37 development requests and 108 candidate labels

This development dataset prepares a tiny claim/hint model to suggest a verification order. It preserves all48,618 bytes of the previous35 rows and appends two Union/JSON requests supported by actual original observation and an independent comparer. No new Fit or model activation occurred.

Measure	Actual value
Development requests / candidate labels	37 / 108
Positive / negative	37 / 71
Frozen inputs	184
All saved candidate observations / selected observations	548 / 539
Positive candidate position0/1/2	6 · 28 · 3
Added requests / labels	2 / 6 (+2/−4)
Project Reader attempts / returns / full matches	37 each
Additional normalized-input comparisons	37
New Fit / independent sources / protected additions	0 / 0 / 0

The13 new inputs and39 observations were compared against819 frozen predicates; a separate implementation matched every verdict. Union's cached cursor and JSON's per-object decoded-key policy passed every frozen condition. The other four candidates have concrete counterexamples. JSON Map's68 unknown predicates remain preserved. Combined with the previous35's post23/selected scope46, the linked scoped total is114, not an entire-history count. The39 pre-AFTER unknown SourcePanic diagnostics also remain in original evidence but are not added to Wanted predicate counts.

ROOT-QUALIFICATION records label authority; ROOT-QUALIFIED-ROWS contains the two unit-weight rows. MATERIALIZATION and post-close readback record actual Go generation and full Reader correspondence. The Reader does not assign semantic truth. Only request/candidate text enters features; labels, groups, sources and evidence stay separate.

37 is not the number of independent evaluation samples. Both requests retain existing groups79/83 and development_train. Union reuses catalog19's ID. JSON is related to an existing top-level JSONL envelope duplicate-key task; this request covers recursive per-object rejection before collapse and ordered prefixes. New independent groups, the new20→60 cohort and protected2400 evaluation additions are all zero. Candidate-position bias remains substantial, so this dataset alone demonstrates neither learning success nor cost savings.

Read data/train.jsonl through the project Reader or reproduce the file and receipt with:

bash scripts/verify-next60-thirtyseven.sh

Go1.27.1 is required. Temporary files are generated from previous35 and pinned qualified rows, then checked byte-for-byte and against all Reader input/supervision/metadata values. Original tasks and models are not rerun. Passing these checks does not automatically authorize a new Fit or protected evaluation.

HF37 immutable release is published. Fixed-commit readback matched61 new/changed files plus attributes; all37 viewer rows matched full values/order. All721 old paths/nine tags remain. New Fits/independent families:0.

Issue19 · 한국어

PR #132: 필수 CI 통과·동일 트리 병합 / CI passed; identical tree merged.

학습 없는37개 비교 / Nonlearned37 comparison · HF37 게시 증거 / HF37 publication evidence.