Summary / Harness arms
A3 hg (gate bundle)
h0 plus every gate at once: tool and bash allowlists, protected test paths, edit-before-read, lint feedback, the verbatim-lift rule, budgets of 150 turns and 800K tokens.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
What is built now
Dev (09-08): 114/150 against h0's 115, Δ −0.7, CI [−8.0, +6.7], regressions unchanged, +$0.001 per task. Regression set: 23/30, one harness reject, and it was a true positive.
What is missing
Measurement of each gate alone (h2 was smoke-tested, not run), and a re-run on the current model.
Reviewer Claude, 2026-10-04
Reviewer's read: Gates cost nothing measurable
Bundling was the lean path agreed on 09-07: if the whole bundle is neutral, no gate needs its own run to justify safety. That is a safety argument, not a capability one. The gates prevent damage, and the bench confirms they do not cost resolves.
How it is compared measured (old model)
Bundle against h0 on dev; regression set as the guard.
Adoption rules shared by every arm: How the tests work.
Research & attempts
No documents yet.
Score history
Dashed line: the suite's baseline. Shaded band: the ± interval. Hollow dots are anchor items carried over from an earlier suite. Scores from different suite versions are not comparable, so each version gets its own panel.
| Date | Suite | Type | Score | ± | Live use | Cost | Notes |
|---|---|---|---|---|---|---|---|
| 2026-09-08 | v1 | sealed | 76 | — | — | — | hg on dev, k=3: 114/150. vs h0 Δ −0.7, CI [−8.0, +6.7]. Regression set separately: 23/30, the one harness reject a true positive. |
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.