experiment: add q128 gap-response screen - #29
Draft
edwards365 wants to merge 72 commits into
Draft
Conversation
Bootstrap ECT environment, dataset verification, and AMP compatibility
…ration Integrate reproducible Day 2 engineering workflow
Validate reproducible clean-container environment and FP16/AMP resume workflow
Integrate deterministic fixed-seed sampling, evaluation protocol, tests, and compact smoke evidence.
Add the clean Role B paired-training runner, telemetry, resume identity gates, evidence collector, tests, and A100 smoke artifact.
Add the loss-EMA feedback-adaptive schedule, fixed-schedule parity tests, adaptive telemetry, exact resume support, activation gating, canonical dataset identity, and clean A100 activation evidence. Role B reviewed head c138bd7: 81 tests passed with no blocking issues.
Add canonical-runtime paired sigmoid/adaptive activation and fresh 16 kimg stability evidence, including comparison summaries, controller telemetry, and checkpoint identities for Role D evaluation.
Add the canonical fixed-vs-adaptive NFE=1/2 evaluation evidence, deterministic manifests, method grids, same-seed comparison grids, and final provenance documentation.
Archive the frozen 16/32/64 kimg Fixed sigmoid vs Adaptive v1 evaluation, provenance, stability records, and guarded negative conclusion.
Archive the frozen Role C paired analysis, guarded decision logic, provenance manifests, and final negative conclusion for the 16/32/64 kimg matrix.
Archive the 64 kimg blind-package provenance, fixed-layout qualitative grids, presentation assets, and a showcase that links to the authoritative Role C conclusion without duplicating it.
…24k-results Archive paired 1024 kimg seed0 evaluation
Add global gap calibration, geometrically normalized local gap control, factorial experiment tooling, compact result summaries, diagnostics, tests, and method documentation.
…rmatory Role C: complete seeds 4/5 confirmatory 256k paired runs
…enance Fix Role E source-provenance reference hashes and classification
Freeze and execute the q=256 confirmatory evaluation matrix
Add versioned formal evaluation artifacts for the q=256 fixed vs global-only matrix. Include all 24 metric records, 12 seed-level paired differences, paired statistics, and portable environment/checkpoint provenance. Replace server-specific paths with logical identifiers and link the formal record to the new package.
Add path-free prospective matrix specifications for q=256 budget and fresh q=128 experiments. Freeze seeds, methods, NFE settings, per-budget metric contracts, pairing keys, and global-only delta direction. Document that quick results cannot select formal cells and require evaluation of every predeclared checkpoint.
Rename formal q=256 results to the release package path. Add frozen q=256 budget and q=128 confirmatory matrix specifications, evaluation job counts, and an evidence-based storage estimate. Update protocol references and matrix tests for the final deliverable names.
protocol: deliver q=256 formal results and freeze prospective matrices
Add publication-ready q=256 figures and LaTeX tables, separate same-protocol budget curves from formal endpoint comparisons, extend the result collectors, and add protocol-aware tests and appendix diagnostics.
Merge the completed six-cell q128 training handoff and supporting provenance artifacts.
Merge the completed 12-job q128 1024 kimg evaluation matrix and paired results.
… doc - theory/: toy_core.py, scan_toy.py, plot_toy.py + toy_hessian.csv, toy_finite_budget.csv, toy_hessian_power_terms.csv - figures/: toy_condition_number.pdf, toy_error_vs_g.pdf, toy_gstar_vs_budget.pdf - Novelty statement and three propositions (Prop A/B/C) - Handoff doc for 2026-08-04 Role C work
…s, ADCM separation Addresses REQUEST_CHANGES on PR #33: P0-3 (stop-gradient): fix A_g to sigma_d^2 E_t[[t,t^2]^T v_g^T] with J_t=[t,t^2] (not v_g). Verified by finite-difference: per-sample gradient expectation matches A_g beta to rel-err 7e-12 (analytic E_z); population loss curvature == H_g confirmed. P0-2 (LR artifact): add scan_controls.py with eta_scale sweep {0.05..1.0} and LR-matched eta_g = eta_1*lam_max(H_1)/lam_max(H_g). Result: under LR-matching the noiseless U-shape COLLAPSES (spread~0) -> it was an effective-learning-rate artifact; noisy case is monotone (noise amplification, no internal min). P0/separation: add separation.py — symmetric vs stop-gradient losses share identical H_g (same instantaneous criterion) but differ in Sigma_g; with LR-matched eta and fixed noise RMS, g_K^star differs (K=50: 1.20 vs 1.30; K=200: 1.45 vs 1.35), converging at K=1000. P0-1 (novelty): rewrite novelty_and_propositions.md to instantaneous-vs- finite-horizon framing; drop false 'ADCM=stationary/orthogonal' claims; separate g_rho^star (budget-independent) from g_K^star. Minor: add test_toy_core.py (7 unit tests, all pass); move training scripts into scripts/; report clipping polynomial residual; fix eta*lambda_max doc value (1.40 -> 1.00).
…ev.2)
Per review, separation.py is NO LONGER Monte-Carlo SGD on two loss types
(which had 3 errors: missing H_g drift, non-zero-mean noise, and treating
sym/stop as having the same average operator). It is now an EXACT
second-moment recursion M_{k+1}=B_g M_k B_g^T + eta^2 Sigma_g, E_K=Tr(M_K).
Two settings, both with identical H_g (same instantaneous criterion):
- realistic: Sigma^(1)=H_g (g-dependent, ~g^2, symmetric-loss noise feature
v_g contains Delta~g) vs Sigma^(2)=sigma_d^2 E[[t,t^2][t,t^2]^T]
(g-INDEPENDENT, stop-gradient online Jacobian [t,t^2] has no g).
-> g_K^star DIFFERS (env1->small g, env2->large g) at fixed eta. The
separation is exactly the g-dependence of Sigma, not a learning-rate
effect (eta fixed) and not pure direction.
- trace-matched (equal-Tr rescale): separation COLLAPSES -> honest negative
result: pure noise direction (equal power) is insufficient; the separation
genuinely needs the g-dependent vs g-independent noise structure.
novelty doc sec 4 rewritten accordingly; adds the LR-matched caveat (there
B_g is g-independent under H_g~g^2 H_2 so the fixed-eta separation also
collapses -- counterexample stated for the real-training fixed-eta regime).
Minor: scan_controls.py docstring drops the unimplemented 'gradient-RMS-
matched' claim (subsumed by LR-matching); add test_toy_core separation test
(exact recursion vs 1-D closed form, 8/8 tests pass).
…tion theory: linear-Gaussian toy model for gap calibration
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Findings
Limitations
Validation
git diff --cached --checkpassed before commit