Decision and routing service with cited evidence and abstention; held-out routing 30/32.

Oracle7-Trafficcop 2.0 (internal)

Status: public release. Prometheus7 Research Institute, 2026-09-25. Source SHA-256 (the trafficcop/*.py files in C-locale sorted order, concatenated; frozen copy at box four /opt/oracle7-trafficcop-2.0): 4a17af931863c351619efce513f94b4e099067a7d192a37c2ef371e60ce1ce4d. The results beside this card were re-run under exactly this SHA on 2026-09-25.

What it is

A decision service: given a state and a set of options, it returns one choice with a receipt. It contains no neural network. It works in this order:

1. Admissibility preconditions strike options that cannot apply.
2. Obligations bind what must be done.
3. An evidence table of attested outcomes ranks what remains.
4. A short microfold lookahead breaks ties.
5. When nothing is warranted, it abstains to collect_evidence.

It replaces Laya, an LLM decision service, behind the same DecisionRequest contract. Standard library only; about 0.1 ms per decision.

Results

Routing (Chameleon repair and MetaFold decisions)

	Dev (16 = 8 cases × 2 option orders)	Held-out (32)	Order-consistent
trafficcop	16/16	30/32	dev 8/8, held-out 16/16
always first	4/16	8/32	0/16
always last	4/16	8/32	0/16
random	3/16	7/32	3/16
always abstain	4/16	8/32	16/16
label-blind	4/16	8/32	16/16
Laya (recorded)	3/16	—	—

• The 30/32 is from the first run on the held-out cases, which were never tuned.
• Both held-out misses are abstentions to evidence (case h-behavior-2, in both orders), not wrong commitments.
• Structured MetaFold states: 2/2. That includes the case with no eligible operator, where Laya chose an operator whose precondition failed.
• Latency: 0.085 ms per decision; Laya took 7.7–16 s.
• Every mutant scores below the real model on held-out, so the instrument is valid.

Arcade (frozen RetroV1 decision panel, plus live episodes)

	Frozen panel	20 episodes: mean score (best)	ms per decision
trafficcop	24/24	41.75 (48)	0.445
greedy, no lookahead	—	22.15 (37)	0.280
random	—	0.15 (1)	0.178
Laya (recorded)	best 16/24	development episodes 0 / 0 / 1	—

Clone-capture triage (Chameleon)

The development card is tuned on 64 human-scored captures:

• Approval-candidate precision is 72%.
• The failure mechanism agrees with the human label on 8/28.
• All 156 failed captures are bucketed.

Because 64 of the 66 human labels that existed were used to tune it, this is a development number, not a claim.

Independent panel: PENDING. 60 catalogue sites were sampled at random (seed 20260925) from the 1,429 captures in needs_review that had never been human-scored. trafficcop's predictions were written and sealed before any label existed:

• Sealed file SHA-256: d9367a4dee474a4ded1b633746852cec0ce5e4a4b502f1b8dbaf36d109dc487e.
• The file is root-owned with mode 0400. The labelling service runs as a separate user and cannot read it.
• Predicted split: 28 candidate_for_approval, 10 review_visual_mismatch, 13 repair_missing_assets, 4 repair_readiness, 5 collect_full_evidence.

Ben labels the 60 sites blind (0–4). score_panel.py refuses to run if the sealed hash has changed. It reports:

• approval precision and recall (approve = human score 4),
• mean human score per predicted class,
• the approve-everything baseline.

This section is filled in from that run and is not estimated in advance.

Failures, stated plainly

• The routing held-out cases (32) were written by the builder before the first run. They are held out, but they are not independent. The sealed 60-site panel exists to close that gap.
• The capture-triage precision (72%) is on its own tuning data.
• Coverage is the decision families that were built: Chameleon repair routing, MetaFold operator choice, RetroV1 arcade, topic routing for the Expert Ring. Nothing is claimed outside them.

Reproduce (box four)

cd /opt/oracle7-trafficcop-2.0
python3 eval_routing.py
python3 eval_arcade.py
# after labels: /opt/trafficcop-panel/venv/bin/python /opt/trafficcop-panel/app/score_panel.py sealed/predictions.json labels/labels.json <out>