Not a chatbot. Not a search box. A command center over every fact the company has ever produced — papers, patents, books, failed experiments — where every answer carries its source atom and nothing ships untraced. We dogfood it on our own estate. The ledger on the right is what that looks like.
Read top to bottom — that is the direction data flows. Nothing enters the graph unverified, nothing leaves it untraced.
PDFs, DOCX, code, emails, notes. Documents are not chopped into blind 1,000-token windows. They are split into atoms — discrete units of knowledge, 582 chars on average — because the right unit of retrieval is one testable mechanism, not an arbitrary window. Anything that fails source verification goes to quarantine. It is not deleted. It waits.
A bi-encoder scores two vectors apart — fast, and blind to the difference between an answer and a keyword match. We rerank with a cross-encoder that reads query and candidate together, on top of dense + sparse BM25 hybrid so exact terminology never gets lost. Bench separation went 0.035 → 3.5. That is the 100×.
Four atoms drawn from four pools — everyday, library, pool_d, legacy — and forced into combination. The transport-prohibition gate kills any mechanism that quietly assumes solid-state conduction: the combination must move energy or matter through a named path — plasma, corona, ionic, contact transfer, capillary. Fiction that survives physics becomes an invention candidate.
Draft (7B / DeepSeek) → filter (cheap_rate discards garbage before a human sees it) → judge (GLM, external, names the exact faults) → rewrite → trace gate: every sentence is sourced to an atom or a logged fault-fix, or it is flagged → template gate: output matches the target format — NASA proposals, SSRN papers, patent claims. The gates are why nothing untraced ships.
Six of these are ours — produced, shipped, costed. Six are engineered projections we have not yet run on a client corpus. The tag tells you which.
● PROVEN = produced in our own runs or shipped artifacts. ● HYPOTHESIS = engineered and costed, not yet run on a client corpus. We will not blur the difference. Neither should your current vendor.
No waitlist form. No CRM sequence. No discovery-call funnel. Send your data problem to david@pietarien.com — you get a scoped build plan and a fixed quote within 48 hours, from a human.
Your corpus, atomized. Cross-encoder retrieval + hybrid search, basic API. No engines. Fixed scope, fixed price.
Pilot + Pieta cross-pollination + the 6-stage Refinery. External reviewer integration. 30-day run-in with your team.
Full Brain, multi-tenant, custom engines, roadmap seat. For companies that want the machine inventing on their estate, permanently.
Modes are for scoping only. Every build gets a written scope and a fixed price before a line of code ships.