210 batch-15 tasks used 119 ad-hoc family names. Each now carries one of the canonical root-cause families (11 of the 19 are used), with version + 1 and a spec.md note, and was recalibrated through both gates. Claude checked that only family, version and spec.md changed, that every family is canonical, and the new trace6 Docker receipts against current digests (210/210 USABLE). Squashed from the 36-local branch batch20-family. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
4.6 KiB
quant-wide-decode specification
Route C original task. source: hwe/gi-quant-scale-overflow. Family: numeric. C++17. Difficulty: UNMEASURED. Status: READY_FOR_CALIBRATION.
The HWE material supplies only quantization range/precision subject matter. Its purported reference performs int32 subtraction before casting and is not a correct reference. This package is independently authored and does not claim a reproduced field incident. Software substitution replaces BPU outputs with deterministic signed 32-bit frame samples; the same numeric reliability failures remain, with no board, network, timing or memory corruption requirement.
Two independent defects: signed 32-bit difference evaluated before widening; narrowing decoded arithmetic into float. The public demo exercises the first; other allowed profiles expose the second. only_fix_difference and only_fix_precision each retain one defect. Gold widens to int64 then computes binary64; valid_signed_magnitude uses a separate ordered signed-magnitude construction and stages a per-channel table before publishing an accepted frame.
Acceptance is entirely the complete public contract in template/docs/contract.md and C++17 public header. The final criterion-to-exact-line mapping is appended after contract freeze. All candidate behavior is observed by a public stdin/stdout observer and compared to an independent Python model, never a gold executable. Defined result prefixes only are compared; stderr is unrestricted bytes. Grader safety checks cover identity, actual child deadlines, descendants, cleanup and independent strict lstat absence. Positive and negative fixture classes each repeat five times at the local and pinned-image formal entry points. No real Agent rollout, board test or measured model difficulty is claimed.
Acceptance mapping (exact public basis)
| Criterion | Public evidence |
|---|---|
| Build, submission file, normal return and unrestricted stderr | template/docs/contract.md:3 |
| Full signed-domain decoding, finite scales/bias/range and numeric tolerance | template/docs/contract.md:5 |
| Construction/reset defaults and retained history on profile replacement | template/docs/contract.md:7 |
| Profile validation, mutation precedence, removal, configured and revision counters | template/docs/contract.md:9 |
| Begin/append validation and capacity, active/pending state and input order | template/docs/contract.md:11 |
| Atomic keep/reject/clamp policies, outside flags and accepted/rejected history | template/docs/contract.md:13 |
| Cancel and frame closing | template/docs/contract.md:13 |
| Per-channel and total statistics, accepted arrival ordinals and binary64 fold | template/docs/contract.md:15 |
| State/profile/statistic/value identities, valid prefix and nonmutation | template/docs/contract.md:15 |
| Public command arity, lexemes, whitespace, invalid input and -9/-1 behavior | template/docs/contract.md:17 |
| Exact raw observer row fields, order, integer/float grammar and termination | template/docs/contract.md:19 |
| Repeat/reset, all channels, empty/full frames and other allowed configurations | template/docs/contract.md:21 |
Public API signatures and field types are frozen in template/include/engine.hpp. Observer context is frozen as literal strings in check.py. No private expected value or assertion is compiled into the observer. Private model keeps an accepted history journal and recomputes statistics independently of engine mutation, using Python arbitrary-precision subtraction before binary64 multiplication/addition. The two numeric mutations are exercised independently by wide difference and range/precision configurations.
Author validation scope
Six private author unit tests cover public lexical values, atomic model behavior, live PID/start closure, fresh signal identity, and metadata rejection. Revised fixed-image trusted core._grade diagnostics cover 21 actual packages, including production deadlines (compiler20s, candidate8s, cleanup4s, absence4s), 17MiB stderr, signed 5000-leading-zero output, setsid/closed-stdio descendants, per-iteration empty-OSError signal retry/reap, kernel EACCES versus simulated OSError, partial success flags, missing wire field, bool UID and nonzero wrapper. Prior incomplete draft evidence remains excluded. No full repository suite or real Agent sampling was run by this author; root owns the shared suite/review. PG absence is derived from owned ancestry closure; supervisor identity is observed, separate leader credential sampling is not claimed. Formal large IO may retain length/hash rather than original wire; diagnostics explicitly identify saved wire/raw bytes. 族名规范化:numeric → numeric-reliability(2026-10-27)