RIDR1: Recurrent Revision Depth

Status: closed negative on development; holdout remains sealed.

Hypothesis

The qualified Qwen3.5-9B IDR1 reviser may implement an iterative correction operator rather than a single-use rewrite. Applying the exact same trained reviser to its own completed first revision could repair residual errors without new weights, training data, sampling, tools, or external verification.

One Changed Factor

Depth one is the immutable IDR1 development result (589/1,289). Depth two replaces the original internal draft with the exact depth-one completion for the same identity. Model root, B1 warm start, IDR1 checkpoint, source, prompt constructor, greedy decoding, 768-token budget, evaluator, seed, and batch geometry remain fixed.

Only development is opened. A pass may unlock exactly one depth-two holdout run. A failure closes recursive depth for this reviser; there is no depth-three, prompt, seed, token-budget, or threshold rescue.

Frozen Gate

All conditions are conjunctive:

• depth-one receipt exactly reproduces 589 overall, 223 math, 349 logic, and 17 code;
• depth two reaches at least 615/1,289, a gain of 26 answers (2.02 percentage points);
• math, logic, and code are each nonnegative versus depth one;
• at least 98% of depth-one-correct examples remain correct.

The comparison reports repaired errors, newly broken correct answers, persistent errors, preserved correct answers, generated tokens, token exhaustion, peak memory, aggregate GPU time, and critical-path time.

Interpretation

A pass supports useful model-owned recurrent inference depth for this trained reviser. It does not establish indefinite convergence or architectural novelty. A failure means the learned rewrite is not a stable iterative operator under this exact contract.

Result

Eight single-H100 shards (747426--747432, 747442) completed the exact development pass. The first shard-7 allocation (747433) had no visible GPU; two replacement submissions (747436, 747440) failed before model execution because the manually supplied model-manifest hash had one extra character. Debug allocation 747441 identified that custody typo. These failures produced no candidates and did not alter the scientific settings. Merge 747443 and comparison 747444 completed cleanly.

Depth two scored 539/1,289, down 50 answers from depth one's 589. It repaired 15 depth-one errors but broke 65 depth-one-correct answers, retaining only 524/589 = 88.96% of correct answers. Every domain regressed:

All capability gates fail. The merged evaluation used 17,035 generated tokens, 625.83 aggregate GPU-seconds, a 106.17-second generation critical path, two token-exhausted outputs, and 19.877 GB peak allocated memory. Comparison SHA-256 is cccc45c0c517df597c24fb64dc1cfc8aa41b0721806fde61987d013e00989ca1; merged report SHA-256 is 755322624182f0b307f7f21b2bfeb44eb9461a8b99804708cba7d5080554e045.

RIDR1 is closed without depth-three, prompt, seed, budget, or threshold rescue. A future multi-stage design must train each later owner on the distribution of completed earlier revisions and must earn conservative answer retention.