Skip to content

Study RHBI critic-feature transport on the Tier 1 texture gate - #558

Open
martyn wants to merge 20 commits into
developfrom
research/rhbi-texture-feature-transport
Open

martyn wants to merge 20 commits into
developfrom
research/rhbi-texture-feature-transport

Conversation

@martyn

@martyn martyn commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Current result across five synthetic tasks

The opt-in raw-assignment recipe passed clean 32×32 texture, sparse arc and unequal-mass width. It failed the full moving-patch and saddle interpolation gates. Its matched raw RHBI control passed both interpolation gates and missed width. CIFAR quality remains unmeasured. Five-task comparison

Nine new same-source runs used 8,400 updates at protocol seed 0. The candidate passed 19 consecutive width observations. Its moving-patch peak fell to 0.77374 at update 800 against a floor of 0.78. A replay confirmed all 141 final saddle outputs were finite, and 53/71 barycenters qualified against a floor of 78%. Raw RHBI qualified 59/71 on those same triangle row IDs. The single CUDA finiteness rounding error does not change the saddle verdict. Retention measurements, exact replays and training GIFs

The next proposed probe holds a saved state and its assignment fixed to measure how raw and feature attraction affect the missed interiors. The three retention gates were diagnostic and all 11 registered scientific rows were preserved.

Initial feature-transport study

Adding normalized critic readout features to RHBI's marginal transport metric fixed the 32×32 texture failure in this matched study. The candidate passed all five terminal checks. Its clean texture fraction rose from 0 to 0.906 and high-frequency energy ratio rose from 0.0086 to 0.696. Both arms used the same source, initial parameters, real batches, clean-live sampler and 288 updates. The opt-in flag is the sole resolved recipe delta.

Retention gate Full result Observation
Rare-mode mass FAIL Endpoint passed all bounds, with a passing suffix of 1/5.
Sparse arc FAIL Chord quality 0.739 against 0.80.
Moving patch FAIL 2/5 terminal observations passed. Step 767 missed the peak bound.
Saddle FAIL 0/5 terminal observations passed. Step 767 missed barycenter quality.

The four failures ended the study before CIFAR. The retention parent measurements remain on their original source. The readout documents the interpolation scorer's float32 finite-flag rounding and identifies genuine quality misses apart from the sentinel values.

Frozen coupling follow-up

Switching the metric on identical outputs moved 97.56% of transport mass on the blurry texture control, 0.035% on the unequal-mass endpoint and 0.055% on the arc endpoint. No active Gaussian mass crossed mode boundaries. All 18 failed arc interiors landed inside the circle. The endpoint measurements give little support to incorrect matches as the retention failure mechanism. Earlier training dynamics remain open. Readout and reproduction

Scope

This PR includes the revision-18 texture gate and its original control publication, an opt-in scalar marginal GAN loss, a zero-update endpoint probe, a frozen seed-0 comparison and compact research-only receipts. The probe found normalized feature nearest-neighbor distances of 0.405 for real textures, 13.35 for blurred textures and 15.39 for flat means. The matched source regenerated the parent's near-zero texture energy. No default promotion follows.

Research readout and actual-training GIF · Matched comparison provenance

Verification

476 local tests passed (44 RHBI and texture tests plus 432 Forge integration and reporting tests), including the objective and derivative control, rejection of missing features and exact continuation in a fresh trainer. Forge independently graded all six paid attempts. Forge declaration validation, lint baseline, and the compiled research memory integrity check passed.

The follow-up passed 531 local tests covering the compatibility fixtures, historical comparisons, Atlas coverage and RHBI/texture integrations. Seven historical incompatibility controls retained their expected XFAIL. Two fresh processes reproduced the saved-state probe byte for byte.

Matched training trajectories

Four seed-0 runs compare RHBI-v12 and its feature variant on the two selected retention tasks. Each pair shared a frozen source, initializer, examples, particle draws and budget. On sparse arc the parent passed 5/5 terminal checks and the feature arm passed 0/5. Both retained all 12 endpoints. Their executed transport plans diverged at update 29, before their mean error on common edges separated between updates 40 and 50. Switching only the metric on a single feature-arm state moved at most 0.1372% of mass over all 600 updates.

On unequal mass both arms failed, each passing only the final observation of the five required. The historical parent PASS used another source and runtime. All 48 official observations from the earlier feature runs reproduced exactly, as did their final models, optimizers and named streams. The ambient process RNG snapshots differed. The arc pair required reattaching the observer at dispatch to preserve its pinned evaluator.

I would next test features for pair assignment with raw-coordinate generator movement and self-repulsion, starting on the full texture gate. CIFAR remains unmeasured. Matched trajectory report, chart, and four Forge receipts

Feature assignment with raw motion

The assignment-only recipe failed the complete 288-update texture gate. Across the 24 evaluations it produced zero clean textures. The final high-frequency energy ratio was 0.1992 against a floor of 0.55. The matched raw control reached 0.0086 and full-feature transport passed at 0.6959 with 90.6% clean textures. Both controls reproduced each of their 24 previous observations exactly. All three arms shared source, initialization, batches and final named RNG states. Three attempts consumed 143.41 worker-seconds.

The stop rule closed the arc and unequal-mass runs for this variant. CIFAR remains unmeasured. The next ablation worth testing would retain full feature-based cross-attraction and restore only raw self-repulsion. Frozen protocol, results, receipts and actual training GIF

Texture recovery with raw self-repulsion

The recipe rhbi_feature_transport="raw_self" keeps feature-based real–fake attraction and uses raw generated coordinates for its fake–fake plan and physical self-cost. In the matched seed-0 comparison it passed all five terminal texture checks and recovered both orientations. At update 288, 241/256 samples were clean, the high-frequency energy ratio was 0.6700, class mass TV was 0.1250 and phase diversity was 0.8049. Full feature transport passed 5/5 and raw RHBI-v12 passed 0/5. All four controls across the two tasks reproduced their 24 previous observations exactly.

Sparse-arc interpolation remained a numerical failure. Both feature recipes learned all 12 endpoints and passed 51/69 chord interiors. Raw RHBI passed 53/66. The two feature recipes had identical live model tensors at all 101 saved states and identical executed cross-plans on all 600 arc updates. CIFAR remains unmeasured. Results, original Forge receipts and both training GIFs

Raw matching with feature attraction

The recipe rhbi_feature_transport="raw_assignment" uses raw coordinates to assign real samples and slack, feature-based real–fake attraction, and raw self-repulsion. It passed all five terminal checks on both selected tasks. The matched raw-self parent passed texture and failed arc. Raw RHBI failed texture and passed arc. Each control reproduced all 24 earlier observations on each task.

At update 288 the candidate had 256/256 clean textures, both orientations, a high-frequency energy ratio of 0.7055, class mass TV of 0.15625 and phase diversity of 0.7712. On sparse arc it learned all 12 endpoints and passed 56/69 interiors. The candidate and immediate parent selected the same 23 particle-row edges. The parent passed 51/69. One further missed interior would drop the candidate below the 80% floor. The candidate's 600 executed cross-plans exactly matched their same-state raw replays.

The follow-up above tested this global recipe on unequal-mass width, moving patch and saddle. CIFAR remains unmeasured. Six receipts, numerical results and actual-training GIFs

@martyn
martyn marked this pull request as ready for review November 18, 2026 04:53

This branch has not been deployed

No deployments
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