Skip to content

[qcs] frustration-free: sim-to-real quantum gate calibration - #282

Open
jiangweiqi001 wants to merge 54 commits into
QuantumBFS:mainfrom
jiangweiqi001:submission/challenge-113
Open

[qcs] frustration-free: sim-to-real quantum gate calibration#282
jiangweiqi001 wants to merge 54 commits into
QuantumBFS:mainfrom
jiangweiqi001:submission/challenge-113

Conversation

@jiangweiqi001

Copy link
Copy Markdown

Summary

  • implement a differentiable open-loop quantum-control model, Hessian/HVP landscape extraction, and query-only CMA-ES closed-loop calibration
  • add fair full/model-Hessian/random/oracle comparisons with strict query and shot accounting, restartable atomic artifacts, paired statistical analysis, and publication-figure tooling
  • document the verified 9,500-trial production run, reproducible LASG02 Apptainer workflow, scientific interpretation, and prompt history in Chinese

Results

  • recover effective Hessian ranks 3 and 15 for one- and two-qubit gates, consistent with the expected d²−1 curved subspace
  • complete all 180 Slurm array tasks and strictly validate 9,500/9,500 production trials with no missing or invalid artifacts
  • retain the complete production snapshot for restartable bootstrap aggregation and final figure generation

Test plan

  • Existing module-level and integration verification completed during development
  • Representative 881-query pilot passed strict artifact validation
  • Production snapshot: completed=9500, expected=9500, pending=0, errors=[], valid=true
  • Report passes git diff --check

Scope

This is an independent Challenge 113 submission branch based directly on main; it does not modify or merge into PR #152.

Made with Cursor

jiangweiqi001 and others added 30 commits July 30, 2026 19:59
Capture the official problem, pinned reference notebooks, and a verified literature corpus so the sim-to-real design can proceed reproducibly.

Co-authored-by: Cursor <cursoragent@cursor.com>
Define a falsifiable, query-accounted experiment that separates Hessian-informed transfer from generic dimensionality reduction and maps its model-gap failure regime.

Co-authored-by: Cursor <cursoragent@cursor.com>
Stage geometry validation, query-only calibration, restartable sweeps, and paired analysis behind explicit scientific and compute gates.

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Bind gate duration to reproducible system configuration after diagnostics show the original horizon lies below the CNOT interaction-time limit.

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Define the in-process API threat model, exact black-box observation semantics, and failure accounting needed for fair query comparisons.

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Require byte-canonical orphan adoption and hold per-trial kernel locks through execution so restart recovery cannot accept ambiguous artifacts or steal live foreign claims. Tighten durability and validation-order failures to preserve fail-closed accounting.

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Derive final attained bounds from all observed evidence while preserving cache independence, and validate solver termination independently from persisted raw facts.

Co-authored-by: Cursor <cursoragent@cursor.com>
Tie cached solver origins to independent exact evidence and prevent nonfinite solver output from fabricating attained improvements or provenance.

Co-authored-by: Cursor <cursoragent@cursor.com>
Prevent cross-design pooling and pseudoreplication while adding strict schema-v3 metric summaries, confidence intervals, and bounded deterministic bootstrap bands.

Co-authored-by: Cursor <cursoragent@cursor.com>
jiangweiqi001 and others added 24 commits July 30, 2026 19:59
Slice reused full-space diagnostics per stratum and make Wilson validation, canonical distributions, duplicate rejection, and public multi-k pairing independently strict.

Co-authored-by: Cursor <cursoragent@cursor.com>
Source reused full-comparator principal angles from independently matched target-k geometry and fail closed on missing or conflicting production evidence.

Co-authored-by: Cursor <cursoragent@cursor.com>
Render the required figures directly from strict canonical summaries with deterministic manifests and failure-case enforcement.

Co-authored-by: Cursor <cursoragent@cursor.com>
Use censored query statistics, complete stratum identities, invariant method styles, and exception-safe figure cleanup so publication outputs remain statistically honest and reproducible.

Co-authored-by: Cursor <cursoragent@cursor.com>
Pin deterministic publication artifacts and inspect complete uncertainty and geometry artists while isolating rendering from inherited matplotlib state.

Co-authored-by: Cursor <cursoragent@cursor.com>
Provide guarded restartable cluster shards and measured pilot tooling so the full study can proceed only from validated resource evidence.

Co-authored-by: Cursor <cursoragent@cursor.com>
Capture validated p80 pilot costs and the cluster ABI blocker so broad production remains fail-closed.

Co-authored-by: Cursor <cursoragent@cursor.com>
Cache the canonical model origin independently from statistical seeds and harden deterministic shards and deployment metadata checks.

Co-authored-by: Cursor <cursoragent@cursor.com>
Measure the exact canonical model-seed preparation used by the representative production pilot.

Co-authored-by: Cursor <cursoragent@cursor.com>
Keep model preparation fixed at seed five while calibrating and piloting the seed-zero statistical unit.

Co-authored-by: Cursor <cursoragent@cursor.com>
Create nested evidence destinations and serialize calibration records as stable canonical JSON.

Co-authored-by: Cursor <cursoragent@cursor.com>
Bind compact calibration, seed-zero pilot, validation, environment, and provisional projection records to their measured source revision.

Co-authored-by: Cursor <cursoragent@cursor.com>
Assert bitwise-shared model preparation while truth orientations and random search bases vary under their dedicated seeds.

Co-authored-by: Cursor <cursoragent@cursor.com>
Separate effective curvature rank from the complete curvature-ordered basis so every approved high-k model-Hessian trial is constructible and provenance binds the source basis.

Co-authored-by: Cursor <cursoragent@cursor.com>
Reject semantically inconsistent measurements and stale deployment inputs while recording a fresh x64 pilot and a coherent pre-Task10C gate report.

Co-authored-by: Cursor <cursoragent@cursor.com>
Use identical bounded coordinates for model-Hessian k=p and make the documented production gate consume external metadata without dirtying clean source checkouts.

Co-authored-by: Cursor <cursoragent@cursor.com>
Bind production gates to the compute-verified frozen image and make Slurm jobs reuse an offline, hash-validated environment before any physics executes.

Co-authored-by: Cursor <cursoragent@cursor.com>
Bind external metadata by its exact byte hash and require LASG02's verified clean, network-none Apptainer mode before any runtime or physics command.

Co-authored-by: Cursor <cursoragent@cursor.com>
Carry the revision-bound source archive basename through Apptainer mounts so strict metadata validation succeeds without accepting unsafe aliases.

Co-authored-by: Cursor <cursoragent@cursor.com>
Allow one explicitly acknowledged frozen dependency sync, then bind all smoke and job execution to the verified network-none runtime marker.

Co-authored-by: Cursor <cursoragent@cursor.com>
Require launchers to resolve the verified runtime gate from the explicit deployment root so Slurm copies cannot redirect execution through their spool path.

Co-authored-by: Cursor <cursoragent@cursor.com>
Resolve container-owned Python executables only inside the isolated SIF and make every deployment digest failure explicit before runtime execution.

Co-authored-by: Cursor <cursoragent@cursor.com>
Disable pycma's broken bound-range standard-deviation limiter only for one-dimensional spaces so all canonical k=1 trials retain CMA-ES, bounded coordinates, and exact accounting.

Co-authored-by: Cursor <cursoragent@cursor.com>
Present the verified sim-to-real pipeline, reproducibility evidence, and development prompts in a review-ready Chinese report.

Co-authored-by: Cursor <cursoragent@cursor.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant