Quantum Reasoning Skill

A model-facing Agent Skill that keeps several genuinely different hypotheses alive, tests each against evidence and tools, revives alternatives when new evidence arrives, and collapses to the best-supported answer only at the end. It is a metaphor, not quantum computing — "measurement" means letting evidence collapse a branch.

Status: protocol implemented, controller benchmark published, model evaluation pending. The skill contract, the reference controller, the benchmark harness and a deterministic falsification benchmark are here and run. A baseline-vs-skill model evaluation has NOT been published, so this project makes no claim that it improves any language model's accuracy. See Benchmark status before citing any number from here.

Quick start

SKILL.md is the deliverable. Three verified paths reach it:

# 1. from a clone — nothing to install
git clone https://github.com/Furox-Art/quantum-reasoning-skill.git
cd quantum-reasoning-skill
python docs/quickstart.py

# 2. from PyPI — SKILL.md ships inside the wheel
pip install quantum-reasoning-skill
quantum-reasoning --validate

# 3. from npm — also published
npm i quantum-reasoning-skill
npx quantum-reasoning --validate

Copy SKILL.md into your host's skill directory, keeping the YAML front matter. docs/getting-started.md walks the install; prompt shapes are in examples/usage.md, host integration shapes in examples/host-integration.md, and required host capabilities in docs/COMPATIBILITY.md.

docs/quickstart.py verifies the checkout, parses the front matter, exercises the controller and smoke-tests the benchmark harness; it calls no model and invents no metrics. Its verified output is in docs/quickstart.md, which is the single source for it — including the scenario where the thresholds refuse to collapse, the clearest evidence they are uncalibrated defaults. It is not duplicated here, because a pasted copy drifts.

Distribution and the surface it exposes

Three channels ship SKILL.md, all on one version: a git clone (everything); the PyPI wheel (SKILL.md + VERSION as package data, plus reference/, benchmark/, docs/, examples/, and a quantum-reasoning console script); and the npm tarball (SKILL.md, VERSION, index.js, LICENSE, README.md, and a quantum-reasoning bin). Both packaged artifacts are verified against the built artifact, not a checkout.

Supply-chain attestations

Only some published versions carry a build attestation, so check before relying on one:

1.1.1 was published through the long-lived-token path, which cannot produce a Sigstore attestation. Attestations are bound to a publish event and cannot be added afterwards, so this is permanent for that version. The package itself is unchanged and passes the same release contract as every other version; only the provenance record is missing.

If you need to verify how a build was produced, use 1.1.0 on npm. To check an attestation yourself:

curl -sS -o /dev/null -w '%{http_code}\n' \
  https://registry.npmjs.org/-/npm/v1/attestations/quantum-reasoning-skill@1.1.0   # 200
curl -sS -o /dev/null -w '%{http_code}\n' \
  https://registry.npmjs.org/-/npm/v1/attestations/quantum-reasoning-skill@1.1.1   # 404

A missing attestation is not evidence of tampering. Versions are locked across VERSION, pyproject.toml, package.json, CITATION.cff and CHANGELOG.md — read the value from VERSION, not a badge (details).

There is no ReasoningSession class. An earlier README claimed one, it never existed, and asserting it made the release gate unsatisfiable. The real surface is two non-compatible CLIs plus a library:

--path is npm-only; --skill-file and --quiet are Python-only. An unsupported flag exits 2.

Release contract

bin/check_release_contract.py is the gate CI runs before any release. It asserts that SKILL.md resolves inside the installed wheel, that the quantum-reasoning entry point is installed and succeeds against the installed asset, that reference.branch_controller imports with a callable API, and that the version agrees everywhere.

Twelve public names in reference/branch_controller.py are frozen — DEFAULT_THRESHOLDS, Branch, BranchMetrics, BranchState, Thresholds, branch_score, classify_branch, collapse_decision, diversity_ratio, recommended_width, uncertainty_from_branches, update_branch_state. Renaming or removing any one halts the release. WEIGHTS is deliberately not on the list: the scoring model can change without breaking the contract.

A negative-control mode proves the gate is fail-closed across four scenarios — removing the installed SKILL.md, emptying it, and feeding the controller check a stub missing its public API and a module returning wrong values — requiring each to be rejected. It does not exercise VERSION or the console entry point.

Reference controller

The control loop behind the protocol: it scores branches, moves them between active / dormant / rejected, revives recoverable ones, and decides when a search may collapse. It never generates reasoning and never calls a model — it only consumes measurements you supply.

Weights and thresholds are reference defaults, not validated constants — docs/MEASUREMENT.md gives the formulas and the list of values that must be calibrated first.

Benchmark harness

benchmark/ measures whether the skill helps rather than merely producing more text: benchmark/evaluate.py (evaluator), benchmark/validate_submission.py (machine-validates a reproducible bundle), benchmark/paths.py (path confinement that rejects .. traversal and symlink escapes), benchmark/cases.jsonl, and four JSON Schemas — case, result, metadata, comparison.

python benchmark/evaluate.py --cases benchmark/cases.jsonl --skill skill-results.jsonl

Inputs are confined to --cases-dir (default: each input's own directory), --output must stay inside it, and reads are bounded by --max-bytes. Both python benchmark/evaluate.py and python -m benchmark.evaluate work.

Tests, tooling and governance

About 43% of the Python here is tests — 1,575 of 3,635 lines across 9 test modules, run with the standard library's unittest on Python 3.10–3.14 on Linux, Windows and macOS. They cover the controller, the benchmark harness, import hardening, path boundaries, dependency-lock drift and release-gate polling; tests/test_readme_badges.py keeps this file's badges honest. The toolchain is pinned twice — pyproject.toml's dev group with ==, and constraints.txt with == and --hash digests for the full transitive closure (details).

Also here: codemeta.json for machine-readable citation beside CITATION.cff; five issue templates (bug, feature, question, benchmark result, plus a chooser in config.yml) and a pull-request template; a Code of Conduct with a scientific-integrity section; a Security policy.

The protocol

SKILL.md is the normative contract. In short: frame the problem (verified vs assumed) → open branches (genuinely different, not paraphrases) → evaluate independently (on evidence, not eloquence) → cross-check (independently reached conclusions get reinforced) → prune and revive (kill contradicted paths, keep dormant ones recoverable) → collapse (only when one path is clearly better supported). SKILL.md also scopes when not to use it: trivially answerable questions should not pay branching cost.

Benchmark status

What has been measured — the controller, not the model. On 9 deterministic falsification scenarios, the multi-hypothesis branch controller achieved 100% final accuracy versus 22% for a single-hypothesis policy that commits to its first leader (0/4 on falsification scenarios, 0/3 on revival scenarios; both policies scored 2/2 on the control scenarios). It abandoned the falsified hypothesis at the same evidence step the falsification arrived, and revived a rejected-then-supported alternative in 3/3 revival scenarios. Confidence calibration showed no difference (Brier score 0.033 vs 0.033), and the accuracy gain cost 109.6% more tokens per scenario. Methodology and reproduction: benchmark/falsification/METHODOLOGY.md.

These results measure the reference decision controller on synthetic evidence streams, not any language model. benchmark/cases.jsonl holds six deterministic smoke-test cases that prove the harness runs; they are explicitly not sufficient evidence of reasoning improvement. A baseline-vs-skill model evaluation has not been published, so this project makes no claim that it improves any language model's accuracy or costs less compute. Anyone can evaluate on models they have access to; every submitted number must carry provenance (script, skill commit, provider and exact model version, run date, sampling parameters, repetitions, platform), and one positive run establishes no general claim. Rules: CONTRIBUTING.md.

How it compares

This table describes mechanisms, not measured performance. Nothing here has been benchmarked against the rows above it.

Documentation

Everything lives in docs/index.md — getting started, quickstart and its verified output, measurement methodology, host compatibility, report forms. Release history: CHANGELOG.md. License: MIT.