PDCA54: preparing acceptance contracts for two public Go candidates

한국어 · Machine-readable record

Two requests from the 120 public Go source candidates now have locally verified acceptance contracts. This is not evidence of a model solving them or a ground-truth difficulty label. A verifier rejects the original source, accepts an independently authored correct change, and rejects known incorrect changes.

This document records development evidence at preparation time. CI-verified public execution support and publication evidence must be checked separately. Local registry and offline verification integration are implemented, but this record does not yet contain the corresponding public CI or merge evidence.

Category	Count in this record
Candidates in the source investigation	120
Locally verified acceptance contracts among them	2
Actual model attempts for these two requests	0
Training runs or training labels for these requests	0
Requests eligible for protected final evaluation	0

The contracts develop requests already among the 120; the candidate count is not 122. The original PDCA53 investigation and 120-candidate inventory remain unchanged. That inventory's SHA256 is e79cf90505fff4c636de99f322ab82001afd22516216cebcd8249b9b3d61cf67. Separate laya-tools tasks and model pilots are not added to these requests' execution counts.

1. Signed ordinals across the full int64 range

go53-humanize-ordinal64 adds Ordinal64 to the pinned dustin/go-humanize source. For example, -21 becomes -21st and -11 becomes -11th. The minimum value -9223372036854775808 must keep every digit.

The original provides only Ordinal(int). Its negative inputs use th, so Ordinal(-21) must still return -21th; only the new API selects suffixes from the magnitude. The contract checks the callable type func(int64) string; it does not distinguish a declared function from a function-valued variable.

The five copied original files are go.mod, ordinals.go, ordinals_test.go, the required common_test.go helper, and LICENSE. This is a closed ordinal subset, not the entire upstream package. Only ordinals.go is mutable.

The original MIT license is retained unchanged. number.go, with its separate WTFPL origin, is neither copied nor built. The actual source has no NOTICE, and none was invented. This repository's Apache-2.0 notice does not replace upstream MIT.

The oracle uses arbitrary-precision math/big arithmetic and an explicit table of 100 suffix residues. The authored correct control uses decimal string tails, so it does not copy the oracle implementation. Checks cover both signs, the 11/12/13 exceptions, every suffix residue, inclusive int64 endpoints, 32-bit and floating-point precision boundaries, and fixed wide inputs.

Actual offline verification rejects the baseline and accepts the correct control with six required terminal passes covering original and independent tests. Thirteen incorrect controls are rejected. They include a comment-only change with no new API and a wrong-signature compilation failure, so not all thirteen are counted as behavioral detections. An independent reviewer accepts a separate safe unsigned-arithmetic implementation and rejects two controls with incorrect large teen or minimum-value suffixes. Deliberately broken candidate-authored tests are never acceptance evidence.

2. An exact lowercase UUID text parser

go53-uuid-canonical-parse adds ParseCanonical to the pinned google/uuid source. It accepts exactly 36 bytes of lowercase ASCII hex and the four fixed hyphens. Every failure returns an error and a zero UUID, including failure after a valid prefix. Zero, maximum and arbitrary version/variant bit patterns remain valid.

Existing Parse and ParseBytes retain uppercase, URN, raw 32-digit hex, wrapper compatibility and error categories. The brief request in the source investigation is preserved; the development contract separately refines compatibility and the supported implementation shape in its own prompt.

The frozen closure has 23 original Go files plus go.mod and the BSD-3-Clause LICENSE: 25 files. Original Google attribution remains intact, and the actual tree has no NOTICE. Source redistribution must preserve the copyright, conditions and disclaimer; binary redistribution must provide the notices required by the license. This is not blanket clearance for other files, dependencies or future model releases.

The verifier supports the exact 10,254-byte original uuid.go prefix or its fixed 10,251-byte Go1.27.1 formatter result, followed by only one new function. Normalization permits only the verified formatter whitespace differences. Supported code uses local values, loops, the original pure Parse/xtob helpers and selected pure standard calls. Global writes, direct entropy reads, calls through function aliases, pointers/global-array aliases, added helpers, function literals, concurrency, import/init/directive changes and attribution edits are unsupported. Declaration objects distinguish an inner local declaration from a subsequent write to actual global state.

Unsupported code is verifier_unknown, even if its behavior is correct. It must not become a model failure or low-capability label. This is not a general effect checker for Go programs.

The original test suite records 212 terminal passes on the current host. The baseline lacks the new API and fails the independent contract. The correct control passes the original 212 plus six independent tests: 218 total. Fourteen incorrect implementations within the supported shape each terminate all six independent tests and are rejected for behavior. Twenty-one unsupported controls are assessed separately by the shape gate.

Independent CLI review accepts a separate hex.DecodeString implementation with 218 passes and rejects an uppercase-permissive implementation. Direct entropy access, hidden global pool writes and deleted Google attribution become unknown before execution. Fixed vectors, the 36-position × 256-byte matrix, repeated/concurrent calls, randomness-reader and pool-state checks are controls for one request. The 9,216 input checks and 218 terminal passes are not separate tasks. Reader instrumentation alone does not prove absence of every entropy path; direct calls are also blocked by the constrained source gate.

See the separate UUID source and contract record for details.

Checking the contracts and their identities

In the development checkout, build the executable once, then read the specs without launching a model:

mkdir -p .cache/bin
go build -trimpath -o .cache/bin/riido-taskverify ./cmd/riido-taskverify
.cache/bin/riido-taskverify --task go53-humanize-ordinal64 --spec
.cache/bin/riido-taskverify --task go53-uuid-canonical-parse --spec

To verify a candidate, supply the pinned public source and its candidate copy. A trusted, working Go1.27.1 installation is required. Recorded executions use macOS arm64; they do not establish every platform.

.cache/bin/riido-taskverify --task TASK --base-dir PUBLIC_PINNED_BASE --candidate-dir PUBLIC_CANDIDATE_COPY --go-root TRUSTED_GO_1_27_1

The executable above returns exit 0 for accepted, 1 for rejected, 2 for invalid input and 3 for verification unavailable. Original module bytes remain pinned; only the temporary compiler module uses the original module identity and Go1.27.1. Execution fetches no external dependency and inherits no authentication, user home or network access. Test elapsed time is not model latency or cost.

Frozen identity	Humanize	UUID
Spec SHA256	5c6fd6058a48bd0e6441a68c351e7b1ed491aec20baff78c6942e377ae79f5d5	c8e3a2edcc2024ce68ba7ea7b5fed00554b2e1c36a12174f942f32b5a5a5eec3
Definition SHA256	99a7c0723e3c9f40a069003ccf3f9c2e290c3cbe8e67be3a4ced02d4a31a29b8	9413ed54a10ee9a1576f71a68cf664134570cf84e735f8daeb88c66c1f4ad7c8
Contract SHA256	3b1351162b746fa61f7837af2923644eb2f4df100de643114c9d9f0eb064fd06	259f3fcd56ac9bf0d4d307cc1b8ed77968b6b3bf21327160e09efd440f71ffd5
Prompt SHA256	107ae4d6da023114cc615ebdef6709053be8471bb0ee37f9763af73f279a94a9	5c6456db8da12a5e9510cf9f39a448298df4dd73bead5706a41288f567f30cb2

Per-file SHA256, Git blobs and implementation references are in the separate JSON record. Contract tests are finite development evidence, not formal proofs over all int64 values, programs or environments. License review of these copied source scopes is not extended to model training, weights or dataset distribution eligibility.

Work remaining for the 2,400-request target

The target is at least 2,400 distinct protected-final requests per domain. The 120 source candidates and two development contracts are neither an answer set nor achieved final coverage. Material already read in development is excluded from protected final evaluation.

Public CI evidence and a fresh model-attempt plan must first be frozen. Actual whole-request model execution, independent completion checks, usage and failure status are then needed before considering capability labels. Cost and the best routing profile remain unknown. Repositories, copied origins, shared code and parent/child/sibling requests belong to connected groups; train, selection, calibration and final material require separate acquisition. Translations, repeated attempts and residue/byte controls stay with their original request group.

Model upgrades/downgrades, repository selection, task decomposition and small nonbinding hints each need their own requests and oracles. These two Go acceptance contracts do not substitute for those domains' training labels.