PASS G01 strings.gen.ts equals the generator's output from the pinned sources (ast extraction, SHA-256 recorded)
PASS G02 punkt.gen.ts equals the generator's output from punkt_tab/english
PASS G03 coser_reference.py: 17 verbatim copies equal their source blocks (CRLF read as LF, as Python reads source)
PASS G05 protocol version coser-gca-1.1 · 2026-11-01: same in the protocol module, the adapter, this reference, the report and the matrix
PASS V01 test set: SHA-256 eb175779…, git blob edabd925, 200 = 100 id + 100 ood, = test_data id+ood, fields complete, 200 unique ids
PASS G04 the module's SOURCES name the pinned blobs (test set SHA-256 included)
PASS C19 model-family table: 32/32 ids in the expected family; rules equal the official name tests for 7/7 families
PASS V02 agent system prompts and first messages: 1400 setups (200 samples × 7 families), 11348 messages, 0 setups differ (covers 9 samples with the 7 special names, 13 characters without a profile)
PASS V03 seeded choices: identical across 7 actor models for all 200 samples (setup takes no run number); equal to the Python construction (0 differ; 0/800 raw draws differ); another protocol version changes 198 samples
PASS V04a NSP parsing (the text before the first colon, as at 93c1474): 1110 vectors, 0 differ
PASS V04b NSP decisions incl. END before/after round 5 and seeded fallback ({'next': 622, 'end': 37, 'fallback': 2416}): 3075 vectors, 0 differ
PASS V05 simulations {'gpt': 200, 'claude': 25, 'o1': 25, 'llama31-instruct': 25, 'step': 25, 'coser': 25, 'other': 25}: 350 sessions, 13282 model requests (messages, temperature, max_tokens), the records, and the 4 judge prompts and BLEU strings built from each transcript; 0 sessions differ (speaker messages per session 6-20)
PASS V07a fix_repeation: 3905 vectors, 0 differ
PASS V07b tokenize_words (regex module): 3905 vectors, 0 differ
PASS V07c remove_inner_thoughts (624efc5): 3905 vectors, 0 differ
PASS V07d add_speaker_name: 1835 vectors, 0 differ
PASS V07e Agent.chat post-processing per family (own-utterance cut, fix_repeation, none for the NSP): 910 vectors, 0 differ
PASS V07f Python repr of name lists: 301 vectors, 0 differ
PASS V07g repr(float): 620 vectors, 0 differ
INFO V07 remove_inner_thoughts 624efc5 vs 93c1474 differ on 3/3905 inputs (the <…think> tags); the protocol uses 624efc5
PASS V08a judge prompts: 200 samples × 4 dimensions (system + user), R and the BLEU strings; 0 samples differ; literal {major_characters} kept in Character Fidelity: True
PASS V08b judge reading: extract_json (+ fix call), parse_response, severity types, score ({'ok': 28, 'retry': 18, 'missing-dimension': 1, 'crash': 0}; fix calls 5): 47 vectors, 0 differ
PASS V10a dimension score from constructed flaws (int/float/str/bool severities, floor at 0, no ceiling: 51 vectors above 100; the stored ceiling variant): 314 vectors, 0 differ
PASS V10 published exp/evaluation/test_set_gpt-4o.json: 12/12 dimension scores recomputed exactly (Python line 12/12), 3/3 case scores
PASS V11 run statistics: per-run means, their sample standard deviation (ddof=1), per-split means: 4 vectors, 0 differ
PASS V12a published BLEU: platform 3/3, Python reference 3/3 equal to the stored values (0.021624838733610124, 0.023715767261473588, 0.030483934319011176)
PASS V12b nltk 3.9.1 word_tokenize (Punkt + NLTKWordTokenizer) on book text, dialogues and edge cases (lower-cased and as written): 1984 vectors, 0 differ
PASS V12c nltk 3.9.1 sentence_bleu(word_tokenize(ref.lower()), word_tokenize(sim.lower())) with correctly rounded log/exp, bit-exact (equal to this CPython's libm result: 1000/1000): 1000 vectors, 0 differ
PASS V12-log math.log: platform equals the correctly rounded value (decimal, 80 digits); this CPython's libm differs from it on 4: 5269 vectors, 0 differ
PASS V12-exp math.exp: platform equals the correctly rounded value (decimal, 80 digits); this CPython's libm differs from it on 3: 3007 vectors, 0 differ
PASS V12-fsum math.fsum: platform port of CPython's msum: 502 vectors, 0 differ
PASS P01 Python character classes on the 282230 code points assigned in this CPython's Unicode 14.0.0 (platform JavaScript Unicode 16.0; known version differences ['0x10fc', '0xa7f2', '0xa7f3', '0xa7f4', '0xab69'] excluded): mismatches re_s 0, re_w 0, re_d 0, regex_w 0, regex_ps 0, isupper 0, islower 0, isprintable 0
31/31 checks passed
exit 0