Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Strongly-typed multiagent ant-robot colony exemplar — ADTs, session-typed protocols, affine-discipline resource handles, storage-as-functor framing, Active-Inference-flavored decision loop, mypy-as-oracle negative controls.
.claude/skills/docxology-template-formal/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 3% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 4% | 0% |
| case-01 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -11% | 0% |
Project-scoped skill for the in-repo exemplar at projects/templates/template_formal/. Load this when working inside the project.
template_formal exemplar — running scripts, editingsrc/template_formal/, or adding a new typed invariant.
subject is illegal-state-unrepresentable design or a decentralized multiagent simulation.
every ADT/session-type/affine-handle claim needs both a tests/mypy_fixtures/ negative-control fixture (mypy rejects the bad program) and a runtime unit test (the discipline actually raises).
testing, no compile-time linear/dependent-type claims) still hold after changes.
bash# From the repository root uv run pytest projects/templates/template_formal/tests --cov=projects/templates/template_formal/src --cov-fail-under=90 MYPYPATH=projects/templates/template_formal/src \ uv run mypy --strict --explicit-package-bases --namespace-packages \ projects/templates/template_formal/src uv run python projects/templates/template_formal/scripts/02_run_analysis.py uv run python scripts/pipeline/stage_03_render.py --project templates/template_formal uv run python scripts/pipeline/stage_04_validate.py --project templates/template_formal uv run python scripts/pipeline/stage_05_copy.py --project templates/template_formal # Optional formal side-specs (non-default; require lake/elan + a Java runtime) projects/templates/template_formal/scripts/check_formal_specs.sh
has neither. Affine handles (TransactionHandle, protocol-phase objects) are a runtime-guarded discipline — they raise on reuse, they are not type-checked against reuse. Grep for "dependent type" and "linear type" outside the manuscript's explicit limitations/scoping section must return zero matches.
negative-control fixture under tests/mypy_fixtures/ proves what mypy --strict rejects; a runtime unit test proves what actually raises. One without the other is an unfalsifiable claim.
src/template_formal/,not in scripts/.
tmp_path), a realin-process fault-injectable bus, and real mypy --strict subprocess invocations — never MagicMock/unittest.mock/mocker.patch.
Agent owns exactly one on-disk databaseand one protocol endpoint; the colony coordinator only ever holds the public Agent/PheromoneField interfaces.
formal/lean/or formal/tla/ ship, they must be reachable from scripts/check_formal_specs.sh — never a vestigial, unwired file.
AGENTS.mdREADME.mdTODO.mdprojects/AGENTS.md| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-05 | fail→pass | 12,398 | 10,623 | -14% | 1 | 1 | 0% | 2,007 | 2,059 | +3% | 0 | 0 | — |
case-03 | fail→fail | 9,157 | 7,359 | -20% | 1 | 1 | 0% | 1,583 | 2,325 | +47% | 0 | 0 | — |
case-04 | fail→pass | 13,889 | 6,911 | -50% | 1 | 1 | 0% | 1,959 | 2,034 | +4% | 0 | 0 | — |
case-01 | fail→pass | 27,996 | 8,716 | -69% | 1 | 1 | 0% | 2,369 | 2,700 | +14% | 0 | 0 | — |
case-02 | fail→pass | 7,169 | 3,548 | -51% | 1 | 1 | 0% | 1,287 | 1,627 | +26% | 0 | 0 | — |
case-06 | fail→pass | 11,872 | 5,611 | -53% | 1 | 1 | 0% | 1,988 | 1,774 | -11% | 0 | 0 | — |
case-07 | pass→pass | 18,250 | 16,390 | -10% | 1 | 1 | 0% | 1,966 | 2,205 | +12% | 0 | 0 | — |
case-08 | fail→fail | 14,040 | 9,312 | -34% | 1 | 1 | 0% | 2,087 | 2,328 | +12% | 0 | 0 | — |
case-09 | fail→pass | 9,520 | 6,788 | -29% | 1 | 1 | 0% | 1,396 | 1,966 | +41% | 0 | 0 | — |
case-10 | fail→pass | 7,337 | 4,727 | -36% | 1 | 1 | 0% | 1,226 | 1,785 | +46% | 0 | 0 | — |
case-11 | fail→pass | 9,005 | 1,549 | -83% | 1 | 1 | 0% | 1,434 | 1,076 | -25% | 0 | 0 | — |
case-12 | fail→fail | 29,282 | 4,441 | -85% | 1 | 1 | 0% | 2,140 | 1,198 | -44% | 0 | 0 | — |
case-13 | fail→pass | 10,261 | 4,333 | -58% | 1 | 1 | 0% | 1,778 | 1,649 | -7% | 0 | 0 | — |
case-14 | fail→pass | 8,109 | 1,913 | -76% | 1 | 1 | 0% | 1,495 | 1,215 | -19% | 0 | 0 | — |
case-15 | pass→fail | 13,284 | 4,277 | -68% | 1 | 1 | 0% | 2,301 | 1,408 | -39% | 0 | 0 | — |
case-16 | fail→pass | 11,722 | 4,442 | -62% | 1 | 1 | 0% | 1,761 | 1,571 | -11% | 0 | 0 | — |
case-17 | fail→pass | 32,312 | 4,261 | -87% | 1 | 1 | 0% | 1,413 | 1,566 | +11% | 0 | 0 | — |
case-18 | fail→pass | 5,658 | 5,209 | -8% | 1 | 1 | 0% | 844 | 1,764 | +109% | 0 | 0 | — |
case-19 | pass→pass | 17,072 | 8,563 | -50% | 1 | 1 | 0% | 3,255 | 2,493 | -23% | 0 | 0 | — |
case-20 | pass→pass | 15,558 | 13,255 | -15% | 1 | 1 | 0% | 3,397 | 3,680 | +8% | 0 | 0 | — |
case-21 | pass→pass | 14,986 | 12,039 | -20% | 1 | 1 | 0% | 3,054 | 3,411 | +12% | 0 | 0 | — |
case-22 | pass→pass | 9,928 | 6,190 | -38% | 1 | 1 | 0% | 2,088 | 2,085 | -0% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +55 percentage points is the difference between those two pass rates over the 21 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.