The tools on real systems: a multi-agent software system’s records and a quantum computer

Three tests outside simulation: fresh AI auditors read the working record of a multi-agent software system, its logs were audited, and an audit protocol ran on an IBM quantum processor.

Auditing a multi-agent system’s working record (H-0012)

Fresh AI auditors read k of the system’s 19 notes and answered questions about the work they describe. Accuracy rose linearly with k, from 0.09 to 0.86, and the notes plus the git log reached 0.98, with 0 fabricated answers on 128 trap questions. Accuracy is about 0.93 times presence (whether the answer was in what the auditor read): what limits legibility here is what was written, not how well it is read. Small n.observed
Logs beat memories, but only for readers with bounded effort: logs keep records from before the information was mixed away. In Shannon’s terms memories hold at least as much.

Auditing the same system’s logs (H-0017)

Around a hub agent, slicing the record by cause returned a median 0.62 of its 196 commits, while slicing by content returned 0.04, a part about 16 times smaller. Copies make corrections hard (ρ = 0.97); the audit found 14 stale statements in the record.

IBM run 1 (ibm_fez, 83 seconds of quantum-processor time)

Job db6nmvg4qg6s73c52r5g. The scrambler’s average fidelity was 0.897 against 0.966 predicted. Detected errors in the auditor’s model cost success rate (0.85, 0.50 and 0.19 against predicted 0.86, 0.50 and 0.15) while the answers stayed right. Undetected errors cost fidelity. Errors outside the light cone changed nothing (±0.01).
Corrected: the two-qubit rows cannot show that the circuit scrambles, because a SWAP gate passes them. The single-qubit rows (fidelity 0.81 and 0.84, both above 2/3) rule out a SWAP at more than 40σ, for Clifford scramblers under a model that trusts the device. Prediction P2 missed because of coherent errors in the identity circuit.
IBM round 2: pending, held back for a redesign before any data.pending
The audit-rate transition (H-0014): pending. The ring value p_c = 0.1598 is reproduced.pending

H-0012: share of questions fresh auditors answered correctly against the number of notes read, with the presence ceiling and the git-log and all-notes lines.

H-0012: logs against memories for a reader with bounded effort.

H-0017: how much causal and content slicing reduce the record.

IBM ibm_fez run 1: measured results against predictions.

Lab records: H-0007, H-0012, H-0016, H-0017; understanding/08 and 11.

Tags: confidence (proved, verified, observed, suggestive, pending) and novelty (known, narrow, ours), defined on the overview.