How This System Is Built — and Why

The operating system under Kairos · written for the team · 2026-08-10
Kairos runs on one bet: conclusions beat conversations. Chat is where thinking happens; this system is where it lands — dated, attributed, permanent. This page explains the machine under the engagement: what each piece is, and the reason it exists.

1 One shared brain, many agents

Each of us mostly talks to our own AI agent. The agents stay synchronized through one canonical git repository — kairos/kairos on Robertas's Forgejo server. Every change is dated and attributed, the full history is permanent, and any clone can restore the whole.

Why Agents drift apart the moment memory lives in three separate chats. One repository is the cheapest possible synchronization layer that is also a complete audit trail. And it is self-hosted on the client side's own hardware — no third-party platform holds the engagement's memory. The old GitHub copy is a transitional bridge on its way out.

2 Conclusions land, or they're lost

A decision that lives in chat scrollback is gone in a week. So the repo has exactly three landing zones: DECISIONS.md — what's settled and why, so nothing gets re-litigated; STATUS.md — who owns what right now; CONTEXT/ — the corrected facts, each carrying its date and confidence label.

Why The evidence discipline is the product. One month is n=1, never a trend. A self-report is a claim until corroborated. Killed claims stay visible so they cannot return under new wording. The engagement asks Mysterium for decision discipline — it has to run on the same standard itself.

3 Operational language

Four words carry the analytical load here, and they are not interchangeable. Fact — computed from named source rows; carries its table, window, and query. Hypothesis — a falsifiable claim written together with its test: what evidence would confirm it, what would refute it, where that evidence lives. No test named — not a hypothesis yet. Assumption — a premise adopted untested so work can proceed; named at the point where the work rests on it. Hunch — pattern-feel without articulated grounds; allowed in conversation, enters the ledger only under that label.

Why The costliest analytical failure is quiet: an extrapolation sitting beside a computed fact, inheriting its confidence. The labels make that impossible — a reader always knows where measurement ends and judgment begins. Two corollaries: an extrapolation never borrows the confidence of the fact next to it, and market claims come from a live scan, never from a model's memory. Canonical wording lives in the shared repo's operating manual; this page mirrors it.

4 Doors with names on them

Every gated surface — this War Room included — uses individual identities: an email one-time code per person, no shared passwords, one sign-in for every room behind the door. Rooms are scoped to the identity: curated team rooms for everyone behind the gate; the raw working notes open only to Lee and Šaras — the working-papers rule of section 4, enforced in code.

Why A shared password cannot be revoked for one person, leaves no audit trail, and widens by forwarding. Named access widens only by explicit policy — a decision, not an accident. And a new name on the door is never silently a new reader of everything behind it.

5 The private/shared seam

Everyone keeps private working notes. That is normal professional practice, not secrecy — working papers stay private wherever thinking has to be free to be wrong. The system enforces the flip side: any conclusion that is safe for the whole team must land in the shared repo in the same session it was formed. Hooks check this mechanically — an agent session cannot end with shareable conclusions unpublished.

Why Two opposite failures, one contract: indiscriminate copying would leak private material; blanket privacy would leave the shared brain empty. Drafts stay private. Conclusions are an obligation.

6 The machine guards itself

Humans own judgment and approvals; agents own mechanics. After one-time setup nobody has git duties — ever. Validators reject malformed state before it can commit. Backups run in layers: server-side instance backup on the NAS, plus daily self-verifying full-history bundles held off-NAS on Lee's side.

Why A rule that lives in prose gets forgotten; a rule that lives in a hook fires every time. Anything recurring and manual is treated as a bug and automated. People do the thinking the machine can't.

7 Why this much machinery for a two-person unit?

Because the deliverable is not a report — it is decisions Mysterium can trust. Those require an operation where facts carry confidence, decisions carry dates and owners, access carries names, and nothing depends on anyone's memory. This system is that operation, running in the open where the team can see it.

← Back to the War Room
Written and deployed by Lee's agent — Claude (Fable 5) · this page ships through the same machinery it describes.