Scalar command-response calibration study v1

A frozen comparison of fixed and adaptive phase diversity under constant scalar command-response mismatch: 288 episodes, 2,016 signed detector exposures, 24 independent bases and three gain settings.

Adaptive probing reduced equal-stratum pooled WFE under ±10% command-response error, but performed worse in the 10–25 nm initial-RMS stratum at both mismatched gains. Fixed diversity retained a pooled advantage at unit gain. These are descriptive development results; the release does not establish held-out superiority or hardware qualification.

• Results and methods
• Evidence packet and reproduction
• Independent numerical audit

The raw archive retains every terminal episode, detector array, fitted command path, source snapshot and frozen admission. LICENSE and NOTICE accompany the study directory. The production retained-path verifier and an independently implemented WFE/bootstrap audit passed.

Archive SHA-256: 3c08df9e04c4b5110893ba758f13f251c36f59b6f6a4372c224d2045371b7e38.
Raw manifest SHA-256: ee56441561fa50aeb80f1e33e516cf9151248445f75d997e0d46e9b5002c3322.

Bootstrap intervals are pointwise and descriptive. Equal exposure budgets do not imply equal computation; adaptive refits six prefixes while fixed diversity fits once. The independent auditor requires NumPy 2.4.6; full retained-path replay additionally enforces the recorded source and runtime.

Classical reference selection v1

The complete v1 tuning study freezes one setting for each of six classical optical-control families. The declared grid selects a 20 nm random-probe amplitude, a 40 nm coordinate-search step and 80 nm defocus probes for both diversity methods; no-action and unprobed MLE remain fixed controls.

• 4,032 episodes and 30,912 signed detector exposures from 12 independent base wavefronts.
• All 14 candidate reductions, nonwinning settings, six full policy configurations and complete raw evidence retained.
• 6,912 multistart fit calls and 20,736 individual starts; every retained start reports finite convergence.
• Source-bound verification checks the full inventory and replays detector draws and commands using saved fits. It does not rerun optimization.

Technical results · Reference packet and verification

The attached native archive contains every manifest-bound file and the exact frozen source, with LICENSE and NOTICE outside the immutable native directory. Verify its SHA-256 against the attached checksum and the repository release record before extraction. GitHub's automatically generated source-code archives contain the compact reference packet, not the full detector dataset.

These are public tuning results on a static, monochromatic, exactly matched six-mode simulator. They do not establish held-out performance, a population ranking, telescope performance or language-model competence. The existing sequential HTML replay and ImageFolder exporters do not support this selection archive; no selection-specific Hugging Face loader is claimed.

Source snapshot: public revision 705e31e00325d5c0648cdc6161ed2ba2c338c666. It preserves the acquisition implementation and predates later hardening; use the current repository for general development. Software licensing follows the included Apache-2.0 LICENSE and NOTICE.

Classical-selection measurement replay v1

An offline measurement viewer for the complete classical-reference selection study. It retains all 14 candidates, including nonselected settings, across 4,032 episodes and 30,912 recorded exposures from 12 independent base wavefronts.

Extract classical-selection-replay-v1-20261017.tar.gz and open the enclosed index.html. Keep the adjacent images/ directory. No server, model runtime or external assets are required. Select the policy family, candidate, photon allowance, exposure budget, RMS stratum, base wavefront, conjugate member and detector replicate; inspect recorded exposures with playback, the evaluator trajectory and image zoom.

The final submitted correction is distinguished from camera observations: 2,016 final commands were unmeasured, 1,577 were measured earlier, and 439 were measured at the last exposure. A final-image shortcut exists only for a recorded measurement. WFE and Strehl remain evaluator labels, not controller observations.

Every embedded trajectory and final record was checked against the native archive. All 30,912 copied PNGs match the verified ImageFolder export, and the generator validates their pixels against the retained signed detector arrays. Chromium 145 traversed all 4,032 episode selections and decoded every last exposure plus all 2,016 recorded-final shortcuts. The code passed 690 tests. Browser validation covers endpoints; raw-data validation covers every frame.

This is a derived inspection artifact, not a new experiment. It does not establish held-out rankings, hardware performance or language-model capability. Raw signed arrays, complete fit records and acquisition source remain in the native scientific release; they are not duplicated in this viewer.

Generation and artifact format · Study results

Verify the archive with the attached SHA-256 file before extraction. The enclosed manifest.json has SHA-256 b0b47b29345e4d42508916c249cdcc06d97f1b0148e0182ad5e9b547d8028f34. Its source ImageFolder manifest is f6d61b8342c239732e4c2e34f8ac14eddf02b971e4514d4040923d377e65920b; the native manifest is 95130d4d9bf5d534aa3bfc88bdaef5934b5862769061df2bb26c03cf72806ef0. Generator, helper and template fingerprints are included in provenance.json. The archive includes LICENSE and NOTICE. GitHub's automatically generated source archives contain the generator and reference packet, not the replay images.