A model-agnostic reasoning skill that keeps multiple genuinely different possibilities alive, verifies and compares them, suppresses weak paths, revives useful alternatives when evidence changes, and selects the best-supported result only at the end.
This is quantum-inspired classical reasoning, not quantum computation.
The project is centered on SKILL.md. The skill changes the model's reasoning procedure without retraining the model.
problem
-> open diverse possibilities
-> verify independently
-> compare contradictions and shared assumptions
-> score / classify branches
-> allocate compute dynamically
-> suspend or reject weak branches
-> revive branches when evidence changes
-> verify collapse criteria
-> select the best-supported answer
- Reduce early lock-in to the first plausible answer.
- Maintain genuinely independent hypotheses or solution strategies.
- Prefer falsification and tool-based verification over self-confidence.
- Expand search when uncertainty is high and concentrate compute when evidence is strong.
- Detect correlated branches so repeated assumptions are not counted as independent evidence.
- Keep high-information alternatives recoverable instead of deleting them too early.
- Work across hosts that can preserve the skill instruction contract.
- Make reasoning-control decisions measurable without exposing private chain-of-thought.
SKILL.md— model-facing protocolreference/branch_controller.py— deterministic reference scoring, state transition, revival, uncertainty and collapse logicdocs/MEASUREMENT.md— explicit formulas, thresholds and calibration requirementsdocs/COMPATIBILITY.md— capability-based host compatibility contractexamples/usage.md— task examplesexamples/host-integration.md— generic host integration patternsbenchmark/cases.jsonl— deterministic seed casesbenchmark/evaluate.py— baseline-vs-skill evaluatorbenchmark/validate_submission.py— reproducible community-result bundle validatorbenchmark/schemas/— JSON Schema contracts for cases, metadata, result rows and comparison outputbenchmark/README.md— reproducible benchmark protocolCONTRIBUTING.md— contribution and independent benchmark-submission policySECURITY.md— vulnerability-reporting policyCITATION.cff— citation metadatatests/— behavioral and validation tests- GitHub Actions validation on pushes and pull requests
- CI-gated automated tag and GitHub Release publishing
- Adaptive multi-branch reasoning
- Semantic diversity requirements
- Active / dormant / rejected branch states
- Independent verification rules
- Contradiction and shared-assumption checks
- Correlation penalty for duplicated/shared-assumption branches
- Interference-like cross-branch comparison
- Dynamic compute allocation from uncertainty
- Measurable branch revival rules
- Explicit stopping / collapse criteria
- Baseline-vs-skill benchmark telemetry
- Final requirement and contradiction check
Install or provide SKILL.md to a compatible agent/skill system, then invoke it for difficult reasoning tasks. No Python package or model fine-tuning is required to use the skill itself.
See docs/COMPATIBILITY.md before integrating a new host and examples/host-integration.md for generic installation patterns.
The Python reference implementation is optional. It exists to make the qualitative policy auditable and testable:
from reference.branch_controller import Branch, BranchMetrics, collapse_decisionThe repository provides the protocol, cases, schemas, evaluator and submission validator needed to test the skill. Real model evaluations are intentionally community-run: users test the models/providers they have access to and may submit reproducible results back to the project.
The project does not require the maintainer to run every commercial or local model, and it does not treat the absence of maintainer-run model tests as a missing implementation feature.
A valid comparison uses the same model, model version, task set, tool access, temperature/sampling settings and token-budget policy with and without SKILL.md.
python benchmark/evaluate.py \
--cases benchmark/cases.jsonl \
--baseline path/to/baseline-results.jsonl \
--skill path/to/skill-results.jsonl \
--output comparison.jsonCommunity PR bundles are machine-checked. They must contain their own cases.jsonl, raw baseline and skill JSONL files, metadata, the generated comparison output and a short README. CI recomputes the comparison and rejects tampered or incomplete bundles.
python benchmark/validate_submission.py \
--root benchmark/results/community \
--allow-emptyThird-party results remain measurements from their submitters, not automatic project endorsements or universal performance claims. A single positive run is not enough to claim general improvement.
Users can submit results through the Benchmark result issue template or as a reproducible benchmark-result pull request. See benchmark/README.md and CONTRIBUTING.md.
Run the repository checks locally with:
python -m unittest discover -s tests -v
python benchmark/evaluate.py --help
python benchmark/validate_submission.py --root benchmark/results/community --allow-emptyCI validates the skill contract, local links, Python syntax, benchmark seed data, JSON Schema files, community benchmark bundles and behavioral tests. Third-party GitHub Actions are pinned to exact commit SHAs.
Release publishing is repository-native and CI-gated. Update CHANGELOG.md, then change the root VERSION file to a semantic version. A release is considered only after the validate-skill workflow succeeds on main; the publish-release workflow then creates the corresponding v<version> tag and GitHub Release if it does not already exist.
v0.3.1 — hardened measurable prototype. The branch-control mechanism, benchmark infrastructure, host compatibility contract, community-result validation, repository security/citation metadata and CI-gated release controls are implemented. Real-model evaluation is intentionally delegated to independent users and contributors, and the project makes no universal performance claim without reproducible external evidence.