Governed memory · internal benchmark · B5

Two agents. Same hour. Incompatible decisions.

Nobody told either of them. Fifteen seconds after the second one decides, it finds out — and refuses to write the number into the runbook.

Preliminary · internal 5 positives · 10 negatives 64 agent turns no vanilla arm — it cannot play
Run 12 Jul 2026 Model Claude Opus 4.8, both sessions Detection CoordinationCase opened in control Delivery the next turn's own answer
Headline
5/5real conflicts caught
5/5reached the next turn
15smedian to detection
100%precision (0 false alarms in 10)

This is the one benchmark on this site with no baseline arm, because there is nothing to compare against. A vanilla agent cannot lose this test — it cannot take it. Detecting that another session, running right now in another terminal, just decided something incompatible requires a shared memory that both sessions write to and something that reads across them. An agent with a CLAUDE.md has neither.

So the question is not "does it beat vanilla". It is does it work, and does it cry wolf.

What actually has to happen

Four hops, two services, one queue — and the agent is never told mid-turn

Session A decides. Session B, running in parallel, decides something incompatible. Every hop below is asynchronous, and every one of them is silent when it fails. So the benchmark gates on each one: a pair whose chain does not land is scored invalid, never a miss. We refuse to blame the detector for a delivery failure upstream of it.

1Session A's turn closes → captured → finalized into control4s
2The turn is extracted into normalized claims4s
3Claims are indexed — now B's session can retrieve them12–36s
4B's turn closes → its claim is compared against A's → contradiction case opened10–15s
Positives — 5 pairs

It catches them, and it tells someone

5/5 detected · 5/5 delivered · 10–15s

Two sessions each decide a production rollback window — 43 minutes and 16 minutes. Both cannot hold. On its next turn, session B is asked a neutral question: what is the rollback window, and is there anything I should know? It was never told a conflict exists.

The rollback window was set to 16 minutes in this session, but that decision is currently under a coordination hold because it conflicts with an in-flight change in another session. Don't commit the 16-minute figure to the runbook yet. The agent, unprompted, on its next turn
Negatives — 11 pairs

The trap it still walks into

0 false alarms in 10 · after we fixed the cause

A detector that only ever sees real conflicts will score 100% and tell you nothing. So most of this benchmark is decisions that look like conflicts and are not.

Negative archetypeFalse alarms
Different setting entirelyrollback window (min) vs staging deploy timeout (sec)0/5
Unit restatement"18 minutes" vs "1080 seconds" — the same value0/3
Scope carve-outglobal default 31 min + "in the EU it is 43 min" — both true0/2
The finding that mattered

The false positive was never the detector's fault

We first published 83% precision. One negative pair opened a contradiction case for two decisions that are both true at once — and warned the next agent about it. We went looking for the flaw in the detector and did not find one.

The session had said "the rollback window is 31 minutes by default". What our own claim normalizer stored was "the rollback window is 31 minutes in all regions". It had deleted the word "default", turning a defaultable rule into a universal one — and "31 minutes in ALL regions" genuinely is refuted by "in the EU it is 43".

The detector was right about the text it was handed. The text had been falsified upstream, by us. Of three identically-shaped pairs in that run, the two whose normalization kept the word coexisted correctly, and the one that lost it raised a false alarm.

MetricPublishedAfter the fix
Recall5/55/5
Delivered to the next turn5/55/5
False alarms1/110/10
Precision83%100%

The fix is a deterministic guard in the claim validator: if a quote carries a defeasibility marker and its restatement carries none, the claim was strengthened and is rejected. We did not fix it with a prompt — the bug was non-deterministic and would not reproduce on demand, so a prompt change could never have been honestly verified. Recall did not move, which is the number we actually needed to check before claiming the fix was free.

We are leaving the 83% on the record. A benchmark that finds a real bug in the product, gets it fixed, and then measures the fix is worth more than one that scored well on the first run.

How to read this

Meetless · governed memory·run 12 Jul 2026·Claude Opus 4.8