SayDo-Agent

1. 1. Resolve

   Match the issuer in the question; fall back to SEC's ticker map

2. 2. Onboard

   Unknown ticker → XBRL actuals in ~2s, stored with provenance

3. 3. Route

   Rules pick one of four tools; a model is asked only where rules are weak

4. 4. Retrieve

   Exact patterns → broad patterns → hybrid dense+sparse search

5. 5. Synthesise

   One Haiku call over numbered passages, citing by index

6. 6. Check

   Every number matched against the passage it cited

Key design decisions

• The say-do comparison is computed in Python, not by the model. Guidance comes from one filing and results from another, and the model receives the finished delta as a passage to write about — never the arithmetic. A model that computes a delta will occasionally compute it wrong, fluently, with a citation attached, and that is a failure the guardrail cannot see.
• The SEC rate limit lives in Postgres, not in the process. SEC caps automated access regardless of how many machines you use, so a token bucket held in Python memory becomes three independent buckets the moment there is a web surface, a bot and a scheduled ingest. The database is the one thing every process shares.
• Refusal is a prompt instruction, not a retrieval threshold. The best possible similarity cutoff — chosen with the answer key in hand — beat 'never refuse' by two questions out of fifty-eight, because dense similarity tracks topical vocabulary rather than entity identity.
• The model cites by index and the index is resolved on our side. Asked to reproduce an accession number a model will sometimes produce one that is plausible, well-formed and wrong; constrained to an integer in a known range, an invalid citation is caught by arithmetic instead of by inspection.
• The chunked corpus is a cache with an eviction policy, capped at ten issuers. Promoting an eleventh demotes whichever issuer has gone longest without a question — dropping its filings and keeping its XBRL actuals, because those cost SEC requests to derive and almost no disk to hold.
• Embedding moved off the VM because it was measured, not because it felt slow. bge-small on an e2-micro whose burst credits are spent runs at 0.13 chunks per second, which turns a 100-chunk ingest into thirteen minutes and looks identical to a hang; the same model on Cloudflare Workers AI runs at 26 chunks per second, and the ingest that had never once completed finished in 13 seconds. The switch was gated on a check rather than a promise: Workers AI defaults to mean pooling and BGE is CLS-pooled, so the wrong default would have returned well-formed 384-dimension vectors that shared no space with the 2,649 already stored — no error, just gradually worse answers. Cosine against the local model and against stored corpus vectors came back at 0.999998.
• The VM speaks plain HTTP and never faces a browser. TLS and CORS terminate in a Cloudflare Worker, which also means the VM's address is configuration: a new IP is a five-second Worker redeploy rather than a rebuild of a statically exported site.