| Workpaper | Status | Agent findings | Updated |
|---|
The P&L is the source of record for Schedule C expenses — the 41 receipts corroborate it and are not counted again. One unrecognized scan was quarantined to Other rather than guessed into an expense.
| Ingestion & document vision | claude-haiku-4-5 One call per document — a bad scan never sinks the batch. Unknown types quarantine to Other. |
| Planning, drafting, optimization | claude-sonnet-4-6 Workflow judgment: archetype routing, adjusting entries, worksheet assembly, reviewer notes. |
| QA review | claude-opus-5 Second-pass scrutiny before anything reaches the queue. Falls back to sonnet-4-6 if a model id retires. |
| All tax computation | deterministic calc engine Every figure on a return: $0, 0 tokens, exact by construction. TY2025 SS wage cap $176,100. No model ever does arithmetic. |
Shipped-error gate for the seed stage: ≤ 0.5%. The engine holds the math at zero; the gates hold the judgment to a person. See for the running rate.
Paste an Anthropic API key over the placeholder at the top of index.html and redeploy. Every model step in a workflow run becomes a real call — haiku-4-5 parses, sonnet-4-6 drafts, opus-5 reviews — with actual latency, token counts, and cost on each step, and the model's own workpaper note set beneath it. The calc engine and the gates are unchanged: math stays deterministic, and nothing clears a deviation but you.
The key ships in page source, so use one with a monthly spend cap. A full live run costs roughly $0.02–0.06. If a model id ever retires, calls fall back to sonnet-4-6 and the step says so.
| Document | Engagement | Classification | Fingerprint | Extracted |
|---|
| Engagement | Returns | Status | Deviations resolved |
|---|---|---|---|
| SHStarlight Holdings | 3 of 4 drafted | In progress | 11 of 12 |
| CHChen Household | 1 of 1 drafted | Ready for review | 14 of 17 |
| DFDelphi Foods | 2 of 3 drafted | In progress | 6 of 6 |
| MPMeridian Partners | 0 of 2 drafted | Queued | — |