docs: final numbers from the overnight sweep

AI-assisted (Claude).

Co-Authored-By: Claude <noreply@anthropic.com>

docs/COMP-FIDELITY.md

@@ -79,13 +79,20 @@ Same niche (07-vintage-moto-forum), same approved comp C, main skill vs this bra
 | exec cut, packet C, ask names the phased build | branch | 96 / $5.24 (cost cap) | **66%** (drift; plate region 43%) | walked spec → grid → regions → plates → hero gate (72% fail, fix, 77% pass) → sections → motion → responsive; 12 turns lost hunting a screenshot the harness wrote host-side only (fixed in impeccable-evals); the exploded plate was produced (62% vs crop) but the page drew the region in SVG anyway (fixed: hero gate now refuses unreferenced plates) |
 | full journey, comp-led | main | 49 / $3.23 | **56%** (contradicted; two bands `missing`) | dark comp with paper fiche rail; build keeps the section order and flattens the material |
 | full journey, comp-led | branch (before the force/reference fixes) | 61 / $3.04 (turn cap) | **65-66%** (drift) | forced past the plates gate with "single-file delivery" (now refused); hero 44% structure but 83% color, plate placed |
+| full journey, comp-led (after fixes), sample 2 | branch | 60 / $3.00 (30-min wall clock) | **59%** (contradicted; hero gate 61 → 63 → 62%) | three plates produced (paper 52%, carburetor 61%, photo), hero gate failed three times, model asked the simulated user, got "truthful translation", forced with the user's words (recorded), wall clock ended it in sections |
+| full journey, sample 1 | branch | 37 / $2.5 | n/a | direction round, then wrote code with no comp round at all: a routing gap in the direction round that predates this branch (also seen on main in cf-full-branch-07b) |
+| full journey, comp-led | main (2nd sample) | 24 / $1.38 | n/a (no comps: model went code-led) | same routing gap |
 | exec cut, 01-observability composed checkpoint | main / branch | 9 / $0.68 vs 10 / $0.84 | 52% vs 48% | composed checkpoint quotes the OLD visualize.md verbatim into the prefix, so the branch text never reaches the model; not a test of this change |
 
+Sweep totals: 12 runs, about $32 of OpenAI spend (gpt-5.6-sol + gpt-image-2).
+
 What it says so far:
 
 - When the phase machinery actually runs, fidelity moves from the mid-50s to the mid-60s on this comp, and the region rows say why the rest is missing (the exploded plate, the parts table, the CTA treatment). Same model, same comp.
 - Execution-cut packets and composed checkpoints carry the old skill's text in their prefix; a resumed session follows the conversation it is in, not the mounted files. Comparisons of Setup-adjacent skill changes need full journeys or a fresh packet cut on the new skill.
 - Two of the three run-time defects the sweep found were harness (screenshot not visible in the sandbox; packet workspace path) and are fixed in impeccable-evals `paul/packet-niche-execution-preflight`. The third (model forcing a gate, model ignoring its plate) is now refused by the script.
-- Cost: the phased build spends more turns before the first write and more image calls (plates). The 96-turn run is dominated by the screenshot hunt and a font-inlining tangent, not by the gates.
+- Cost: the phased build spends more turns before the first write and more image calls (plates). The 96-turn run is dominated by the screenshot hunt and a font-inlining tangent, not by the gates. The 30-minute wall clock and 60-turn cap in the harness are tuned for the old one-write build; a phased build with three plates and a hero loop needs the execution cut kind's 100-turn budget or a longer wall clock.
+- The hero gate at 72% is reachable (77% on the ask run) but the model's second and third attempts moved the score by one point each: it edits CSS values when the diff says a region is missing. The gate's message now names the region and the failure mode; the next lever is making the region crops the thing the model looks at (it opened the side-by-side once and the crops never).
+- Whether a greenfield session enters comp-led at all is decided in the direction round, before any of this. Two of five full-journey samples (one per skill) skipped the comp round and wrote code; the config default is comp-led and image generation was on. That routing gap is separate from this change and worth its own fix.
 
 Next: cut fresh packets on the branch skill for 02 and 07, sweep both lanes (openai + anthropic) at n=3, add one Operate niche (11-analytics-dashboard, plates near-empty) and one mobile niche (23-transit-mobile, portrait comp).