Probation for the machines.

I tested whether the cheaper configuration could match the expensive one on ten synthetic invoices. The bill was $0.13986 in estimated API usage. The surprise was the punctuation.

Claude Fable 5 returned ten accepted results. Claude Sonnet 4.6 returned ten answers wrapped in Markdown code fences. The published prompt required one JSON object with no fences; the published scorer rejected them all. Six backticks failed the job interview.

That is a strict format-compliance result. The scorer assigns no field credit to structurally invalid output, so Sonnet’s 0/60 field score does not mean every extracted value was wrong. The raw answers are below. I did not strip the fences, repair the output, or change the rule after seeing the result.

Sonnet’s ten calls cost an estimated $0.02556; Fable’s cost $0.1143. Fable delivered the only accepted results in this run, at $0.01143 in API cost each. Sonnet’s cost per accepted result is undefined because none passed. There is no meaningful cost-per-acceptance ratio between them.

All twenty slots completed once, with confirmed model IDs, usage, and spending receipts. Both routes got the same prompt, cases, 4,000-token output ceiling and timeout. Order alternated; no retries, fallback or repair were used. The original protocol and hashes remain unchanged.

Ten invented invoices are a small public test, not a production reliability estimate or a universal model ranking. Internal model defaults differ. Human review time and cost remain unmeasured. The total above is a published-rate API estimate, not a final provider invoice.

The wrapper got a separate hearing. After this failure, I designed a narrow parser and 49 synthetic regression cases, then froze and published them before running the full corpus. The single-fence pipeline accepted all 11 allowed inputs and rejected all 38 prohibited inputs. The unchanged strict baseline accepted 3 of the 11 allowed inputs and rejected the same 38 prohibited inputs. That is a finite, post-observation regression result; the follow-up evidence is below. These original model scores stand. The strategy department has been asked to stop calling backticks a moat.

```json
{
  "supplier": "North Ledger Studio",
  "invoice_number": "INV-001",
  "invoice_date": "2026-09-01",
  "due_date": "2026-09-30",
  "currency": "USD",
  "invoice_total": "120.00",
  "review_fields": []
}
```

{"supplier": "North Ledger Studio", "invoice_number": "INV-001", "invoice_date": "2026-09-01", "due_date": "2026-09-30", "currency": "USD", "invoice_total": "120.00", "review_fields": []}