EPEna Pragma / Infrastructure Report

The shared vault, rebuilt.

The EP company vault went from an un-governed 8,000-file sprawl to a scalable, multi-writer federation that Prism, Echo, and Carl co-own, and it now generates and tests its own new knowledge.

Repo ena-pragma/ep-vault Shipped 2026-07-11 Merged PR #77 · PR #78 Rollback tag pre-migration-2026-07-11

The vault was never empty. It was un-governed: no operating contract, no single home per fact, competing wikis, a dead 31-agent fleet, and no way to search or maintain it. The rebuild gave it a written constitution and a clean shape, consolidated the sprawl, and added a generative loop that proposes new conclusions and refuses to trust them until reality has graded them.

Nothing was lost. Every move is reversible in git, the client deliverables stay in the repo for every co-owner, and one tag rolls the whole thing back. Prism built and merged it; this page is the read-in for the people who share it next.

What changed

From sprawl to a governed shape

The same content, filed by owner and lifecycle instead of by accident. The headline moves:

MeasureBeforeAfter
Operating contractnoneroot CLAUDE.md
Live agent homes31 dirs2
company/ subfolders3931
Canonical wikis2–31 (238 pages)
Remote branches197
Generative loopnonelive + gated

How it functions

Four zones, one home per fact

The old vault tangled four different things into flat peers. The rebuild separates them by who owns them and how they change. Every fact now lives in exactly one place, which is both the filing rule and the reason two writers almost never touch the same file.

Knowledge agents synthesize

Immutable sources plus the one canonical wiki. Our current best understanding, revised freely, never silently overwritten.

raw/   wiki/

Agent runtime each owns its own

Identity, memory, checkpoints, and per-agent specialty wikis. Prism and Echo, and the pointer to Prism's sovereign vault.

agents/   domains/

Company record humans decide

What has been decided, not synthesized: mission, clients, SOPs, ratified decisions. Agents propose; humans ratify.

company/

Plumbing coordination

The board room, shared skills and keys, and the report-only maintenance zone that automation writes to and never past.

office/   infra/   _maintenance/
The load-bearing line: wiki/ is our current best understanding (agents rewrite it, invalidate-don't-delete). company/ is what we have decided (human-owned truth, changed only by a reviewed proposal). That split is what makes three writers work.
ingest a source becomes synthesis query find + file the answer back conjecture generate + test new claims lint health, connect never prune maintain report-only, never edits

The new part

The vault now discovers, not just remembers

The generative loop reads the knowledge graph for conclusions it already implies but nobody wrote down: a page links B, B links C, but nothing links A to C. That missing edge is a candidate. On the live 238-page graph it surfaced 3,392 of them.

A links B links C missing···→ claim?

A candidate is only a question. It becomes a card with a claim, a falsification condition, and a test plan, and then it faces the gate that is the whole point of the design: only a SOUND verifier, reality itself, can promote a claim into curated knowledge. A council of AI critics is a WEAK cross-check and may never promote a claim on its own. An untested conjecture never enters a curated page.

Born-validation · CJ-001 Refuted, gate held

Claim: as general-purpose AI commoditizes, defensible value concentrates at two poles, platform ownership (Apple, on-device) and vertical domination (Shield AI, defense), as two instances of one force.

Weak · council4 of 4 critics refuted it, with 2026 evidence. Suggestive only; cannot promote.
Sound · evidenceConfirmed the parts (both poles are real; horizontal prices are falling) but not the unifying mechanism. Apple even rents part of the pole, paying Google to run Gemini in Siri.
Outcome: the parts are true, the bridge is not proven, so by the card's own rule the claim was refused promotion and kept in the ledger as a stepping-stone. A plausible-but-unproven idea was correctly kept out of the company's trusted knowledge. That is the gate doing its job.

How it scales

Built to grow like a company

Adding a person, an agent, or a client is a template plus one ownership line, not a reorganization. Each writer's high-frequency work (their own domain, their checkpoints, their status) is direct-push with no waiting. The shared truth (the wiki, the company record) is proposal-and-review, so nothing shared changes unseen. Ownership is routed softly by folder and recorded per page, so a three-author vault never deadlocks on a merge, and search is already wired so the vault stays findable as it fills.

Get your data in

What each of you does next

The vault is ready to receive. Two ways in: promote a section of your own vault as a reviewed bundle, and file status at end of day and end of week. Here is the concrete path for each of you.

For Echo

Bring your knowledge in

  1. Confirm your home. Decide whether the shared vault is your live home or whether, like Prism, your live home lives elsewhere and this vault points to it. This is the one open call that shapes the rest.
  2. Claim your domain. Copy the domains/_template into domains/echo/ for your specialty synthesis. It is yours, direct-push, no review needed.
  3. Promote sections. When a page of yours becomes shared truth, push it as a bundle (a manifest-checked, leak-gated export) into the shared wiki/ by pull request. The shared copy carries provenance back to yours.
  4. File status. At each checkpoint, drop a short end-of-day entry in status/; it auto-merges. The weekly rollup gathers them into the company timeline.
  5. Own your lane. Anything under domains/echo/ and agents/echo/ is yours to write freely; the shared core is where you propose.
For Carl

Bring the company record in

  1. Own company/. The decided record, clients, mission, SOPs, ratified decisions, is human-owned. You and Branden hold final say on what merges there.
  2. Add client work. File engagements under company/clients/ by pull request. Deliverables (PDFs, designs) stay tracked so every co-owner gets them.
  3. Ratify decisions. Record decisions and ADRs in their single home under company/; an agent may draft, but a human sign-off is what makes it the record.
  4. Review, do not hand-merge. Everything reaches the shared trunk as a reviewed proposal, so you approve rather than untangle.
  5. Settle the open calls. A short list below needs your and Echo's decision before the last pieces switch on.

Honest edges

Open decisions and what is deliberately teed up

The structure is done and merged. A few pieces wait on one call each, not on more building.