Skip to content

Latest commit

 

History

History
128 lines (104 loc) · 6.75 KB

File metadata and controls

128 lines (104 loc) · 6.75 KB

Experiment 008 result: CORA-C2 rejected before holdout

Status: development gate failed; protected holdout unopened.

CORA-C2 beat static target-Fisher by 26.17% and adaptive MSE by 21.62% on the frozen 16-task development window. It did not beat CQER-32: macro excess NLL was 0.391586 for CORA-C2 and 0.354250 for CQER-32. Confirmation-2 reduced normalized committed-mask churn by 79.99%, but it worsened raw-CORA NLL by 6.96%, above the frozen 1% limit. Five preregistered checks failed. Ranked MBPP window [8, 16) remains unopened.

Date recorded: 2026-07-23

Authenticated diagnostic

The development identity, implementation, six methods, storage contract, and 13-condition advancement rule were frozen before this run. The run used the separate ranked MBPP calibration window [16, 32).

Field Value
Artifact evidence/experiment008-cora-c2-development-16task-69eec866.json
Artifact kind recurquant_cora_c2_development_quality_diagnostic
Clean implementation commit 69eec866b90f2ae5386da23a8f34fba4c428d9e1
File SHA-256 03da6dc27641e7816f6f03bd059876322974530911d28e1d9719dd992af8f9db
Canonical evidence SHA-256 24877f73068f765ef0bae6b2bee54ff95d5204977b3c08dfe95516d651c2fb70
Created at 2026-07-22T20:51:42.559117+00:00
Model Qwen/Qwen3.5-0.8B-Base at dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68
Device and dtype CUDA, bfloat16
Task IDs 666, 795, 944, 653, 857, 884, 878, 822, 687, 820, 920, 771, 869, 851, 728, 704
Aligned scored tokens 798
Bootstrap 10,000 paired task resamples, seed 2339
Repository and source files clean and stable at start and end
Evidence verification valid, with no verifier errors

Independent rechecks reproduced every paired bootstrap contrast, source and selector hash, task and token identity, byte count, quota, and transition handshake. The protected [8, 16) loader guards also passed after the run. This is an authenticated method failure, not an execution failure.

Frozen-primary result

Method Role Macro excess NLL Mean KL CVaR95 KL Top-1 agreement
target_directional_fisher_difference_int4 static Fisher 0.5303981081 0.5286498712 3.3478351496 0.7586527802
adaptive_mse_target_directional_fisher_quota per-write row MSE 0.4996191338 0.4905063435 3.0781028867 0.7698915936
query_ema32_weighted_mse_target_fisher_quota CQER-32 0.3542503491 0.3641314693 2.4055756629 0.8027831651
causal_observability_mse_target_fisher_quota raw CORA 0.3661060743 0.3896415923 2.6112778820 0.8017726764
query_ema32_confirm2_mse_target_fisher_quota CQER-32+C2 0.3951669149 0.4165541902 2.7546808161 0.8111176938
causal_observability_confirm2_mse_target_fisher_quota frozen primary 0.3915858567 0.4186489517 2.8813917115 0.8103738017

Comparator-minus-primary paired NLL contrasts were:

Comparator Mean improvement Paired 95% interval Interpretation
Static target-Fisher 0.1388122514 [0.0436136126, 0.2466757520] 26.17% reduction; pass
Adaptive MSE 0.1080332771 [0.0277902825, 0.1880076991] 21.62% reduction; descriptive pass
CQER-32 -0.0373355076 [-0.1060209079, 0.0192620480] 10.54% worse; fail
Raw CORA -0.0254797824 [-0.0689048139, 0.0198680538] C2 was 6.96% worse; fail

CORA's transition-aware point estimate was also 3.35% worse than CQER-32, so the observability recurrence did not improve the strongest causal selector. The uncertainty interval for raw CORA versus CORA-C2 crossed zero, but the frozen advancement rule used the prespecified relative-worsening threshold as well as the paired comparisons.

Confirmation-2 result

Confirmation-2 behaved as a strong debouncer:

Selector Normalized committed churn
Raw CORA 0.0129384697
CORA-C2 0.0025893428
Relative reduction 79.9873%

That stability was not free. CORA-C2 raised NLL from 0.366106 to 0.391586. The same pattern appeared in the CQER ablation: C2 raised NLL from 0.354250 to 0.395167. C2 improved top-1 agreement for both selectors, but its NLL and tail-KL cost rules it out as the frozen advancement mechanism. This experiment supports treating two-hit admission as a switching-quality trade-off, not as a generally beneficial selector.

Gate decision

Frozen check Observed Result
Lower NLL than static, adaptive, and CQER CQER 0.354250 < primary 0.391586 fail
NLL reduction vs static 26.17%, minimum 20% pass
NLL reduction vs adaptive 21.62%, minimum 5% pass
NLL reduction vs CQER -10.54%, minimum 5% fail
Paired lower bound vs static 0.043614 > 0 pass
Paired lower bound vs CQER -0.106021 <= 0 fail
Top-1 margin vs static/adaptive primary was higher pass
Top-1 not lower than CQER 0.810374 >= 0.802783 pass
CVaR95 margin vs static/adaptive primary was lower pass
Raw CORA NLL reduction vs CQER -3.35%, minimum 3% fail
C2 churn reduction vs raw CORA 79.99%, minimum 50% pass
C2 top-1 not lower than raw CORA 0.810374 >= 0.801773 pass
C2 NLL worsening vs raw CORA 6.96%, maximum 1% fail
Exact bytes, quotas, handshakes, finiteness, identity all exact pass

The packed state used 2,564,096 bytes. CORA-C2 used another 152,064 resident bytes for its FP32 observability diagonal and bit-packed prior raw masks, for 2,716,160 bytes including selector state. This is not an equal-total-memory, peak-memory, latency, throughput, or fused-kernel result.

Interpretation and decision

The negative result isolates two useful facts:

  1. The frozen diagonal causal-observability proxy was not better than the much simpler query-energy EMA on this development window.
  2. Requiring two consecutive raw-mask hits removes most switching but is too sticky for the quality objective at this token scale.

Experiment 008 stops here. The independent numerical verifier is not built, the protected [8, 16) window is not opened, and no formula, threshold, or ablation will be tuned and relabelled as Experiment 008.

The underlying observability-weighted-error idea also has classical finite- wordlength state-space precedent. CORA is best understood as an experimental Gated DeltaNet adaptation of that principle, not a newly invented principle. Any successor must use a new protocol and development identity, replace or substantially strengthen the failed signal, and compare against CQER-32.

Experiment 008 makes no novelty, state-of-the-art, deployment, or breakthrough claim.