GridGxly/qscore-runs

Gatesmith circuit improvement attempts

Gatesmith scores quantum circuits and evaluates optimizer proposals under recorded conditions. Open-weight models are called through Hugging Face Inference Providers using pinned model/provider pairs. The engine, rather than a model claim, determines correctness and improvement.

Current status

Integration setup is in progress. leaderboard.parquet currently contains a valid typed schema and zero benchmark rows. The successful API smoke is an unverified proposal, and is excluded from benchmark ranks. No accepted circuit improvement or qscore-compatible certification is claimed yet. This dataset is private during development.

What an attempt records

An attempt is a candidate returned for a circuit-region request, with job/request identity, layer/model/provider/prompt provenance, validation, verification status, scoring evidence and acceptance decision. Call-level token usage and estimated cost are recorded once per API call. Missing accounting remains null. Worker exports will populate attempts after the verifier and loop are connected; this initial setup contains no fabricated attempts.

Leaderboard fields

Groups use (layer, model, provider, prompt_version). Columns are requests, proposals, parse_rate, verified_rate, verified_improvement_rate, median_latency_ms, prompt_tokens, completion_tokens, cost_usd_estimate, preceded by the four group identifiers. Counts are integers; rates, latency and cost are numeric. Unknown rates/accounting are nullable, never substituted with zero.

Parse rate is parsed proposals divided by returned proposals plus calls with malformed output. Empty responses are not malformed proposals. Verified rate is verification passes divided by parsed proposals. Verified improvement rate is distinct requests with at least one accepted proposal divided by distinct requests attempted; request identity includes job ID. Only compare runs under identical fixtures, budgets, seeds, targets and engine/score versions.

Verification and caveats

Schema-valid output is not proof of equivalence. A candidate must be spliced into the original circuit, verified using the declared method, then re-scored under matching settings and satisfy acceptance constraints. Inconclusive verification is never a pass. Estimated API cost is advisory; provider billing is authoritative. The read-only leaderboard Space performs no inference or circuit execution.

Versions, sources and license

Initial scoring baseline: Gatesmith commit 6b42f722730f19a5c3d9b72e1a19cf8d54a0fedd; score version qscore-2026.2. Future records must carry their actual versions. Our authored records and example circuits use Apache-2.0. Model licenses are recorded separately in the pinned registry; no model weights are redistributed here.

Source and verification specs · Leaderboard Space