NDR1: Natural Draft Revision

Status: complete and closed negative on development. Holdout remained sealed.

Pre-Output Custody Correction

The first prepared source (e9965343...4389) was never consumed by a draft owner or fit. A read-only comparison against the original IDR1 source questions found 34 exact development overlaps, 26 exact holdout overlaps, and additional split-unique word-13-gram collisions. Filtering that already-selected file would retain 10,918/11,276 rows but disproportionately remove science rows, so it is not an admissible training source.

The corrected source must be rebuilt from the same raw corpora, tokenizer, seed, token budget, and domain weights. Before tokenization and deterministic selection, it removes exact word-normalized source questions and every word 13-gram occurring in exactly one row within either bound evaluation split. Repeated split boilerplate is not an exclusion gram. Selection then backfills from remaining candidates until every original group token quota is met. The report must bind both reference paths and SHA-256 values, exclusion-set sizes, global and per-source drop counters, and the final output hash. Draft generation remains closed until an independent audit reports zero exact and zero protected 13-gram overlap.

The first hash-bound rebuild (750210) met every domain token quota, but its independent audit (750214) failed before GPU release because punctuation- equivalent source questions remained duplicated under the builder's legacy whitespace-only identity. The corrected filtered path uses the same word- normalized identity for exact-overlap filtering and source deduplication. This is an input-custody repair with unchanged corpora, tokenizer, seed, quotas, and capability gate; the rejected build is not admissible training data.

Hypothesis

CFR1 failed because appended synthetic faults and clean-copy presentations did not resemble the errors produced by the B1 draft owner. NDR1 changes the data mechanism, not a CFR hyperparameter: B1 generates one deterministic draft for each short, fully verified, source-disjoint training problem. The revision owner then sees the original source plus that exact model-owned draft and must emit the untouched verified solution. There are no synthetic faults and no clean draft equals target presentation.

Matched Control

The control receives another source's same-domain, nearest-length B1 draft. Source, verified target, row order, target-token multiset, initialization, optimizer, update budget, trainable parameters, and evaluator are identical. No source receives its own draft. The comparison isolates information in the actual draft from generic verified source-to-solution training.

Data Gate

The source is a fresh Qwen-tokenizer-exact 4M-target-token mix with 1,536-token source-plus-target admission and 40/10/40/10 math/code/science/procedural weights, rebuilt prospectively under the corrected overlap boundary above. Sixteen deterministic B1 shards generate at most 768 tokens per source. The merged curricula must retain at least 90% of source rows at 4,096 tokens, have exact target multisets, zero source/donor identity matches, exact source/draft/checkpoint hashes, and no holdout use. Failed or missing shards fail closed.

Capability Gate

Train exactly one aligned and one shuffled arm from immutable B1 update 256, using identical final-four rank-8 LoRA, 512 updates, batch 1, accumulation 8, 4,096 context, learning rate 2e-5, and matched seeds. Evaluate once on the existing 1,289-row source-disjoint IDR1 development board. Promotion requires all of:

• aligned at least 603/1,289;
• aligned at least +10 answers over shuffled;
• math at least 223, logic/science at least 349, code at least 17;
• aligned 768-token exhaustion no more than 400 and no more than shuffled plus 25; and
• complete matched token, parameter, memory, latency, and protected-hash receipts.

Any miss closes exact NDR1 without source-size, generation, update, rank, layer, seed, decoding, parser, or threshold rescue. Holdout remains sealed until a conjunctive development pass.

Result

The corrected source passed independent custody before model work: 11,220 unique normalized questions, 4,000,845 charged target tokens, zero exact development/holdout overlap, zero protected split-unique word-13-gram overlap, zero normalized duplicates, and zero admitted truncation. Sixteen H100 shards generated one natural B1 draft per source. The merged curricula contain 11,220 rows per arm, exact target multisets, no source/donor identity matches, and same-domain nearest-length shuffled donors with character-length delta p95 of six.

The matched aligned and shuffled fits each completed 512 updates, consumed 1,479,584 charged target tokens, and trained exactly 2,704,896 parameters from the same B1 update-256 checkpoint. On the frozen 1,289-row development board:

Aligned loses 37 answers to the matched shuffled control and 34 answers to the unchanged reference. It repairs only 46 of 767 cases where the base and expert references are both wrong, versus 68 for shuffled, while generating 75,804 more tokens and exhausting 111 more rows. Only the code floor and receipt gate pass. The exact comparison SHA-256 is ae15a15b5cf8761cf9c12c5188ea06b85437f429f0553a29d4b31016f0379361.

This is a decisive negative, not an ambiguous near miss. Under ordinary full-trajectory CE, the revision owner can learn source-to-solution behavior without extracting useful information from its aligned natural draft; in this run the aligned draft is actively harmful. Exact NDR1 closes without a retry, and holdout remains unopened. A successor must make draft-dependent state or execution causally necessary rather than repeat source-plus-draft full-answer regeneration. Machine-readable evidence: docs/research/SHOHIN_NDR1_RESULT.json.

Read-Only State Attribution

An identity-exact post-closure attribution joined the immutable development data and both candidate files without model execution. Training contained only 1,405/11,220 = 12.52% token-exhausted drafts, whereas development contained 910/1,289 = 70.60%. On the 871 wrong exhausted drafts, aligned wins 11 pairwise cases and shuffled wins 44: a -33 delta that explains 89.19% of the total -37 deficit. Aligned also generates 823,764 tokens versus shuffled 747,960. Attribution SHA-256 is 5a7c5c6fc1e17c470e64a926d1b6df5c0c22840f45d94eb601424e211616e6e4; result: docs/research/SHOHIN_NDR1_STATE_FAILURE_ATTRIBUTION_RESULT.json.

This identifies a termination-state/interface mismatch but does not authorize an exhaustion-reweighted NDR1 retry. KCR1 is a structurally different successor in which model-owned cutoff state selects an executed KEEP, CONTINUE, or RESTART branch.