Amendment A3 hard stop
$46.75 of $100
47% used · $53.25 of headroom
Updated 2026-10-03: plus the two attempts at the small rented test run of 2026-09-25 (about $0.50 and $0.0525), which the ledger charges here; the first has since been billed at $0.5162, now annotated in the ledger (2026-10-06). That leaves about $53.25 under the stop. The rest of this note is as written on 2026-09-22. All vendors, cumulative, and added up row by row from the compute ledger, which is the system of record: the learnability pilot that opened Amendment A3 ($13.92), the short test of whether a rented machine can delete itself ($0.29), seeds 1 and 2 run as one wave ($20.10), the run that asked whether the control question can be learned when it is taught properly (about $9.90), the recovery of that run's final checkpoint ($0.07), the two follow-up reading runs of 2026-09-20 — the matched control and the standardised refit — which were created on rented machines, refused by the instrument's own reproducibility check and deleted (a measured $1.90), and the measurement rehearsal of 2026-09-21 (about $0.02). Corrected 2026-09-22 from $44.3: that figure counted the two follow-up runs as costing nothing, but the account balance moved $79.7159 to $77.8119 while they ran. That leaves $53.80 under the stop. Option D would draw $27–39 of it if charged here.
MVM-0a registered design envelope (Amendment A2)
$229.62 of $450
51% used · $220.38 of headroom
Updated 2026-10-03: plus about $0.55 for the two attempts at the small rented test run of 2026-09-25, so headroom is about $221.85. The successor's two releases are re-priced at $161.90 from the measured training speed, which would put the programme at $422 to $434. The rest of this note is as written on 2026-09-25, and its headroom figures are against the old $400 ceiling. Raised from $400 to $450 on 2026-09-25 by page 6 of the Weekend 1 queue ruling (docs/rulings/2026-09-26-weekend-1-queue.md), to hold the successor's base plan plus one extension; the compute ledger carries a dated ceiling note saying so. Before that, raised from $200 to $400 on 2026-08-16 at measured 5090-secure pricing. Includes the lost 30M pilot ($97) and the 30M re-run ($18.9). Two corrections got it here. The figure was $215.7 until 2026-09-21, when two ledger rows turned out to have been written without a programme running total, which had left out the run that tested whether the control question can be learned (about $9.90) and the recovery of its checkpoint ($0.07). It then read $225.7 until 2026-09-22, when the two follow-up reading runs of 2026-09-20 — described in several places as costing nothing — turned out to have moved the account balance from $79.7159 to $77.8119, a measured $1.90 that the ledger's own correction of 2026-09-21 already carries. Headroom is therefore $172.40, not $174.30, and that gap is worth stating: the two spending releases John staged on 2026-09-21 come to about $175, which is about $2.60 more than the programme has left rather than the seventy cents over that was reported at the time.
Account balance
$72.32
Updated 2026-10-07, corrected later the same day from the handoff for the registration draft: the balance read $72.3178 after the four 10-million development runs of 2026-10-04 (the $73.7742 first quoted was the reading just before the last machine; the runs cost about $1.47 together, under their $10 line); the programme stands at about $229.62 of $450. The rest of this note is as written on 2026-10-04. RunPod prepaid credits, auto-reload off, never loaded past the $80 account limit John set; topped up in increments he chooses. The balance is the last one the compute ledger records being read from the account, after the second rented test run on 2026-09-25 ($75.8645). Before that it read $77.3451 on 2026-09-21, having fallen from $79.82 because the two follow-up runs of 2026-09-20 cost a measured $1.90 and the storage volume drips about a cent an hour whether or not anything is running. Anthropic and Gemini API judge calls are billed separately and cost cents per run.
$10.30 lost to idle billing so far. Four occurrences of finished pods billing idle; the reason every launch now carries a terminate window and a watchdog.