Direction after rereading Laya57

On2026-10-02, reviewed SDK commit 4aa6761be8173de4ce6d92c31b3e40b6eaf59a7c. Bodies of19 public files matched their Git blobs. No model execution or weight download occurred. This review is not a measured performance improvement.

Our target and the reference implementation

The original Router selects a Laya checkpoint from explicit model/task and language signals. It is not a trained predictor of Codex completion or usage. The typed decoder provides choice argmax, an expected ordinal score and noul P(true). This differs from free-form generation; rounding a score need not produce its argmax.

Riido prioritizes repeated hints about candidates to verify using extremely little resource. Preserve all candidates and fallback; existing verification tools retain truth and acceptance. The next criterion is reducing checking work beyond BM25 across diverse public behaviors, rather than increasing the number of model variants.

Three inexpensive reference points

Reference	Subsequent scope	Verification
Score meaning and abstention	Use the confidence implementation to distinguish probability/heuristic, calibration and passed/abstained/unevaluated in small metadata	An unconfigured gate is not passed; nonfinite/missing scores cannot pass
Support checks before heavy loading	Review consistent English checkpoint language guards across repository preview and task routing	Unsupported scripts cause zero loader calls; preserve explicit override; Latin script does not certify English support
Repeated-result caching	A separate bounded experiment informed by the cache example, keyed by raw text, ordered candidates, implementation/model revision and temperature/budgets	Negation, </<=, candidate order and revision changes miss; test mutation, byte limits, eviction and lookup costs

These are proposals. Do not retroactively modify public evidence or claim cache, speed, memory or fee savings. A normalized feature SHA can collapse punctuation and case semantics. Cache lookup might even exceed the current microsecond Go scoring cost.

answer_confidence=max(p) already existed in the earlier SDK pin. It differs from entropy-based choice/score confidence; noul uses max(p). Neither is domain success probability without independent checkpoint/option-count/domain calibration. Earlier entropy descriptions apply only to that choice/score field; preserve historical snapshots.

Do not rebuild existing Go choice/noul, truncation checks, resident ORT, BPE cache, none/ambiguous, fallback or JSONL as a new engine. Go Predict currently lacks score support, so not every reference primitive is ported. Do not remove locks protecting mutable BPE state and Predict/Close lifetime merely because of the reference Router's lock placement. Its Agent has a separate inference read lock.

English421M/multilingual322M are total model sizes; their backbones are395M/307M. A small ternary head does not remove retained backbone RAM. Multiple questions reuse state tokenization while retaining question-specific encoder inputs. Do not cache question-dependent hidden states by state string alone.

Benchmarks and licenses

The original tier-routing example is a policy example. The application benchmark routes399 GSM8K/MBPP/AG News domain labels, not actual Codex completion or costs. BENCHMARKS explicitly records missing committed evidence for a fine-tuned row and weak base results. Timing reported on another GPU is not our Mac/Go performance.

The SDK LICENSE declares Apache-2.0. Preserve license, applicable notices and modification notices when copying or translating. Model-card declarations were checked separately.

Card	Reviewed revision	Card declaration
English	55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851	Apache-2.0
Multilingual	e4e9ddf21a7b1903b7acffd8814ad4307bf63a67	Apache-2.0
Typed decisions	1a793eb568e6718f15941d08f85432581df534e3	Apache-2.0

This verifies public declarations, not all training-data rights. Do not import teacher labels as independently established truth. New paid model calls, fits, weights and final evaluation remain0. Preserve at least15 groups,5% necessary utility and2400 distinct protected-final requests per domain.