local leCore BM25 retrieval + direct provider calls · corpus (plan + 144 Python files, 396,489 tok) retrieved per-ask instead of re-sent · FINAL RESULTS (frozen) — run completed ·
| model | state | runtime | with leCore | without (full-context) | save× | DHH paid | files | lines | %of port | effects | $/1M shipped | $/1M corpus-read | last | proj. beat/loss vs DHH ($) | proj. beat/loss vs DHH (time) |
|---|
solid = actual cost WITH leCore retrieval (only relevant slices reach the model) · dashed = the same asks WITHOUT it — the full 396k-token reference carried in every prompt, priced from measured provider bills · bar: DHH's published port is 66 files / 21,565 lines / 37 effects — Fable $550, Grok 4.6 $55, GPT-Sol $43, DeepSeek-V4-Max $23 · grok routes direct to xAI · answer key quarantined, models see only plan.md + the Python source · click any model name to read its generated code live · during-mortem: every dollar of self-inflicted cost, and why the numbers are conservative — plus per-ask ledgers (raw receipts)