A team renames its core object. The owner rules on it. But the code, the tickets, and the standups keep using the old name — the rename never really lands. Now an agent has to finish a piece of work that uses that name. Reading the raw notes, it ships the dead name 16 times out of 17. Governed memory — same notes, plus the verdict — ships the one that's in force.
Both arms hold the same notes. Both can read the owner's rename. The difference is that governed memory also records which decision still stands — so it ignores the habit and ships the current name every time but one. The ungoverned agent reasons over the same contradiction and, ~94% of the time, concludes the superseded name is current, because that's what the latest note and the sheer volume of usage tell it. It is slower reaching the wrong answer than governed memory is reaching the right one.
There is no tidy "renamed X to Y" decision doc sitting in a decisions/ folder — that clean artifact is the thing governance produces. Real history is a mess: an old standard, a rename, and a later shrug that the rename never took. Governance is the layer that says which one is accepted.
The old name is prevalent throughout — ~59 of 60 notes use it, including the most recent ones. Frequency, recency, and the "reality check" all point at the superseded name. Only the governance verdict points at the one the owner ruled in force. The ungoverned arm has every one of these notes on disk and reads them; it is not missing information. It is missing the verdict.
This is why the ungoverned arm is a coin-flip and not a floor: it can get it right (1 of 17), it just usually doesn't. The whole outcome turns on which note it decides to trust — and nothing in the text tells it which one won.
"A rename to Transfer Log was decided in 2024, but a 2025 note says it never landed… almost everything still says Handoff Note… treating that as the working term." ungoverned · shipped the SUPERSEDED name · wrong (the common case, 16/17)
"2024-06-16 — Owner decision: the canonical name is now Transfer Log. No later note reverses it." ungoverned · trusted the owner's ruling · right (the rare case, 1/17)
"Checking governed memory before I finalize… 2024-06-16 — ACCEPTED: the canonical name is Transfer Log. The draft's 'Handoff Note' is the deprecated term. The naming evidence is decisive." mla · queried the verdict · right, every trial but one (16/17)
| rename | ungoverned | mla |
|---|---|---|
| Handoff Note → Transfer Log | 1/5 | 5/5 |
| Work Ticket → Task Card | 0/3 | 3/3 |
| Change Request → Delivery Slip | 0/3 | 3/3 |
| Case File → Record Card | 0/3 | 2/3 |
| Job Sheet → Run Card | 0/3 | 3/3 |
| total | 1/17 (6%) | 16/17 (94%) |
On every trial the health signal confirms the ablation: mla grounded (used the governed verdict), ungoverned unused (reasoned from raw notes). The one mla miss — Case File — we keep: even a retrieved verdict lost once to the contested context. ~94%, not a suspicious 100%.
| metric | mla | ungoverned |
|---|---|---|
| ships the current name | 16/17 ✓ | 1/17 ✗ |
| agent time (median) | ~110 s | ~142 s |
| used governed verdict | every trial | n/a |
| read raw notes to decide | no | every trial |
The pull arm doesn't fail fast. It investigates — reads the code, greps the notes, reconstructs the timeline — and then trusts the wrong signal. It pays more to be less reliable. Governed memory asks one question and stops.
# reproduce cd tools/bench-mla node run.mjs --tasks canonical-deprecation-use --arms vanilla,static,ungoverned,mla \ --corpus 60 --trials 5 --substrate neutral