Costs

Every cost figure in the record, including the budget's three successive ceilings and the $175 ledger gap that a caveat failed to travel with.

Costs

Two rules govern this page, both of them the project's own. First, from qwen35/HANDOVER.md: do not quote a total from prose — run check_plan.py, which refuses to print the bad sum. Second, from qwen35/plan.json#actual_provenance_rule: "null means NOT MEASURED, never zero", and "if you cannot state which read produced a number, you do not have the number." Every figure below carries its file.

The budget had three ceilings, in sequence

plan.json#constraints_note insists the two buffers do not add: contingency is already inside total, and unallocated (127) is the only room above it. It also separates the approval ceiling (the budget) from the OpenRouter pot, a real shared account balance that this file is forbidden to carry, because the experiment "can stop for want of GRANT while still being inside BUDGET".

Pre-zoo spend

No total is given for the gradient probe, deliberately. gradprobe/cost_ledger.jsonl carries 23 rows, of which five are marked pre-ledger; from run stdout (facts $0.0107 and $0.0013, docs $0.0356 and $0.0003, verify_final $0.0076) and the rest are metered rows written by the scripts. The two sets overlap on the same tags, so adding every row would double-count. No file states the probe's total, so this page does not either.

| teacher screen, generation | $0.119453 over 2,880 rows | teacherscreen/results/gen_cost.json | | teacher screen, judging | $8.820759 over 5,760 judgements | teacherscreen/results/judge_cost.json | | sweep100 pair generation | $1.82, 26,523 calls, 23m15s, zero errors | harness memory, "Sweep execution log (2026-08-15 00:00-00:05)" | | sweep100 steering run | GPU $1.23 (2113 s A100-40GB at $2.1/h), judge $3.365 (2,830 calls), total ~$4.60 | sweep100/results/steer.md, "Cost and wall time" | | text-to-LoRA hypernetwork runs | $1.12 (1917 s) and $0.91 (1557 s) | harness memory, 2026-08-16 entries | | expanded 148-trait sweep training | $2.77 | harness memory, 2026-08-17 | | hole-word constitutions | $0.013221, 3 calls to anthropic/claude-sonnet-4.6 | qwen35/constitutions_cost.json |

The ledger correction of 2026-08-23

This is the central cost episode and it is recorded twice — in qwen35/plan.json#ledger_defect_note and in qwen35/HANDOVER.md's Money section, whose first line is a supersession banner.

Superseded: "$195.90 actual against a $1000 ceiling". That is the recorded sum of per-phase actual values.

Current: measured realised spend was ~$370.65 (Modal ~$308.83 plus OpenRouter $61.82) at the 2026-08-23 00:05 read. Remaining against the $1,000 ceiling was therefore ~$630, not ~$804.

Three mechanisms, kept separate because they need different fixes:

1. Billed is not incurred. A phase closed on a live billing read captures only completed intervals. bin/modal-phase-cost's own docstring was written about exactly this, using phase 2's $7.60 as its example — and that $7.60 was still sitting in plan.json. The tool written to stop the error had already produced the number nobody re-read.
2. App-name collision. train_qwen35.py defaulted PC_APP_NAME to pc-qwen35-phase2, so any run not overriding it billed under phase 2. There is no pc-qwen35-phase5 app: the 134-run main sweep is inside phase 2's $69.33. Phases 2 and 5 together recorded $62.85 against $69.3254 measured — 10% out, not the 9.1x that phase 2 alone appears to be (plan.json#joint_billing_groups). This is why per-phase splits must not be quoted.
3. A phase costs its failures. Phase 3 recorded $111.04 against $206.885 measured — 1.86x. The logs say why: phase3_{shuffled,permuted,seedpaired} each have a .FAILED-datarace and/or .RSLORA-INVALID predecessor in qwen35/. What was recorded is what the successful attempt cost. See Lesson: a phase costs its failures.

Two gaps, not one — always say which. At the 00:05 read: recorded Modal actuals $173.89 against measured Modal $308.83 = Modal gap $134.94; recorded all pots $195.90 against realised all pots $370.65 = total gap $174.75. Phase 3's $95.80 shortfall is 71.0% of the Modal gap and 54.8% of the total gap. Both $134 and $175 are right; naming the denominator is what stops a later reader concluding one of them is $40 out. The Modal gap also rises while a phase runs — $137.52 by the 00:12 read — so quote a gap with its read time.

plan.json#modal_measured records total_pc_qwen35: 311.4096, read_at 2026-08-23T00:12:00+00:00, window 2026-08-01 to 2026-08-24, FINAL: false because phase 10 was still running. Its own annotations: the total is solid because it depends on no attribution being correct; the per-phase splits are unreliable; and while phase 10 runs the total is a floor.

How it was reported matters. Commit f2e02b4 (2026-08-23 00:40) is titled "HANDOVER: delete the bare $195.90 rather than annotate it — a caveat does not travel with a number". The bare figure had been sitting directly under a banner saying it was wrong by $175, and HANDOVER now says so explicitly: "a caveat does not travel with a number, only the number's own text does. Same reason check_plan.py refuses to print the sum."

The per-phase ledger

From qwen35/plan.json#phases. actual: null means NOT MEASURED.

Phase 7 (steering) carries a realised block: OpenRouter $40.0529835 (from results/steer134_judged.json's cost block, independently summed from steer134_judge_cache.jsonl, 15,944 lines, agreeing to the last digit) plus Modal $26.1433 (app pc-qwen35-phase7-steer134), for $66.196 against a $35 nominal — 1.89x the whole phase. The earlier framing "over 100% on the OpenRouter side" understated it: the unmetered Modal half was 39% of the true cost. The judging figure was later corrected upward again to $44.33 (16.28M tokens, judge openai/gpt-5.6-terra, rubric v3-steer134) once the file was read.

Phase 10 (OCT stage 2) carries the sharpest entry. Its realised block gives modal_failed_attempts: 27.9 and modal_six_app_total: 37.93 at read_at 2026-08-23T05:16:00+00:00, with bought_nothing: "~$28 of the ~$38 spent to date — 74% — bought no result." Version 1 of the run (ap-apV1QCxfYlgrNhCVJ9kn6U, launched 2026-08-22 23:42) died at about 04:25 with a FunctionTimeoutError at 10800 s having billed $27.90 and produced nothing; oct_stage2.py:749 records that the same three-hour cap had already killed two earlier full-scale runs the same day. This is git commit 9d5785e: "phase 10: v1 died at a 3h function timeout having billed $27.90 — 'a phase costs its failures', live, four hours after I wrote it up as history".

The phase then settled: $65.27 across 6 apps, FINAL: true, in_flight: [] — the first non-lower-bound figure the phase had. The corpse stays a fixed $27.90, so productive spend is $37.37 and the corpse share closes at 43%, down from 74%. Two warnings attached to that number: the gen and merge stages were skipped and served from cache, so $65.27 is the cost of training and releasing three traits, not of taking three end to end; and three prior projections given to Samuel ($95-285, ~$330, $75-95) were all wrong because they extrapolated a container-startup burst of about $21/h against a marginal rate of about $5/h.

The ledger's own conclusion about phase 10, from plan.json (janitor's correction at 05:16): the failure cost was not invisible — all six instances bill under one app name, which is exactly why the tool reads $37.93 and not ~$10. The blindness is in a per-instance read and in the ledger. The rule is therefore quote the tool, never an instance — see Lesson: quote the tool, never an instance and git commit cab9aff.

The unresolved charge

App oct-continue-persona-curvature, $8.7978, 2026-08-21/22, is excluded from the Modal total and remains measured but not placed (plan.json#modal_measured.unresolved). It is seven app ids sharing one name, not one app: six did real work back to back on 08-21 between 13:00 and 17:00 UTC (IxOhD1 $3.1880, H7pE2M $1.3856, 5hBdB6 $1.0805, JZ5PAR $1.3905, 8g94qk $1.3767, Jhj15P $0.3529), and the seventh, NqZUko, is worth $0.0236 and its whole log is a container-start crash. Reading the logs alone would have produced the false and quotable finding "a failed run billed $8.80". An app name is not an app — group by object_id before reading any log.

Its entrypoint module oct_continue_modal is not in the tree, so it was launched from this box from a file since deleted; 08-21 was the harness blackout (631 turns attempted, all failed at zero cost), so no agent could have launched it, and six sequential launches over four hours is a person iterating. The retained logs cannot settle which model arm it belongs to and it is not inferred.

Cost models the project derived

• $10.17 per app plus $0.333 per run, not a flat per-run rate; verified out-of-sample to 0.8% on the 134-run sweep (qwen35/HANDOVER.md). "A flat rate is only ever correct at one job size."
• 4.77 GPU-hours per trait for full-scale introspection SFT, the three-trait average — correcting a 4.20 figure quoted from warm alone while it was the only finished run. organized ran about 1.5 h longer than warm on nearly the same step count.
• $2-3k order of magnitude to take all 134 traits through OCT end to end, and the verdict was "no for the sweep".
• $2.10 per GPU-hour for an A100-40GB, the rate zoo40_meter.sh integrates against.

The running total

Stage-2 cost is about $15 per trait at the planning rate (SFT alone averages 3.2 hours) — this figure is maintainer-reported, and the nearest written source is .garden/journal/2026-09-04.md, which records the same number as a correction to a $20-25 quote given to Samuel, and prices 10 traits at $151.

Two rules earned, at cost

• A carried ask is a live claim about the world. On 2026-08-23 an OpenRouter top-up request was withdrawn: the pot read $143.87 remaining of $958 granted, the grant had risen from $908 to $958, and the $99.36 in the pending batch was stale. Re-sending it would have asked Samuel for money he had already sent. Re-measure the ask, not just the answer.
• A mechanism that fails in the flattering direction fails silently, because the flattering direction is where scrutiny stops. Three instances in one day (2026-08-19): an estimate too small to question ("four dollars felt right and I did not check it, precisely because it was small enough to be reassuring"), a baseline that only ever moved up, and a projection whose optimistic form nobody would have interrogated. The $420 charge that was avoided was caught by a $0.63 smoke test, not by any ledger guard.

Related: Infrastructure, Phase 2 — the recipe search, The teacher screen, Method lessons, Timeline.