Skip to main content

AI expert

APPROVED, IN BUILD — this is an approved design for the Binary Networks platform's Standards Library. It is not merged application behavior. The work file explicitly says design only; the AWS evidence says the API role has no Bedrock permission. A:.set/work/P2.22.md:3-10 E:evidence-aws.md:28,87-88

The design reads text PDF, CSV and JSON after upload. DOCX, XLSX and scanned PDFs remain unsupported. Extraction does not generate AI answers. The ask path sends the square's prepared text and has no tools, URLs or S3 credentials. Inactive squares disable the expert. No chat history is retained. A:.set/work/P2.22.md:30-47,75-85,117-183

Boundary or settingApproved design and honest limit
DBstandard_pages RLS and filtered security-barrier views use set.standard_id. This prevents missing filters; hostile SQL can change the setting. Ingest metadata is broader. A:.set/work/P2.22.md:79-82,235-271
S3Extraction uses a read-only square STS session. The extraction task can request other square sessions if compromised; there is no malware scan in this release. A:.set/work/P2.22.md:82-85 A:.set/work/decisions-2026-10-26.md:14-15,33-34,50
Model and regionProposed Claude Sonnet 5.5, eu.anthropic.claude-sonnet-5-5, EU-only profile from eu-north-1. The app pack verifies active eu.anthropic.claude-sonnet-5-5 and eu.anthropic.claude-opus-5-5 profiles, but says the app does not use them. Invocation access, evaluation and actual processing locations remain NOT VERIFIED. E:evidence.md:42 A:.set/work/P2.22.md:34-39,94-106,168-175 E:evidence-aws.md:87-88
Input budget300,000 tokens per question. Extraction and token estimates must be verified in implementation. A:.set/work/decisions-2026-10-26.md:62 A:.set/work/P2.22.md:138-139,160-164
SpendEvaluation: $20 total and $10/day hard caps. Real operating caps follow measured spending. Design reserves worst-case spend before each attempt, settles from usage and charges ambiguous failures at worst case. A:.set/work/decisions-2026-10-26.md:22-23,62 A:.set/work/P2.22.md:293-355
Emergency stopCap zero is Ron's approved kill switch; a deployed test is still required. A:.set/work/decisions-2026-10-26.md:68-72 A:.set/work/PS.14.md:50-52
QuotesNormalize Unicode/whitespace, remove Hebrew marks and bidi marks, require at least 20 characters and a substring on the cited page. This proves quote provenance, not correctness or whether the interpretation follows. A:.set/work/P2.22.md:84-85

The latest evaluation material decision is IPL 13 only, superseding ISO 27001 plus IPL 13. Ron must approve the 50 questions before the evaluation. No evaluation was run during documentation work. A:.set/work/decisions-2026-10-26.md:62-64 A:.set/work/P2.22.md:37-39

NOT VERIFIED — merged migrations, deployed extraction, EU routing, model access, spend enforcement, quote-verifier results and the Sonnet-versus-Opus evaluation. Verify with code, isolation tests, a deployed digest and sanitized app evidence. Neither model availability listings nor Terraform grants prove a working expert.

APPROVED, IN BUILD — the later app pack records expert implementation work on item/W3-expert at 0d30a21, including schema, spend controls, verifier and extraction. It is not integrated into this documentation source branch or deployed. The current API definition has no EXPERT variables, Bedrock policy or extraction task family. This updates the work file’s earlier “design only” wording without promoting it to merged. E:evidence.md:18,27-33