artifact, and is not the active push/PR execution path. Some implementation files still use the historical level3 / daee_level3 names until a deliberate harness-rename migration:

python skill/scripts/daee_level3.py --input <input.md> --out <run-dir>

It produces features.json, route_plan.json, reconstruction/validation verdicts, and, when the route is valid, execution_prompt.md. If route generation writes execution_blocked.md, return that visible PARTIAL/block note and do not execute an ordinary answer. After the model answers from a valid execution prompt, validate the answer against the route plan:

python scripts/check_execution.py --route <run-dir>/route_plan.json --output <run-dir>/output.md

If execution validation returns partial or fail, a maintainer-facing report should include:

PARTIAL - script-harness execution check: <specific defect>

Honest maintainer claim: the repository includes deterministic routing (route.py) over span-backed feature extraction (diagnose.py), with route plans validated by reconstruction (reconstruct.py) and post-output execution checked (check_execution.py). That route/check harness is excluded from the canonical user-facing package unless a future dev artifact is explicitly created. Routing is deterministic given features. Feature extraction has interpretive components with input-span validation. Transformer execution remains probabilistic; highest-complexity burdens remain bounded by model capability even under optional script-harness routing. Pure-Hermes parity, fixture-18 resolution, catalogue-wide executable harness coverage, codons, and owner packs are not claimed.

Integration Boundary

New background material should be integrated only when it improves routing, restoration, scope control, or terminology discipline.

The goal is not to accumulate study notes.

The goal is to extract durable distinctions and convert them into reusable architecture: new Case Modules, tighter Tactic or Technique criteria, clarified Glossary entries, sharper confidence rules, or better routing boundaries.

If material does not alter how the skill classifies, sequences, or restores, it should usually stay outside the live skill surface.

Package Boundary

The canonical user-facing upload name is daee-epistemics.skill. For the v0.4.x release line, the package artifact is built from atomics through generated local/CI skill/ and recorded in docs/release-artifacts.md. GitHub Releases are the binary distribution surface; older v0.3.1.0 assets and smokes are historical evidence, not current-package evidence for v0.4.0.0. package.ps1 emits a local .skill.zip archive because it is a zip payload with the skill root at archive root. Publish/upload the same checked payload as .skill; do not publish both .skill.zip and .skill, and do not re-zip it.

Binary skill archives and generated skill/ runtime output are not committed to this repository. Build locally or in CI from atomics/skill/** into generated skill/, then package that runtime, or use the verified public GitHub Release asset.

The canonical archive root must contain SKILL.md, references/, compiled-module-map.json, build-manifest.json, and README.md directly. It must not contain the repo/dev harness roots data/, scripts/, or tests/. This returns the user-facing package shape to the scriptless runtime boundary used before the optional route/check harness was introduced, while retaining the harness in the repository for maintainer validation. Do not zip the whole repo root, and do not produce a bundle whose top level is skill/. Package the canonical contents selected from the generated skill/ directory, not the directory itself.

For path fidelity, build the archive with the manifest-backed package script. It validates the generated skill/ tree, rejects unexpected packageable files, and writes slash-safe archive entry names for skill hosts that inspect the bundle structure directly.

On Windows, package.ps1 calls the Python packager in tools/package_skill.py. For a v0.4.0.0 local package rebake, the command is:

powershell -NoProfile -ExecutionPolicy Bypass -File .\package.ps1 build\daee-epistemics-v0.4.0.0.skill.zip
Copy-Item build\daee-epistemics-v0.4.0.0.skill.zip build\daee-epistemics-v0.4.0.0.skill

From any folder, open a Bash-compatible terminal and paste the following if you want a clone-and-package flow. The command clones the repo into a temporary subfolder, builds daee-epistemics.skill from the generated skill/ contents, and removes the temporary clone so the folder you opened ends with only daee-epistemics.skill.

repo="https://github.com/theislampill/daee-epistemics.git"
tmp="daee-epistemics-src"
tmp_zip="daee-epistemics.tmp.skill.zip"
out_skill="daee-epistemics.skill"

rm -rf "$tmp" "$tmp_zip" "$out_skill"
git clone "$repo" "$tmp" &&
(cd "$tmp" &&
  python tools/build_framework_pipeline.py &&
  python tools/build_compiled_runtime.py &&
  python tools/check_compiled_runtime_freshness.py &&
  python tools/check_package_shape.py &&
  python tools/package_skill.py "../$tmp_zip") &&
mv -f "$tmp_zip" "$out_skill" &&
rm -rf "$tmp"

If you open daee-epistemics.skill, you should see SKILL.md, references/, compiled-module-map.json, build-manifest.json, and README.md at the top level of the archive. You should not see data/, scripts/, tests/, atomics/, tools/, docs/, build/, .git/, smokes/, .daee/, level3-runs/, __pycache__/, route_plan.json, features.json, validation.json, reconstruction.json, execution_verdict.json, execution_prompt.md, execution_blocked.md, partial_banner.md, retry_prompt.md, output.simulated.md, or output.model.md.