Short-claim experiment56b: preparing behavioral truth and connected groups

한국어 · Acquisition direction · Recorded56a results · Prepared inputs

This stage prepares truth/group auditing. No official audit result has been observed yet. The execution-plan schema is riido-typed-truth-execution-plan-v1, state is preparation_truth_groups_only, and policy is truth_groups_only_no_roles_no_fits_no_models. Freeze inputs, sources and the execution plan before outcomes; then perform one official audit and CI replay. Role assignment, fitting, model calls and performance measurement are outside this stage.

56a implemented a small Go tool without a large encoder, but its48 authored requests formed11 connected groups, below the operational minimum15. Increasing the sentence count alone does not erase that failure. This stage prepares separate contracts for error identity, state commits, memory ownership, lifetime and graph behavior that integer arrays could not adequately express. Preserve the original56a text, truth and observations.

Scope fixed before observation	Meaning
24 new requests +48 legacy requests =72 total	Original requests from the same coordinated authoring process, not72 independent samples
6 new behavioral families	Different concrete Go function types and observations, not6 independent populations
24 new source functions	6 correct and18 wrong controls,4 per family
80 new literal vectors	35 state-family and45 flow-family vectors; neither request count nor training-sample count
320 new control checks planned	Apply each family's vectors to its4 sources: 80 × 4; no official result yet

New and legacy data come from one coordinated synthetic authoring process. They do not establish diverse authorship or production code from multiple people. Review fidelity of the English requests; this Korean/English documentation does not add evaluation examples in another language.

Each request and candidate stays within512 bytes for both raw and normalized text, and32 normalized words. Do not hide scope through automatic summaries or silent truncation. Preparation uses nonlearned, CPU-only Go1.27.1; a256MiB Go-heap soft limit is not a hard cap on total OS RSS. Distinguish the small request tool's fixed arrays from metadata maps used by offline go/types provenance checking. Labels and source information do not enter ranking features.

Family	Behavior to observe against independent literal tables
Atomic configuration commit	Commit only successful results; distinguish malformed input by errors, zero result and preserved destination
Error identity	Observe Is/As through wrapping/joining and original cause values; equal messages alone do not establish identity
Owned snapshot	Values, length, nilness and mutation isolation in both directions; observe input mutation during copying and later aliasing
Cancellation/lifetime	Acquisition, result and cleanup order/counts; prior cancellation, repeated/post-terminal events and status
Quoted/escaped delimiters	Quotes, escapes, empty tokens, input/token limits, error kinds and partial output
Cycles/shared DAGs	Distinguish reachable ancestor cycles from shared nodes; duplicate edges, unreachable cycles and unchanged input

Actual behavioral truth and short-caption fidelity are separate checks. Literal expectations are authored independently so candidate functions cannot generate their own expected results. A safely observed candidate panic violates a no-panic contract; unreliable observers or unsupported scope remain unknown. Never convert unknown into failure or success labels.

Independent reading marks5 families as development-caption READY. Keep delimiter code controls, but prepare all4 delimiter requests as unknown. Three cannot express the complete contract and use unknown-incomplete-caption; one has an ambiguous policy and uses unknown-ambiguous. Do not infer aggregate known/no-answer/unknown counts or actual group counts from these preparation decisions. The official audit will retain the original denominator and exclusion reasons.

The source-closure audit pins referenced authored types, globals, methods and helpers alongside function bodies. Unresolved authored references or unsupported initialization stop the audit. Standard-library and observer support are excluded only from group relations under an explicit policy, while remaining in provenance. Transitively join shared helpers, copied behavior, prototypes, counterexamples, candidates and parent/sibling relationships. More conservative combined grouping does not overwrite the historical56a result.

Require at least15 each for total connected groups and connected groups carrying clear labels. Labeled groups must include known/no-answer data; unknown-only groups do not count. Publish the two gates separately. If insufficient, stop preparation without splitting groups or lowering minima. Passing both does not establish statistical independence, training eligibility, generalization or performance; roles and fits remain0 at this stage.

The execution order is:

1. Commit the new source closure, literal tables, inputs, caption review, group policy and execution plan before outcomes.
2. Perform one official truth/group audit with the fixed Go executor. Separate the legacy48-request record from the new combined72-request relation report.
3. Publish integers, acceptable sets, unknown reasons, relations/groups and input/source digests, then replay them in CI. Record actual commits, manifest digests and CI links in the subsequent results document.
4. If checks fail, record the cause and a new freeze boundary when correcting it. Do not adapt truth, the primary criterion or minimum groups to observed outcomes. CI remains the verification gate; add no separate human-approval flow.

Even after preparation/auditing, keep training roles0, fits0, new weights0 and actual Codex calls0. Do not change the cumulative paid-call ledger. These data are not final-evaluation eligible. Preserve the existing separate minimum2,400-request final-evaluation goal, protected final and CoSQA reserve; neither read nor score them. Publish no model bodies, private code, real work inputs, personal paths or raw traces in this stage.

Compressing a whole-function contract into32 words can omit exceptions, error kinds or supported scope. The next hypothesis is to ask about one property with a fixed scope, rather than the entire function. A separate contract might ask only whether valid shared children are allowed without a cycle rejection, or whether a delimiter inside quotes is preserved. This remains a proposal: freeze a new input contract, independent truth, unknown conditions, source closure, caption review and group relationships first. Passing a partial property never certifies whole-function correctness.

Preserve relationships between prototype parents, translations, paraphrases, counterexamples and sibling requests. In a later stage ready for fitting, INT8/PTQ children of the same parent fit and different-seed/STE siblings are not independent data. Do not execute the existing56 proposal of at most8 fits/16 derivative artifacts until a separate training execution plan is ready.

Resident --stream, identical/distinct requests, pipes, queues, open-loop load and cache costs belong to a separate resource stage. Do not reinterpret56a's sequential16-child replay as concurrency or resident-load evidence. The kevent profile attribution remains unresolved. This plan creates no new CPU/RSS, speed or LLM-savings evidence.