Fusion receipts — "selective fusion with receipts"
Cette page n'est pas encore traduite : vous lisez l'original en anglais.
Chimera's reasoning core blends a panel of models (panel → judge → synthesizer). Fusion buys quality but costs more tokens, so the honest question is never "is fusion good?" but "was it worth it, here?". Receipts answer that with numbers instead of a claim.
Every fusion run can be priced into a receipt: what each advisor (panel member), the judge, and the synthesizer cost — each at its own model's rate — plus whether selective mode short-circuited the panel. Persist the receipts and you get a publishable cost × quality curve.
Try it
# Show the itemized per-advisor cost of one run:
chimera fuse "Explain CAP theorem simply" --show-cost
# Append each run's receipt to a JSONL, then summarize the curve:
chimera fuse "..." --receipt runs.jsonl
chimera fuse "..." --receipt runs.jsonl --selective
chimera fusion-receipts runs.jsonl
fusion-receipts reports the fusion rate (how often the full panel actually ran vs. a selective
short-circuit), the mean/total cost over the runs that had a known price, and — when receipts carry
a pass/fail quality signal — the pass rate and the dollars per passing answer.
Honesty rules (by construction)
- Tokens are measured; dollars are estimated. Token counts come from the provider; the dollar figure is computed at approximate public list price, so a receipt is an estimator, not a bill.
- Unknown model → unknown cost, never zero. If any stage runs a model with no price on file, the
receipt's total is
None(unknown), so a missing price can't masquerade as "free". Prices are overridable in code (chimera.fusion.set_price). - Per-advisor attribution. The panel cost is broken out per model (
receipt.advisor_costs), so you can see which advisor earned its keep — the substance behind selective fusion, not a slogan.
Why this exists
The field moved toward routing/cascades (spend more only when the stakes justify it), and away from always-on fusion. Receipts are what let Chimera fuse selectively and prove it paid off — the cost×quality curve is the evidence, published including the runs where fusion did not help.