Tiny hints: actual preparation69 and new roles70

The goal is to cheaply suggest which evidence candidate an agent should inspect first. This encoder-free claim-hint experiment is separate from training the Laya backbone. Sources and requirements still verify each suggestion; remaining candidates and fallback stay available when early suggestions are wrong. This does not authorize actions or prove correctness.

For “find Go files in nested directories,” several function captions may be available. The model proposes inspection order. Success means fewer checks before finding an acceptable candidate on identical requests. This remains a development proxy; actual LLM usage requires separate observation.

Completed

• First actual Go array preparation69 on original72: Project1, retries0, fits0. Preserve90 train/36 validation rows, including18/1 zero-loss-weight rows. Keep63 unknown candidate truths nullable and calibration outside fitting.
• Author-side and separate saved-array reviews passed original supervision/column checks. Preserve the checker's candidate-count assumption failure and later clerical correction. This is not feature re-extraction or a model-quality result.
• Append four finite upstream68 requests, then recompute whole76/226-candidate/19-family roles. Fixed seed1729, first Assign1, retries0. Known-containing families11 train/4 validation/3 calibration; unchanged coverage passed. Additions were not attached to historical66 roles.

Reading measurements

Observation	Actual value	Scope
Owned payload of original72 preparation	About1.69MiB	Returned arrays and retained strings; excludes temporary allocation
Same whole-process peak RSS/wall	About53.95MiB /0.466s	One development observation including reading, validation, arrays and JSON output
Raw output / public summary	About7.7MB /97KB	Summary omits full features and cannot replace a training asset
Full76 role-preparation peak RSS/wall	About51.19MiB /0.478s	Metadata and assignment, not fitting or inference

CPU1, soft Go heap256MiB and outside300s wall were configured. Soft heap, payload, JSON size and RSS are different quantities. GPU, Go heap, Codex savings and AI-assisted preparation/review costs remain unmeasured.

First fit

Freeze new arrays for whole76 roles, then run one primary8192-dimensional FP32 fit. Compare against one strongest nonlearned control on identical validation; retain zero-loss-weight candidates in ranking evaluation. Calibration stays outside fitting and epoch selection. Both added upstream families actually landed in train, so this validation is not generalization testing on those libraries.

The prepared FP32 format is32,792B; decoded float64 coefficients occupy65,536B. Whole-process memory and repeated serving cost need actual post-fit measurements. Once FP32 development utility is established, independently compare INT8/PTQ children or ternary STE siblings. At least2400 distinct protected-final requests per domain remains a separate unfinished objective.

Records and usage

Actual preparation69 · Whole76 roles70 · Earlier PR92 CI/Wiki publication proof.

The available nonlearned Go command is riido-shortclaim. Start with request/candidate JSON in usage56. These records do not activate a trained default model. Coefficients stay outside GitHub; only licensed, verified, CI-gated scoped releases go to Hugging Face. 한국어.