Repository navigation
Conversation
Add explicit nominal observations, analytic seven-parameter likelihood gradients, and focused regression tests. Development-only, offline calibration; no changes to frozen evidence, existing policies or the reserved evaluation cohort.
Executed local development comparison — 31 October 2026The focused prototype at
At gains 0.9/1.1, mean reductions are 68.00%/70.11% against fixed and 48.13%/53.09% against adaptive. At unit gain, the joint estimator is 0.1190 nm (7.81%) worse than fixed. Individual regressions are retained. This corrects a specified model misspecification, not a novel algorithm or proof of universal controller superiority. Design: 24 independent base wavefronts (eight in each of three RMS strata), reused across three gains; 72 base/gain cases and 216 method outcomes. Seven exposures, 14,000 expected photons per method. Joint and fixed use exactly the same camera arrays. Adaptive has the same photon/exposure budget, not the same computation budget. Six-mode, static, matched pupil/camera model apart from constant scalar command response. Known flux/read noise. Final correction is evaluator-scored without an additional camera image. The mismatch contrasts pass paired Bonferroni simultaneous 95% intervals across six contrasts, plus a pre-acquisition point-reduction screen of at least 20%. The nominal screen used a provisional, non-customer-approved 0.5 nm margin. All 576 selected fits reported successful finite optimization. Three local starts are not proof of global optimality. Fresh RNG roots have not been admitted through the upstream exclusion ledger. Verification: 16 focused tests actually executed and passed in 0.61 s. A separate audit implementation, importing neither ProbeBench nor the acquisition harness, used direct Fourier matrices to reproduce all 1,008 retained signed camera arrays exactly and all physical WFE values exactly; maximum Strehl difference 1.23e-15. This is an independent implementation by the same assistant, not third-party replication, optimizer refitting, cross-simulator or instrument validation. Timing boundary: median joint offline fit time was 1.960 s on this local concurrently used CPU. Not a real-time claim. Baseline timings include simulated acquisition whereas joint starts from retained fixed observations, so these are not controlled end-to-end speed comparisons. The complete local source/data packet is supplied to the repository owner in the current ChatGPT conversation as Evidence bindings (SHA-256):
|
Purpose
The scalar-response development study changes physical command gain while the existing estimator assumes unit response. This draft adds a narrowly scoped alternative: estimate the six phase coefficients and one positive scalar command gain jointly, then compensate the final nominal command.
This is actuator command-response calibration, not pyramid-WFS optical-gain detection. It does not resolve the separate Noether AO detector failure or establish runtime safety, RL superiority or a new scientific method.
Changes
NominalObservation: immutable copied camera data, expected photons and nominal commands; no evaluator truth fields.Existing policies, defaults, frozen results and reserved evaluation cohorts are unchanged.
Executed verification
16 focused tests passed locally using byte-verified upstream
optics.py,likelihood.pyandestimation.pyat base9406f5b5b04f9692cd194599277109423a2dbee6. Tests cover every gradient coordinate, the unit-gain reduction to the nominal objective, immutable observations, rejection of unchanging commands, compensation algebra and a bright synthetic recovery fixture.The full repository suite, CI, supported-version matrix and hardware were not run in that local environment. The complete separate development acquisition packet is retained outside this PR; this code-only PR does not publish or promote a benchmark result.
Before promotion
The nominal estimator can remain better at unit gain. Three optimizer starts, bounded gain and passing unit tests are not global identifiability or deployment guarantees.