⚠ working notes — what we've learnt so far · updated 12 July 2026 difficulty–response curves

Finding 7 · The router: making the card earn a living

Everything on this site condenses to a practical question: given a request, which model should answer it? We turned the card into a router and replayed it against our own recorded data — hundreds of real attempts on shared puzzles, with real per-call costs.

First, the idea that died (reported per the house rules): using a single call's token count to judge whether that answer is wrong. Measured AUC ≈ 0.5 — a coin flip. The invoice locates frontiers in aggregate; it does not grade individual answers. Bring a verifier.

Then, the replay that pays — policies over the same instances, real costs (cheap rung: gpt-oss-20b; "premium": o4-mini):

policysolve rate$/solved
always cheap (gpt-oss-20b)0.65$0.00888
cheap, retry ×3 on verified failure0.776$0.0135
cascade: cheap → premium on failure0.692$0.0395
always premium (o4-mini)0.363$0.15475

Two rules fall out, both violations of routing folklore. Retries beat escalation: near-frontier failures are coin-flips (finding 1), so re-rolling the cheap model solves more, cheaper, than calling in the big one. Premium ≠ stronger: on this task family the $4.84/M model solved half as much as the $0.22/M one — the only ladder that means anything is measured frontier on your task, which is what the card is.

Live validation (fresh puzzles, real spend)

25 fresh instances, levels [3.0, 3.6, 4.2, 4.8, 5.3], via OpenRouter; total spend $5.963.

policysolved$/solved
always_gpt_oss_20b12/25$0.0074
always_o4_mini9/25$0.1434
formguide_router24/25$0.1854

Use it

The router ships in the repo under integrations/litellm: a LiteLLM-compatible FormGuideRouter (cheapest capable rung → mist retries → frontier escalation, per-rung token budgets, both max-token spellings), a machine-readable formguide.json generated from the measured card, and recalibrate.py for the drift problem (finding 2): frontiers moved a full dial unit in twelve hours in our data, so calibrations carry dates and a re-probe costs cents.

The card is a snapshot; the router is the reason to keep the snapshot fresh. One load-bearing caveat: this router had a free referee — SAT answers are checkable. Whether the trace can stand in for the verifier got its own round of experiments: finding 8. (Spoiler: bring the referee.)