# SEC Filing Workbench — interview brief

**The problem.** A plausible answer with the wrong row or period costs more trust than a visible refusal. This demo helps an analyst inspect a small filing pack, trace reported values and perform a few defined comparisons. It does not attempt enterprise archive coverage.

**The design choice.** Preprocess each PDF version once with pdfplumber, keep its hash, text, page image and coordinates, and build a small fact cache. Use lexical search for passages and exact identities for financial cells. Embeddings could help find a topic; they cannot establish which net-income row, scale or quarter is correct. No vector index, unrestricted SQL, OCR service or MCP wrapper is in the live path.

**The division of work.** The model chooses useful tools and proposes evidence. Code owns admission, arithmetic and final rendering. Standard uses Terra without reasoning and our Responses loop. Thinking uses Astra medium with the fast service tier and the hosted Agents loop. Python Decimal is available in Thinking; the backend recomputes every admitted calculation with decimal.js. More tool calls are not a confidence score.

**The proof to show.** Ask for Q2 revenue, open the selected quarter column and the excluded H1 candidate, then expand the final checklist. It records the actual source, label, header, signed token, dates, exact fact identity, requested operands and answer-binding checks. A lookup does not claim arithmetic. Unused visual or reconciliation tools do not earn a pass. For margin change, show both income/revenue pairs and the unrounded-ratio difference in percentage points.

**The important correction.** The earlier validator checked evidence but allowed a wrong sentence beside it. Four deliberate prose mutations demonstrated that weakness. New numerical answers are rendered from validated objects; the model's numerical prose, period description and calculation note cannot supply the headline. The historical failure remains available beside the corrected controls. Old conversations have not been retroactively relabeled as validated.

**How to defend the numbers.** Exact identity means source, label, metric, signed amount, units, basis and start/end/duration. Wrong period is wrong. Decimal results use a 1e-8 evaluation tolerance. The runtime accepts a model receipt within 0.005 and then replaces it with the Decimal result; reconciliation separately allows 1.5 million for rounding. See the current run's per-case results, errors, refusals and latency in the readiness page. The source key was inspected again by the assistant, not independently signed by a person.

**How failure behaves.** A saved function result has a session/turn/call identity. Its exact delivery payload and idempotency key are durable; an uncertain acknowledgement reuses them. A paid visual read is not silently repeated. A controlled image-only page makes the native-text parser abstain, exposing a useful future OCR/Docling test. Reader agreement and a balancing equation are consistency checks, not independent ground truth.

**Where the claim ends.** Explicit financial language and six consolidated metrics are supported; narrative entailment, alternate layouts, amendments, segment/cash-flow identities and filing authority remain limited. A numeric-claim guard can withhold a useful narrative answer, so refusals count against coverage. No small benchmark establishes enterprise accuracy. The most valuable next evidence is an independently reviewed, held-out set of unfamiliar questions and layouts, followed by an analyst usability trial. Enterprise work adds accounts, access control, retention, monitoring, version policies and enforceable spending controls.

**Eight-minute demo.** Spend two minutes on question → excluded candidate → source; two on Decimal and the checklist; one on the historical mutation and corrected output; one on an actual failed/refused run; and two explaining the limits, current measurements and the next experiment. Practice one unfamiliar change yourself: a polished app and agent-written tests do not prove hands-on understanding.

**Current public transport finding.** Standard passed a four-operand margin comparison. Thinking subsequently returned a correct bound lookup after reading the original PDF, while its managed query_facts call failed with HTTP 424. The failed tool is visible; do not describe it as a successful hosted function call. Use Standard for the live walkthrough. The separate local-to-cloud delivery recovery test does not settle this production failure.