THE FOLD / RESPAWN / ROLLBACK / THE FIRST MUTATE
THE FIRST MUTATE
the last instant a rollback was still free
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Most measures of whether an agent worked carefully are advisory: read the transcript, form an impression. David’s root0-i13n replaces that with one metric, and it is answerable: did anything that could falsify the answer fire before the first action that mutates? Every tool call is classified verify, mutate or neutral, and the first mutate freezes the answer for the session. Nothing said afterwards can improve the score. Crucially, unknown tools count as neutral — never as a control — because, as the pack puts it, a metric that flatters itself is worthless.
LIT verified live: the metric is ABSORBING — over 4000 traces ending in a mutate, appending any number of later verifies never changes the verdict; counting unknowns as neutral is strictly CONSERVATIVE across 6,000 random traces, never scoring higher than counting them as controls and strictly lower on some; the flattering variant inflates the pass rate from 2701 to 3567 of 6,000, a swing of 14.4 points; and the verdict depends only on the prefix up to and including the first mutate.
LIT verified live: the metric is ABSORBING — over 4000 traces ending in a mutate, appending any number of later verifies never changes the verdict; counting unknowns as neutral is strictly CONSERVATIVE across 6,000 random traces, never scoring higher than counting them as controls and strictly lower on some; the flattering variant inflates the pass rate from 2701 to 3567 of 6,000, a swing of 14.4 points; and the verdict depends only on the prefix up to and including the first mutate.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) built the Hermes super pack — 4 tools, 6 hooks, 10 bundled skills, 42/42 selftest checks against a fake context — and made two decisions that carry the whole design. First, unknown counts as neutral. Second, the installer deliberately refuses to arm the veto or the A/B split, because “they change whether commands get refused, and that is not an installer’s decision.” Seated at ROLLBACK, because the first mutate is precisely the last instant a rollback was still free.
AVAN (AI) measured the flattering variant rather than asserting it was worse, and the size of the gap is the interesting part: 14.4 points of pass rate, on traces where nothing about the actual behaviour changed. That is the entire distance between a metric and a decoration. Worth being exact about scope: what is verified here is the arithmetic of the metric — absorption, conservatism, prefix-dependence — on synthetic traces. Whether the classification of any real tool is correct is a separate question this page does not touch, and it is the harder one, because a tool misfiled as a control would corrupt the measure without changing any of these properties.
AVAN (AI) measured the flattering variant rather than asserting it was worse, and the size of the gap is the interesting part: 14.4 points of pass rate, on traces where nothing about the actual behaviour changed. That is the entire distance between a metric and a decoration. Worth being exact about scope: what is verified here is the arithmetic of the metric — absorption, conservatism, prefix-dependence — on synthetic traces. Whether the classification of any real tool is correct is a separate question this page does not touch, and it is the harder one, because a tool misfiled as a control would corrupt the measure without changing any of these properties.
3 ONE DIMENSION
One session on one axis. Everything after the first mutate is commentary.
4 TWO DIMENSIONS · INTERACTIVE
Build a session and watch the moment the answer freezes.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: many sessions, each frozen at its own first mutate.
AVAN’s addition (the inverse-companion): the forward reading is “check before you change.” The inverse is that the metric is really about when evidence stops being cheap. Before the first mutation, a falsifying test costs nothing but time; after it, the thing you would have tested no longer exists in the state you would have tested it in. So the ordering is not a discipline imposed on the work — it is the shape of the work’s own economics. Read backwards, “look before you touch” is not a moral instruction at all; it is the observation that touching destroys the cheapest evidence you will ever have.
LIT the metric is ABSORBING — over 4000 traces ending in a mutate, appending any number of later verifies never changes the verdict; counting unknowns as neutral is strictly CONSERVATIVE across 6,000 random traces, never scoring higher than counting them as controls and strictly lower on some; the flattering variant inflates the pass rate from 2701 to 3567 of 6,000, a swing of 14.4 points; and the verdict depends only on the prefix up to and including the first mutate
FIG What is verified here is the ARITHMETIC of the metric — absorption, conservatism, prefix-dependence — on synthetic traces. Whether the classification of any real tool is correct is a separate and harder question this page does not touch: a tool misfiled as a control would corrupt the measure without changing any of these properties. The 14.4-point inflation was measured rather than asserted; it is the distance between a metric and a decoration.
FIG What is verified here is the ARITHMETIC of the metric — absorption, conservatism, prefix-dependence — on synthetic traces. Whether the classification of any real tool is correct is a separate and harder question this page does not touch: a tool misfiled as a control would corrupt the measure without changing any of these properties. The 14.4-point inflation was measured rather than asserted; it is the distance between a metric and a decoration.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of ROLLBACK · David Lee Wise (ROOT0), with AVAN