Bob's Big Brain — how the machine works

Your team knowledge system in one page: what proposes, what decides, what's on the receipt — and exactly where an LLM is (and is not) in the loop.

The rule the whole system is built on: the model proposes; the deterministic kernel disposes. An LLM writes suggestions. Plain code — no model, no network — decides what becomes durable memory, writes the hash-chained receipt, and answers searches. Today's whole-machine digestion ran entirely without an LLM: raw import in, deterministic govern out.

The pipeline

Propose

You + Claude sessions LLM

brain_capture — a session writes a candidate fact ("we decided X", "the fix for Y"). It's a proposal, nothing more.

Propose

Nightly compile LLM

ICO reads the day's work and derives knowledge — summaries, concepts, contradictions — keeping raw and derived strictly separate.

Propose

Bulk import NO LLM

Raw documents in, verbatim (today's 17,165-candidate machine digest). No model touches the content.

all proposals land as JSONL in 

Buffer

The spool — ~/.teamkb/spool/

A waiting room. Nothing in here is trusted, searchable, or durable yet.

brain_govern drains it

Decide

Govern kernel NO LLM

dedupe → policy & secret-scan → promote or reject. Deterministic code (this is what ran for ~6 hours today, zero network connections). Every single decision appends a SHA-256 hash-chained receipt.

only promoted facts become durable

Durable state

Curated memories — teamkb.db

The team brain. Lifecycle-managed (active / superseded / archived), never edited by a model.

Evidence

Audit chain — the receipts

Append-only, hash-chained by protocol (prev_hash per event) — tamper-evident, verifiable with brain_audit_verify.

indexed for retrieval

Retrieve

qmd BM25 index NO LLM

Keyword search, zero ML. That's why full sentences return nothing but "backup", "deploy", "shipped this week" hit.

Answer

brain_search → cited results

Every answer carries a qmd:// citation back to the governed memory it came from. No citation, no claim.

So — is an LLM part of any of this?

Where an LLM works (proposal side)

  • Claude in your sessions, writing brain_capture candidates
  • The nightly ICO compile passes (deriving summaries from raw)
  • The nightly review agent that reads the capture inbox and drafts a digest
  • Slack alert normalization (Groq/NVIDIA) — cosmetic, ops-side only

Where an LLM never touches (control side)

  • Dedupe, policy, secret-scan, promote/reject — the govern kernel
  • The audit chain — every receipt is plain crypto, no model in the loop
  • Search — BM25 keyword matching, no embeddings, no reranker (by ratified decision, until the eval demands more)
  • Backups — age-encrypted archives to the VPS + R2, restore-tested nightly

Today's whole-machine digestion, traced

17,165candidates imported (51.7 MB spool)
6,501auto-superseded by deterministic dedupe
10,751active memories after promotion + cleanup
11junk survivors archived, each with a receipt
23,791hash-chained receipts on the audit trail
✓ Digestion complete (2026-07-16). Dedupe ran deterministically, the junk sweep is receipted on the audit chain, and everything above is searchable. Search tip stands: two or three strong keywords, not sentences — and skip dots/hyphens in terms ("currency fixes", not "CLAUDE.md currency fixes").