QuantumMindLite

QuantumMindLite is a lightweight sequential workflow for producing structured, source-audited hypotheses about whether a represented public problem supports an asymptotic quantum-speedup claim at a stated scope.

Completed runs can also be compiled into a Quantum Acceleration Evidence Graph (QAEG). This is an offline, deterministic sidecar: it reads the saved public state.json and decision.json plus the runtime registry, and it neither changes the workflow nor calls an LLM or API.

It does not prove a new quantum algorithm, end-to-end implementability, optimality, or global novelty. The deterministic validator proves only a machine-tested consistency/safety contract relative to the curated primitive registry, the public problem representation, action ownership, and absence of gold/evidence data. The formal boundary is in docs/soundness.md.

Install

python -m pip install -e ".[dev]"
python -m pip install -e ".[dev,live]"
python -m build

The package is quantummindlite; the distribution is quantummind-lite. Prompts, primitive/source catalogs, and PaperBench fixtures are packaged under quantummindlite.resources, so CLI commands work from outside the repository after editable or wheel installation.

From Anaconda Prompt, using the project environment:

conda activate quantummind
cd /d %USERPROFILE%\projects\quantummindlite
python -m pip install -e ".[dev]"

Run the full engineering checks in that environment:

python -m ruff format --check src tests
python -m ruff check src tests
python -m mypy src tests
python -m pytest -q
python -m quantummindlite.cli validate-paperbench
python -m quantummindlite.cli benchmark-all --provider mock --output-dir runs\mock_qaeg_smoke
python -m build

The mock benchmark is a fixture/plumbing smoke test, not a model experiment or performance claim.

Architecture

• One deterministic Orchestrator.
• One generic Agent.
• One typed Python ActionSpec table for action order, role, prompt, schema, readable context, and merge policy.
• Strict Pydantic action outputs plus explicit monotone state merges.
• Deterministic B validator with exactly ten public checks.
• D maps validator verdicts to routes and never upgrades them.
• Offline QAEG compiler/verifier over completed runs. It is a sidecar, not a graph workflow, planner, GNN, or learned judge.
• Downward-only QAEG-screen-v0.1 research screening after the unchanged graph verifier.

CLI

python -m quantummindlite.cli analyze --input C:\path\to\public-case.yaml --provider mock --output-dir runs
python -m quantummindlite.cli benchmark --case-id QM-PB-001 --provider mock --output-dir runs
python -m quantummindlite.cli benchmark-all --provider mock --output-dir runs
python -m quantummindlite.cli benchmark-family --family-id PB-001-family --provider mock --output-dir runs
python -m quantummindlite.cli benchmark-family --family-id PB-006-family --provider mock --output-dir runs
python -m quantummindlite.cli validate-paperbench
python -m quantummindlite.cli freeze-paperbench --confirm
python -m quantummindlite.cli inspect-run --run-dir runs\<run_id>
python -m quantummindlite.cli count-loc

--provider mock is the default. Live calls are explicit and disabled unless QUANTUMMINDLITE_LIVE_OPENAI=1 and an OpenAI model/API key are configured:

$env:OPENAI_API_KEY = "<set locally>"
$env:QUANTUMMINDLITE_LIVE_OPENAI = "1"
python -m quantummindlite.cli analyze --input C:\path\to\public-case.yaml --provider openai --model <model>

Do not use mock scores as model-performance or scientific-reasoning claims. Mock benchmark output is labeled fixture_self_test; live output is labeled live_model_run and records provider/model.

Offline QAEG

For each completed run, QAEG validates that the saved decision still equals a fresh deterministic B1-B10 decision, then projects the state, decision, and public runtime registry into typed nodes and edges. The verifier traverses those nodes and edges to check the claim support path, registry-derived obligations, barriers, graph integrity, and novelty boundary. It also reports a non-authoritative generic-wrapper diagnostic.

When per-run output is enabled, the sidecar writes:

• evidence_graph.json
• graph_verifier_report.json
• graph_summary.json

graph_status describes the graph layer (PASS, WARN, or FAIL); it is not the scientific verdict. In particular, a correctly represented negative run can have graph_status: PASS while claim_accepted remains false. claim_accepted can be true only when the authoritative B verdict is POSITIVE and all five hard graph checks pass. QAEG never upgrades B or changes D's route.

QAEG-screen-v0.1 is a separate deterministic screening layer. It consumes only the public RunState, the v1 graph/decision projection, and the public registry; it does not read input.json, trace.jsonl, or PaperBench gold/evidence. It classifies original-versus-candidate output alignment, access upgrades, oracle construction risk, classical-baseline status, and the candidate universe. graph_status, claim_accepted, and research_disposition are independent fields: screening may demote or defer a candidate, but it cannot upgrade or alter B, D, or the graph report.

Re-screen all 628 completed discovery/probe runs from Anaconda Prompt. This is zero-API replay and does not invoke a model:

conda activate quantummind
cd /d %USERPROFILE%\projects\quantummindlite
python scripts\regraph_runs.py ^
  --runs-root runs\final_discovery_run_v1 ^
  --overview-csv runs\final_discovery_run_v1\fdv1_probe_overview_rows.csv ^
  --output results\fdv1_screening_v01_summary.csv ^
  --family-output results\fdv1_screening_v01_families.csv ^
  --top-output results\fdv1_screening_v01_top9.csv ^
  --top-k 9 ^
  --no-write-per-run

Re-screen the 106 copied review candidates separately:

python scripts\regraph_runs.py ^
  --runs-root runs\final_discovery_run_v1_review_candidates ^
  --overview-csv runs\final_discovery_run_v1_review_candidates\manifests\combined_review_manifest.csv ^
  --output results\review_screening_v01_summary.csv ^
  --family-output results\review_screening_v01_families.csv ^
  --top-output results\review_screening_v01_top9.csv ^
  --top-k 9 ^
  --no-write-per-run

The family key is the normalized parent algorithm, selected primitive, original output type, provided access, and candidate universe. A missing parent keeps each run in its own family. Canonical selection is deterministic, and the top CSV contains only canonical PASS rows whose disposition is KEEP_FOR_EXPERT_REVIEW, LITERATURE_SEARCH_FIRST, or REFORMULATE; demoted, rejected, source-repair, and invalid rows are excluded. The legacy graph_value_label remains only for v1 comparison. There is no single weighted triage score.

A batch continues after an individual bad run, reports collected errors after writing successful rows, and exits with code 2 if any completed run failed or no completed run was available.

If --overview-csv is used, it must point to public-only metadata. Do not join PaperBench gold/evidence files into QAEG batch output.

These outputs are graph plumbing and research triage over existing artifacts, not model-performance evidence, independent replications, human labels, or proof of a new quantum algorithm. No expert-review accuracy metric is valid until blinded ratings exist. Reproducibility hashes are in results/qaeg_v1_baseline_manifest.json and results/qaeg_screen_v01_manifest.json; see docs/QAEG_METHOD_NOTE.md for the exact method.

Run Artifacts

Each run directory contains input.json, trace.jsonl, state.json, and decision.json; benchmark runs also write score.json. Trace rows record provider/model, action schema, prompt/input/output digests, latency, usage when available, attempt count, and parse/refusal/incomplete status. Hidden chain-of-thought and secrets are not persisted.

PaperBench

validate-paperbench checks public/gold/evidence separation, source mappings, freeze-manifest digests, and the B-rule implementation identity. The manifest is refreshed only by freeze-paperbench --confirm.

PB-007, PB-008, and PB-009 are represented as known quantum pathways with CONSTANT_FACTOR_ONLY speedup class. They therefore receive a NEGATIVE verdict for asymptotic speedup, not a “no primitive exists” label. PB-006 and PB-010 remain conservative no-asymptotic-speedup cases under their represented models.

Registry entries with speedup_class: NONE are diagnostic patterns, not selectable quantum pathways. Barrier agents receive only the catalog subset relevant to the plausible or selected pathway; canonical public access, promise, and output facts deterministically discharge catalog conditions where specified.