Public behavior verification57 results

Collected62 actual original Go API observations:59 matched prewritten finite expectations and3 disagreed. Two cover the same large-prerelease ordering failure in opposite directions. The other contrasts early glob success with a stronger validation policy. These are not model accuracy, general function correctness or2400 independent task results.

Findings

Observation	Prewritten expectation	Original result	Meaning
1.0.0-99999999999999999999 vs 1.0.0-100000000000000000000	SemVer numeric precedence: smaller(-1)	Larger(+1)	Both strict parses succeed; uint64 overflow enters the comparison's string branch. Preserve this finite normative discrepancy.
Reversed pair	+1	-1	Confirms the same pattern, without adding an independent bug or task.
Match("{a,[}", "a")	Stronger whole-pattern policy: false + ErrBadPattern	true + nil	The first alternative succeeds before the malformed later alternative is checked. Do not generalize this into a universal upstream API obligation.
ValidatePattern("{a,[}")	false	false	A separate observation. No validation gate was inserted before Match and no result was synthesized.

These findings show why broad candidate claims such as numeric version comparison or pattern-error handling need small claims with explicit input scope and failure conditions connected to real verification. Tiny-model semantic understanding and cost savings remain separate questions.

Counts and independence

• Strict parse16, metadata comparison5, prerelease comparison9, path matching9, escape/character matching8, error matching6 and pattern validation9:62 operation observations.
• Requested entry APIs90: StrictNewVersion44, Compare14, Match23, ValidatePattern9. Getter/String/Original calls224 are separate. Upstream internal helper calls are not counted.
• Normative expectations14:12 matched/2 disagreed; descriptive API46:all matched; stronger policy2:1 matched/1 disagreed. Known library errors17 were accurately observed; unknown/panic/unsupported0.
• Verified28 Git blobs before observing. Official collection1, official retries0, exit0. A plan preparation attempt rejected the symbolic HEAD value with zero observations; it is not an official retry.
• Two source families and six finite property contracts. Inputs, aliases, reversed pairs and contract variants do not add independent final tasks. New model inputs/captions, ranking evaluations, fits, paid calls, weights and protected-final tasks0.

Preserved15 original MIT files totaling99,035 bytes and both full licenses. Original module declaration bytes live under metadata/upstream.go.mod.txt; parent go.mod/go.sum are unchanged. These upstream behaviors already exist and are not six newly completed coding tasks. See the source manifest, prewritten plan and authoring/review record.

Freezes and reproduction

Evidence	Bytes and SHA256
Probes	218,192 / c4deb9d9e33dca8ff4498531d898ea82ee05559727cbc4f00d99455910722574
Execution plan	5,472 / dfc17915b37d4ce007f48a3a8d4e6fcc693f926f6a789562f6a8fcc68c347d7b
Original observations	262,559 / 850c60266eceb613565390430adca12cea9b1a56207b803b06cb681f563978c5
Actual Darwin/arm64 Go1.27.1 CGO0 binary, excluded from Git	5,096,818 / 09e5f6079b8a436c432d313a884376a5bde2267e7afdde88c5d422f1777cc5d7

Source freeze 9e99914f6d1e35aa9d97413f80fc6e03768e97ce; input freeze 1462c705087f24666049f8bd31d317490c05dbeb. The three JSON files total486,223 bytes. Verified19 embedded source/upstream artifacts and7 support artifacts before observing. Post-collection publication_test.go and audit_test.go are not retroactive pre-collection pins. Regression replay adds no official observations or independent samples. Usage explains platform boundaries.

Scheduling1 and heap soft limit256MiB are settings, not measured RSS/CPU/GPU/latency. This run executed no Laya/GPU and adds no dependencies or locks to the production hint path. Preserve earlier56a/d measurements and pending/unknowns. Cache and score-metadata suggestions in the Laya retrospective are not achievements of this run.

Next acquire accurate requests/captions and candidate relationships across more sources, then freeze and assess model-free ranking utility while preserving every candidate. Do not claim240 acquired sources, minimum15-group/5% necessary utility or2400 distinct protected-final requests per domain. Training eligibility remains false; no new HF weights were released.