Datasets

Round 2 (dataset version 1.1, corrected), on 243 in-scope events (Feishu group after the round-1 cutoff, with the bot’s replies present, and the operator’s Claude Code sessions): f_v5 matched the wanted mode 25/243 exactly (0.1029, Wilson 95 % 0.0707 to 0.1475), 54/243 with M1 and M2 read as one (0.2222), 101/243 across the five classes (0.4156, 0.3555 to 0.4784); the bot itself matched 34/70 of the corrections where its mode could be read; on the 173 controls the operator let stand, f_v5 would have changed 108 across the five classes. Label reliability: on the operator’s own answers the models’ consensus matched 6 of 9 exactly (Wilson 95 % 0.3542 to 0.8794), 8 of 9 with M1 and M2 as one, 8 of 9 across the five classes; with the delegated answers included, 16 of 21 exactly, 20 of 21 with M1 and M2 as one, 20 of 21 across the five classes. Truth is the operator’s own answer on 9 in-scope events, a delegated answer on 12, the models’ merged label on the rest. Result case A. The canon is unchanged.

One hundred anchor scenarios for ten dimensions of the iLang v5.0 judgment vector: five levels per dimension, one Chinese and one English sentence each, drafted by Meta’s Muse agent from the canon’s definitions and reviewed by the authors (nine changes). The open set of Part II §5, published here because the canon’s v5.0 text is sealed.

Three wordings of the iLang canon, one model, one route, 320 cases each. The 26-statement rewrite that leaves decisions to code took execution passes from 10 to 25 of 100 and authority self-assignment from 86 to 68 cases (p = 0.006 and 0.002, paired); the merged canon, which adds one rule about the runtime’s budget line, took them to 68 of 100 and 2 cases. Grammar and judgment stayed within the noise of a single run; refusals and filters stayed at zero.

Every decision two production bots were about to make, judged by the iLang v5.0 reference function and published daily on Hugging Face, one line per event, without the conversation text: the eleven-dimension vector, its probabilities and confidences, the f_v5 mode, the deciding step, the model’s own direct five-class answer and whether it disagreed with the mode (an internal-consistency measure, not a second judge).

f_v5, computed from four models’ weighted reading of the eleven dimensions, matched what the operator wanted on 13 of 157 in-scope events exactly (0.0828, Wilson 95 % 0.049 to 0.1365) and on 60 of 157 across the five classes (0.3822); the bot itself matched on 7 of 30 corrections where its mode could be read; on the 112 controls the operator let stand, f_v5 would have changed the bot’s decision on 102 exactly and on 63 across the five classes (five-class agreement 0.4375, Wilson 95 % 0.3492 to 0.5299, against a baseline of 1 by construction). Result case A+C, CX-007 registered against f_v5_STEP-4_S_bands; truth is the operator’s confirmation on 17 events, the models’ merged label elsewhere. The weighted measurement of T2 did not win its first round: MAE 0.1027 against 0.1043 for a plain mean and 0.0989 for the best single model. The canon’s normative text is unchanged.

45 model runs against a 320-case deterministic conformance suite, 18–20 September 2026. Best weighted total 0.8417; no run reached the L1 gate. 93.5% of all execution-rule violations fall on a single rule: acting with authority the model does not hold. Every run’s per-track scores and per-rule failure counts are published.

Three days of a production agent’s messages judged twice — by a cheap always-on judge model and by the operator’s written rules. 7,940 messages judged, 2,648 scored against the reference. Agreement 68.4% → 75.1% → 80.9%; 87.2% in the judge’s top confidence band. Includes a worked correction: the rule-compliance gain everyone would have quoted, 69 → 7, is 13 → 7 once both days are measured with the same criterion.

Is the v5.0 judgment mapping — 11-dimension vector to one of eight decision modes — a learnable surface or an arbitrary table? A plain gradient-boosted tree recovers it at 0.9653 against a 0.3528 majority baseline, and its predictions pass the official JCS gate at 0.9861. 24,000 pairs, seed fixed, predictions published.

24-day longitudinal record of AI agent judgment calibration in production. 152 operator messages (41 explicit tuning instructions), 1,062 user messages, 2,162 API calls, plus a 14-day WeChat calibration log. Every figure traceable to a file.