Finite-input claim audit56e: plan before outcomes

Before fitting a tiny model, align the promise in the text with the observation scope and derive limited truth from actual source behavior. Preparation56c checks nine specific inputs while its pending prose describes general behavior. Do not approve that version in place. Add a separate finite-input v2 contract through the Go finiteproperty package and riido-propertyaudit command. Useful tiny-model training and validation remain unfinished.

Separate contract and public prose

The directly authored request states exact ASCII inputs in hex. For example, 0x22612c62222c63 spells the bytes of "a,b",c. The existing punctuation-removing normalizer retains distinct hex words. This does not prove that a model understands them or that normalization limitations are solved. Preserve56c's prohibition on literal features: v2 is a different problem with a visible finite input set in the request, not a repaired score for old prose or a general function certificate.

The prose is AI-assisted public synthetic material, independently reviewed by agents reading the source. Do not claim human authorship or approval. The evaluator never automatically inserts hidden literals or expected observations into text. Candidate captions contain no source IDs, expected arrays, control roles or observed success/failure. Only request/candidate prose is prospective feature input. This execution has no ranking, numeric feature encoder or learned model.

Fixed scope and truth

Preserve original candidate IDs, order, code and source-closure hashes. Four original parents times three properties produce12 derived rows and36 candidate captions. Reuse the existing authored2/2/5 inputs and expectations from scopedproperty.Definitions() unchanged. Its pending describes the original record, not this separate prose review. Decode the publicly declared hex and require the same input count/order/bytes. Never fill missing prose from metadata or automatically generate the request.

Execute only four registered sources through typedbehavior.ObserveQuotedProperty:36 actual source/input calls, referenced by36 candidate labels in12 rows. Whole-function correct/wrong roles and original acceptable sets do not supply property truth. Compare exact Error, Count, all eight Tokens and Panicked. A syntax input requires the exact syntax error; OK with zero output mismatches. Count0 with a retained token and an observed panic also mismatch. Unsupported observers and corrupt bindings stop rather than creating false labels. Agreement applies only to the listed input set.

All sources connect through the same scanQuoted helper family. Add zero independent parents/families. Preserve56b's72 parents,17 total/16 labeled groups and four unknown delimiter parents. Neither12 rows nor36 calls are new independent requests or2400 protected-final examples.

Freeze and execution order

1. Commit implementation, plan, independent prose review, preparation unit controls, modules and historical56c artifacts as the source freeze. New controls validate exact comparisons, tampering, bounds and scope without candidate calls. Existing regression tests already observe historical controls; this is not a source never previously observed.
2. Build Go1.27.1 CGO0 with -trimpath -buildvcs=false -p=1 once. Put the actual executable SHA and size in the plan, never the executable in Git. Verify21 compiled sources and13 support artifacts against files and actual Git blobs.
3. prepare emits dataset/plan with no candidate observations. Freeze them as probes-56e.json and execution-plan-56e.json in a separate input commit.
4. Before observations, audit --input-commit verifies the same executable,34 source/support blobs and two input blobs,36 in total. Canonical JSON byte equality rejects duplicate keys, extra fixed-array entries, unknown fields, null substitutions and trailing objects. Automatic retries0.
5. One official audit preserves inputs/expectations/observations for all36 calls and12 row results. Subsequent CI replay checks that record, adding no samples or official attempts. Reusing an output directory fails before observation. Executing into a different directory is a separate attempt and must not be confused with the official ledger.

GOMAXPROCS=1 and256MiB soft Go heap are settings, not measured RSS/CPU/GPU or a hard bound. Each artifact is limited to1MiB. Each prepare/audit preflight performs192 length-validation normalizations: typed24×4 + pending12×4 + finite12×4. Numeric encoders/scoring/ranking0. Source, parent and evaluator metadata are not inserted into normalized text.

Unchanged training requirements

Official ranking/utility, roles/splits, fits, new weights, model calls, paid calls, protected-final/CoSQA reads and activation0. Retain minimum15 total/labeled groups,5% necessary utility and at least2400 distinct protected-final requests per domain. This finite family alone does not establish training eligibility. Next acquire licensed fixed-revision/closed-source independent public contracts toward120→240 and separately freeze fixed/BM25/lexical/rule/oracle utility while preserving every candidate. Existing upstream behaviors are semantic-observation material, not evidence of newly completed coding tasks.