Small Go claims: implementation and nonlearned diagnostics56a

한국어 · Usage · Numeric result · Resources · Profile summary

An encoder-free Go preview is implemented and measured. The48 authored requests form11 connected groups, below the minimum15. Partitions, fits, weights and model calls are zero; no actual LLM savings are claimed. Preparation56 remains the historical snapshot with zero data and measurements at its freeze.

Implementation

riido-shortclaim suggests an independent verification order for bounded requests and candidate descriptions. It validates lengths, Unicode, duplicate JSON keys and counts while retaining every candidate. It provides a resident JSONL --stream, pkg/shortclaim, and four nonlearned controls. Python, model downloads and external API keys are unnecessary. Existing Laya difficulty routing and Codex defaults remain unchanged.

The controls are fixed order, BM25, word/ordered-bigram matching and an explicit behavior-v1: grammar. Rules are not guessed from prose. All48 natural-language development requests are out of grammar scope, producing 48/48 BM25 fallbacks for narrow rules. Fast finite-grammar matching does not establish general language understanding.

Pre-outcome freeze

Commit45b2ad3 sealed inputs, executors, sources and the nonlearned plan before official score/performance collection.

Item	SHA-256
Execution plan	9c8461b9fe0a80a0f0976c8f07a68f3a47f0c534800a0a303e1cbd8a613caaeb
Original48-request fixture	0fe97dd65606f4239ef9187f1fc4d9c8dc2fcaebf342612b2071220896b92df0
Go functions and independent literal tables	b377343a64c019c205f69865286a21ce5d91dda3051c31aa5c69b75a40300559
Eight-candidate performance file	bc41dc423139be4adebc0687ffd30237f22458ea58fccf58771e2003276b5155
Measured Mac binary	5f5463005f9dd482247a5f028af9c5c6e6d7cec2c04aa18e4605e8a45a278fb2

The plan also pins eight runtime sources and Go1.27.1. Builds use CGO0, trimpath and buildvcs=false. The driver checks binary bytes, Go identity and package ID, then executes a private copy of those verified bytes. Audit decodes the exact input bytes whose digest was checked. Public output excludes personal paths, full environments and raw traces. Output bounds, a shared60-second replay deadline and15-second child deadlines apply.

Data and truth

The authored dataset contains 48 parents,144 candidate descriptions,12 prototypes and11 connected groups, all English from one author. It covers conditional retention, authorization negation, inclusive boundaries, summed quotas, stable filtering, first removal, adjacent compaction, rotation, nondecreasing order, clamp, prefix balance and arithmetic order. Variable/constant changes, translations and model settings are not independent requests.

Twelve correct and24 wrong Go functions are checked against68 literal vectors not generated by candidates. Applying each prototype's vectors to three functions yields204 base control checks; candidate-set checks are recorded separately. Source and normalized-code hashes, core templates, shared requests and negative candidates join transitively. Closed-window and clamp-window share a closed-boundary core, merging12 prototypes into11 groups.

Natural-language fidelity was independently read apart from Go tests. Under-specified acceptance and adjacent-compaction phrases were strengthened before scoring. Finite vectors and one independent reading do not certify general semantic truth. Source, labels, evaluator modes, roles and IDs are excluded from scorer features; only request/candidate descriptions enter scoring.

There are24 answerable requests,12 verified no-answer sets and12 unknown ambiguous policies. Unknowns remain in the original48 denominator but are excluded from utility and Top1/3. No-answer sets exhaust all candidates for both baselines and oracle. Acceptable sets may contain multiple candidates.

Nonlearned observations

These are unpartitioned development diagnostics, not holdout or actual Codex task results.

Control	Checks across36 known requests	Checks across24 answerable requests	Top1 /24	Top3 /24
Original order	72	48	4	24
BM25	59	35	15	24
Word/ordered-bigram	60	36	14	24
Narrow grammar→BM25 fallback	59	35	15	24
Oracle knowing acceptable candidates	48	24	—	—

BM25 is strongest, averaging59/36≈1.639 checks including no-answer sets, against the oracle48/36≈1.333. Maximum possible relative improvement is (59−48)/59≈18.64%; answerable-only gap is (35−24)/35≈31.43%. The oracle is not achieved model performance, a deployable policy or a feature. It clears the necessary5% upper-bound screen without proving learnability. Authored candidate order and short control construction affect results. Perfect Top3 also reflects small3–4-candidate sets and duplicated acceptable candidates; do not generalize it to semantic competence.

Because11 groups are below15, no roles are assigned and components are not split. Stop reasons are insufficient_connected_groups and no_partition_or_training_execution_plan. FP32/INT8/STE/PTQ fitting, HF weight releases, final scoring and production activation remain zero. Protected final and CoSQA reserves were not read.

Mac resources

Hardware is Apple M4 Pro,24GiB, darwin/arm64, Go1.27.1. GOMAXPROCS=1 limits concurrent Go Ps, not measured OS-thread count or affinity. No GPU, ORT, Laya encoder or result cache is used.

Each control has three fresh cold processes and one warm process on the same authored eight-candidate input. Each of five warm stages has20 warmups,101 allocation probes from AllocsPerRun, and2,000 timed real-work samples. Repetitions are not dataset samples. The request normalizes to32 words and candidates to at most32; this is not a worst-case test of every512-byte input, language or shape.

Control	Full Go request warm p95	Repeated benchmark child peak RSS	Full request Go allocations/op
Fixed order	0.063417ms	10.5625MiB	291
BM25	0.069667ms	10.2500MiB	291
Word/ordered-bigram	0.077042ms	10.421875MiB	291
Narrow rule	0.079250ms	10.468750MiB	291

Full requests include JSON preparation, validation, ranking, digest and output serialization to io.Discard. Ranking combines revalidation, features, scores and ordering. HeapAlloc is a Go snapshot including uncollected objects, not RSS. Peak RSS covers the entire repeated benchmark child, not instantaneous single-inference RAM. The32MiB/1ms targets pass on this fixture without replacing semantic-quality gates.

The first of12 cold processes took370.932333ms; the other11 took2.564958–5.232667ms. The first sample is preserved without asserting its cause. OS/file caches and initial-launch effects remain unverified. Cold RSS ranged roughly4.53–4.69MiB. Replay wall was1.269162916 seconds. The four warm children used user CPU0.184671/0.200684/0.214853/0.232318 seconds and system CPU0.003933/0.003496/0.003057/0.003607 seconds. JSON retains every observation. Arrival-rate open-loop load, concurrent requests and stream latency remain unmeasured.

Profile and optimization judgment

The frozen plan permits one separate narrow-rule profile with10,000 repetitions per stage; it is excluded from headline latency/RSS. Across1.34 profile seconds and1.06 sampled seconds, runtime.kevent accounts for0.88 flat seconds(83.02%), runtime.madvise0.07, BM250.02 and lexical tokens0.05 cumulative. Do not call kevent an application CPU bottleneck or shared-lock contention. Go issue80921 describes a separate go1.28-devel Darwin syscall stack-attribution report; it was reviewed as context, neither reproduced nor established as the cause of this go1.27.1 observation.

Candidate arrays bound8 slots, word arrays32, and frequencies use small contiguous arrays. There are no runtime maps, shared mutable state or shared locks. Selected Rank computes only its requested control. Normalization, validation and JSON still allocate, so this is not a zero-allocation tool. SIMD/SoA gains and an application hotspot are unproven; first validate profile attribution and stage resource costs.

Reproduction and next work

The performance plan pins a Mac arm64 binary. Do not substitute another platform's hash into the same plan. Linux/macOS CI replays nonlearned truth/groups/order and integer metrics exactly, with1e-12 relative score tolerance only. CI quality gates automatic merging without a separate human approval step.

CGO_ENABLED=0 go build -trimpath -buildvcs=false -o .cache/riido-shortclaim-56 ./cmd/riido-shortclaim
go run ./cmd/riido-behavioraudit --plan-sha256 9c8461b9fe0a80a0f0976c8f07a68f3a47f0c534800a0a303e1cbd8a613caaeb --out .cache/my-short-audit
go run ./cmd/riido-shortperf --plan-sha256 9c8461b9fe0a80a0f0976c8f07a68f3a47f0c534800a0a303e1cbd8a613caaeb --out .cache/my-short-performance.json

Outputs must be new. Publish authored fixtures, numbers, hashes and function aggregates only; raw pprof, personal paths and real-work inputs stay local. New Git JSON data/records excluding the historical preparation snapshot total about260.6KiB, comfortably below64MiB including code/docs. Model binaries are excluded from Git. Next acquisition56b examines atomic updates, error identity, snapshots, lifetimes, escaping and graphs, requiring typed truth contracts and transitive grouping first. At least2,400 distinct final requests per domain remains the separate target; current final eligibility is zero.