TTFX GOLDRUN · DURING-MORTEM · WRITTEN MID-RUN, NOT AFTER

every dollar of self-inflicted cost_

This run's harness was improved while it ran, and improvements cost money. None of that waste is edited out: it is in the ledgers, the totals, and the projections you see on the dashboard. This page lists it, so the conservatism is checkable, not claimed.

cost-heavy changes pushed mid-run

1. The render gate (mid-run): "kept" was retightened from compiles to compiles AND emits ANSI-styled animation. Already-paid unstyled effects were dropped and re-earned at full price (ds lost 7 in one stroke). Every re-earn dollar stays in the totals.
2. Reasoning-budget truncation, three variants: deepseek (0 chars at 8k), grok (output pinned at exactly its 16k cap), sonnet-5 (Anthropic max_tokens covers thinking+text — files cut mid-fence). ~$7 of paid, unusable responses before each was diagnosed. All on the ledgers.
3. grok paid ~$1.90 asking a model to write a file the harness generates for free (effects/mod.rs). Harness bug, our cost, kept.
4. Duplicate-process era: a bad kill pattern briefly ran two copies of some models — double-bought cores. Kept.
5. Restart re-cores: fable bought its core 3x, sonnet 2x, across harness fixes. Kept.
6. Fat-chunk era: early asks shipped 22–33k prompt tokens (60-line chunks) before rechunking to the sidecar's ~700-char geometry. Those expensive asks are in the ledgers.

why the projections are conservative

• Projections use the blended rate — total spend (all waste above included) divided by kept effects, scaled to 37. Marginal rates are far lower: sonnet-5's last effects cost ~$0.12 against a blended ~$0.60. A clean run hits the projected number with a wide margin.
• The "without leCore" column is measured, and measured LOW: real cold+warm corpus sends, priced by provider bills — and the measurements found no cache discount for Anthropic-by-default or DeepSeek-via-OR (warm = 1.0x cold). We still credit the full-context arm with the same keep-rate we achieved, though several models cannot even fit the 396k corpus in their windows.
• DHH's side is his published successes only — $23–550 per model. His failed runs (DSV4 Flash, GPT Luna) cost real money too; we count none of that against him.
• All measurement overhead is ours: the $31.76 cold/warm corpus measurement and every probe came out of our account and none of it is billed against the comparison.
• The quality bar rose mid-run and totals were not reset. The strictest gate is applied retroactively to spending accrued under weaker gates.

receipts: per-model ledgers · run logs · generated source · account drain sampled from the provider every 60s (header of the dashboard). swap 'sol' for any model tag.