amos 4f17b2afe5 Batch 20 (20261027-C-family-20): map batch 15's ad-hoc families to canonical ones
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>
2026-10-27 05:58:17 +08:00

15 KiB
Raw Blame History

dev-code-window — author specification

Version 1; route C original C11 engineering task; source hwe/gi-cache-jit-code-flush; family code-publication; difficulty UNMEASURED. The root screened source authorization, heldout exclusion and cold/local byte identity before assigning this sole base ID. Source scripts were not executed. Source material supplies only the direction; this is not a replayed incident.

Admission: ACCEPTED as a deterministic software substitute. An offline interpreter consumes an execution view of integer instruction words, while publication writes a separate data view. Missing synchronization reproduces stale instructions without machine code, executable mappings, architecture-specific behavior, real time, devices, or network. Exact half-open synchronization ranges and per-slot publication generations remain observable reliability requirements. The task does not claim CPU cache/barrier or hardware validation.

Scope: only tasks/dev-code-window/, this base's subagent directory, and its run directory. Candidate file window.c; public engine, store, interpreter, command reader and driver are frozen grader context. No platform edits or global test runs. Status DRAFT until local and Docker formal gates each repeat all fixtures five times; then CALIBRATED only. No Pi/Qwen sampling.

Two independent initial defects are planned: full installation writes the data view but skips synchronizing the execution view; patch optimistic concurrency compares a global publication serial to a slot generation, rejecting a valid patch after another slot publishes. The latter is outside the default one-slot replay. Separate only-fix fixtures will prove independence. All rejected state transitions, range behavior, generation behavior and interpreter results must be specified publicly before implementation.

Acceptance and evidence matrix will be completed against exact public line references after the contract is frozen. The Python oracle will generate fixed-seed command sequences and calculate expected observations independently; neither gold nor candidate C will calculate grader expectations. The child driver will only emit raw observations. All files and input must exist before starting any untrusted child; afterward parent check will use only captured pipes. Cleanup, if present, runs as candidate child and its return code is checked.

Environment: fixed runtime ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:1519f8daa403998e5401eeb70b973e8aac8346e25a937033bc502dc5de1641a1; image identity confirmed equal to this digest. A read-only tool probe confirmed gcc, g++ and make. Initial /home free space 88 GiB. Public reproduction will use gcc/make only. Scratch root is owned by UID65534 via wrapper hook and will never be chmodded by check; new support subdirectories receive their permissions before the first child.

Public evidence matrix

All paths below are relative to template/. Each row is a separate checked behavior; test combinations and fixed random values are private, while all expected behavior is public. The oracle uses only these clauses. It compares active valid ranges and permits stderr.

# Checked behavior Exact public basis Evidence mechanism
1 Four slots, capacity twelve, zero reset docs/window.md:3 W1 Reset and all-slot state rows
2 Only active valid-range word elements are compared; tails unspecified docs/window.md:3 W1; docs/console.md:6 C4 Variable-width D/X rows; valid fixture poisons tails
3 Every publication advances only its slot generation and global serial docs/window.md:4 W2 Per-command B/S rows
4 Other slots do not invalidate expected generation docs/window.md:4 W2; docs/window.md:6 W4 Cross-slot directed prefix and interleaving
5 Install accepts 1..12 words and replaces existing slots docs/window.md:5 W3 Minimum/maximum and long-to-short installs
6 Install validates the whole program before effects docs/window.md:5 W3; docs/window.md:8 W6 Invalid final, early HALT, opcode and rotation
7 Install writes count words at zero and synchronizes [0,count) exactly once docs/window.md:5 W3; docs/window.md:10 W8 Data/execution rows and write/sync counters
8 Install sets active and length and needs no generation token docs/window.md:5 W3 Metadata rows and replacements after drops
9 Patch inactive error has precedence over token/range/program checks docs/window.md:6 W4 Inactive slot with conflicting token and invalid edit
10 Patch token compares the selected slot generation, returns -1 on conflict docs/window.md:6 W4 Matching, stale, future and uint64 maximum tokens
11 Patch matching active slot checks first+count <=length docs/window.md:6 W4 Last-word, nonzero-offset and out-of-live-range slices
12 Patch validates the whole resulting program, including unmodified regions docs/window.md:6 W4; docs/instructions.md:3 L1 Early HALT and terminal replacement edits
13 Patch preserves all words outside its slice and active length docs/window.md:6 W4 Complete valid data rows after each operation
14 Patch writes count words and synchronizes [first,first+count) exactly once docs/window.md:6 W4; docs/window.md:10 W8 Execution rows and exact store counters
15 Drop checks active before expected, publishes inactive length zero docs/window.md:7 W5 Repeated drop, conflicts and all-slot metadata
16 Drop writes/synchronizes nothing and preserves generation for reuse docs/window.md:7 W5 Drop/reinstall sequence and cumulative counters
17 Declined publications and inactive execution have no state effects docs/window.md:8 W6 Full before/after state except rejected count
18 Active execution uses execution view, increments executions once docs/window.md:9 W7 O rows, executions counter, mask_symptom rejection
19 Execution leaves publication metadata and word arrays unchanged docs/window.md:9 W7 Full state observations after repeated execution
20 Normal mutators and declined commands have zero value/retired docs/window.md:11 W9 O rows on every command
21 Whole program length, terminal zero HALT, no early HALT docs/instructions.md:3 L1 Whole installs and merged edits
22 uint32 modulo arithmetic, signed byte conversion and retired count docs/instructions.md:4 L2 Zero, UINT32_MAX, high bit and random inputs
23 ADD uses signed byte, XOR uses unsigned byte docs/instructions.md:5 L3 Fixed boundaries plus fixed-seed instruction combinations
24 MUL wraps and zero is allowed docs/instructions.md:5 L3 Boundary instructions and random combinations
25 ROT accepts 0..31, rotates left, zero unchanged docs/instructions.md:5 L3 ROT0/ROT31, invalid32 and fixed-seed operands
26 SET uses signed byte and AND uses unsigned byte docs/instructions.md:5 L3 Positive/negative byte extremes and combinations
27 Opcodes7..255 and nonzero HALT operands rejected docs/instructions.md:5 L3 Invalid install/edit program rows
28 Malformed view faults are reported by public interpreter; correct publications avoid faults docs/window.md:9 W7; docs/instructions.md:6 L4 Raw public interpreter output; expected correct views always valid
29 Blank/comment lines ignored; sequence advances only for processed lines docs/console.md:3 C1; docs/console.md:5 C3 Directed blank/comment prefix and per-line sequence
30 Whitespace, CRLF and final line without LF accepted docs/console.md:3 C1 Tabs, CR and no-LF final command
31 Overlong line rejected once then drain and continue docs/console.md:3 C1 1024-byte line followed by valid reset
32 Decimal-only bounded fields and exact arities; no extra tokens docs/console.md:4 C2 Signed/prefixed/overflow/missing/excess tokens
33 Leading zeroes allowed; expected full uint64 allowed docs/console.md:4 C2 Zero-padded install and maximum token
34 Nonzero rc increments rejected once; R resets counters without resetting sequence docs/console.md:5 C3 S rows after rejects and phase resets
35 Exact observation field order, slot order, numbers and active-range arrays docs/console.md:6 C4 Every O/S/B/D/X row compared to model
36 stderr diagnostics allowed and normal EOF exits zero even after rejection README.md:5; docs/console.md:5 C3; docs/console.md:6 C4 Return code check, no stderr gate, valid fixture emits diagnostics
37 Build submitted window.c with frozen public C11 support README.md:5; README.md:7; docs/console.md:7 C5 Fresh make invocation, no reused products

Invalid adapter calls and corrupted metadata beyond the console/window contract are not hidden test inputs. The adapter is frozen public support; no arbitrary internal calls or out-of-contract API pointers are injected. All test operations come through documented console streams.

Original engineering and fixtures

The original engineering consists of window.c (publication orchestration), store.c (separate views and counted adapters), instruction.c (bounded integer language), console.c (bounded streaming parser), observe.c (raw observation), driver.c (dispatch/EOF lifecycle), public type headers, and a necessary Makefile. No other new batch task was read or reused. Source HWE reference was only read; no reference machine-code placeholder was copied. Unique code is counted once from the initial template, excluding blank lines, comments, fixture copies, pure example data and documentation.

Fixture Expected Mechanism
gold PASS Whole merged-array validation and separate imperative install/patch paths; both defects corrected
template FAIL Missing install sync and patch using global serial as expected token
noop FAIL Install returns success without any install or publication effects
cheat FAIL Always synchronizes four words, matching only the public demo width; other lengths expose stale views or wrong counts
cheat_read_grader FAIL formally Tries direct grader paths and parent's cmdline grader path; proceeds correctly only if grader readable, otherwise exits71. Local unreadable/readable behavior recorded separately
only_fix_sync FAIL Fixes installation synchronization; retains erroneous global-token patch check
only_fix_generation FAIL Fixes per-slot patch token; retains missing installation synchronization
mask_symptom FAIL Executes data instead of execution view, masking one stale execution symptom; execution view and synchronization counters still wrong
valid_segmented PASS Segmented virtual-program cursor validates without a merged array; table-driven opcode bounds and a transaction descriptor unite commit phases. Distinct data organization and control from gold. Deliberately nonzero unspecified tails and legal stderr test permissiveness

Initial business code: 635 nonblank noncomment C/header/necessary-Makefile lines, measured once in template (no sample multiplication). Per-file count and exact counting script are recorded in report.json and the local code-count evidence. A second conservative count removing brace-only lines is retained as auxiliary information, not used to multiply fixture copies.

Verification architecture and stages

Private files are check.py and context.json. Public support is frozen as strings in context.json. verify.py is byte-for-byte rollout/bos_verify.py. No isolation-blocked marker exists. All input and expected rows are generated by Python before candidate compilation. Candidate bytes are copied into a new scratch build; compilation and execution are argv subprocesses, UID65534 under the formal wrapper. Driver contains no hidden assertions or expected results. The parent does not access candidate-controlled filesystem paths after the first child. Scratch root is never chmodded. Only newly created owned subdirectories receive mode0777 before any child. Cleanup runs as a candidate-identity Python child, and its return code is required before PASS.

Fixed seed 0xC0DE710 combines directed public and alternate prefixes, program/range/parser boundaries, and ten phases of four initial installs plus ninety random interleaved commands. Python model computes all expected uint32 values, generations, counters and active-range views. Six independent model behavior unit tests cover arithmetic constants, validation, cross-slot tokens, atomic edit failure, drop/reuse, and parser/ignored-line semantics. These local tests are not a replacement for the two formal calibrations.

State: CALIBRATED (difficulty UNMEASURED). The exact formal gate facts are recorded below; all digest-covered assets stayed frozen. No real Agent, native tool trajectory, hardware or measured task difficulty is claimed. No global unittest, commit or push performed.

Completed formal calibration

Current task digest 1d2f8eb1490f3ceeb6d3425a0577655bc79839e0531f34a64afef6f79aeef95b is identical in both formal receipts. Image ID sha256:1519f8daa403998e5401eeb70b973e8aac8346e25a937033bc502dc5de1641a1 matches the fixed image. No calibration retry or environmental failure occurred.

Local attempt: runs/20261023-C-hwe-scale-15/gi-cache-jit-code-flush/local/calibration-7a7722618d4243139b19a3119a3eefd5/receipt.json. Actual overall FAIL, cleanup confirmed. All 45 trials completed: gold and valid_segmented each five PASS; each ordinary negative five FAIL; cheat_read_grader five PASS because the trusted local identity can read the copied grader. These are exactly the five explicitly permitted local mismatches; the receipt was not altered.

Docker attempt: runs/20261023-C-hwe-scale-15/gi-cache-jit-code-flush/docker/calibration-e22fe9e55ca84ea4be3d42c333d0f3b1/receipt.json. Overall PASS, 45/45 expected outcomes and cleanup confirmed for the attempt and every trial. Gold and valid_segmented each five PASS; all seven negative fixture classes each five FAIL. Ordinary negatives build and execute, then fail raw observations. cheat_read_grader also builds; its designed unreadable-grader exit71 is separately recorded as a controlled negative behavior, not infrastructure failure or observation mismatch.

Final count excluding public headers is 537 original nonblank noncomment C plus necessary Makefile lines. Header definitions add 98, giving 635 original code/header/build lines. No fixture copies are multiplied. Complete per-file counts, 37-row evidence matrix, 1,025 input lines / 1,018 processed commands / 14,252 raw observation rows, six model test results, exact commands and receipt summaries are in this base's report.json. Read-only self-review found no substantive issue. Final /home free space 87 GiB. No real Agent sampling, full-suite run, hardware evidence, commit or push was performed by this subagent. 族名规范化:code-publication → linux-concurrency(2026-10-27)