G0 authoritative run — measured cost basis for the PI go/no-go (2026-10-07)

Purpose. The frozen protocol (vignettes/articles/devlog-2026-09-28-G0-pi-adjudication.qmd §PI-5) allows the formal G0 verdict to originate only from a docker adapter + cosmic-file provider execution of the primary (or, per §6/§7-R7, the degraded) profile. This memo records what the harness now actually costs per unit of that work, measured end-to-end on the 2026-10-07 smoke run, so the PI gate is decided on numbers rather than on guesses. No protocol constants are changed here.

1. Pipeline status: connectivity proven

The 14:01 smoke execution (results/docker_smoke/, profile=smoke, adapter=docker, provider=cosmic-file, F2-frozen sigfit 2.2.0 + signature.tools STL) closed the last three container-path defects:

defect	class	guard now in CI
u64 seeds overflow as.integer	tool-boundary seed fold mod 2^31	test_seeds golden pins
user catalog never mounted	docker cmd assembly (both mounts)	test_seeds mount asserts
relative --cachedir → named-volume error	Path.resolve() at assembly + CLI boundary	test_seeds.test_docker_cmd_mount_sources_are_absolute

Result: 8/8 cells ok, zero FATAL, intervals.csv consumed downstream, fresh judgment_smoke_20261007-140100.md (kill conditions met: False). Per its own footer this file remains a harness self-check, not a G0 verdict — that status is unchanged and correct.

2. Measured unit rates (this machine, cold containers, N ≤ 1000)

tool	per rep-run wall time	dominant cost
sigfit 2.2.0	≈ 19–23 s	container start + Stan model init
STL (signature.tools)	≈ 112–122 s	container start + R namespace load

High-N calibration (§4, run same day) revises the naive reading. At N = 300 000 on fully loaded counts (sum = 300 000, all 96 channels non-zero, verified estimates up to ~8 × 10^4): sigfit 9 s, STL 38 s per rep-run warm — faster than the small-N cold figures. The smoke per-rep rates are therefore dominated by cold-start tax (container/daemon/library init amortised over 1 rep), not by per-N compute: neither tool shows a super-linear N wall in the measured range. The 1e6 tier is an eight-fold extrapolation of a measured slope and remains cheap.

3. Profile extrapolation (from g0/grid.py frozen constants)

profile	cell-reps/tool	both tools	serial, cold-rate floor	serial, warm-rate floor	×8 parallel warm
primary (200 reps, full 19-point N grid)	20 200	40 400	≈ 791 h	≈ 263 h	≈ 33 h
degraded (§6: 100 reps, 3-point N grid)	2 100	4 200	≈ 83 h	≈ 27.5 h	≈ 3.5 h

Warm rates (§2 calibration) are decision-grade for the 3e5 tier and extrapolated (8×) to 1e6; cold rates bracket the worst case with per-cell container restarts. The degraded profile is a same-day overnight run under any bracket; primary is a weekend-scale campaign, no longer a multi-week blocker.

4. High-N calibration — done

Two-cell execution (sigfit + STL at N = 300 000, cache/calib/, 2026-10-07, immediately after the smoke run) replaced the floor-rate extrapolation with measured warm values; results in §2. The remaining unmeasured tier (1e6) enters the §3 bounds only through the linear warm slope.

5. Decision menu for the PI

1. Run degraded now (§6 is protocol-predefined precisely for a compute-budget-constrained execution; verdict power ≥ 0.95 retained): same-day/overnight window under the measured warm rates, produces the authoritative Framing A/B input at bounded cost.
2. Run primary — now a weekend-scale campaign (§3), feasible on this machine without the M7 pinned-hardware host; the pinned host remains the right venue for the paper’s reproducibility exhibit.
3. Deferred — not supported by the ROADMAP: the G0 gate blocks Framing A vs B, hence the paper’s quantitative spine.

Recommendation: option 1 (degraded) immediately; option 2 opportunistically in parallel on spare cores once the degraded verdict lands.