Employee of the week
The founding interview: 8 manifest questions + a 3-question truth pass. Reading the customer's site settled 6 conditions before anything was asked, so the live interview was 5 manifest + 3 truth. It stops when every condition resolves — never on a counter. The consultant is the Genesis persona, not a seat: it interviewed, proposed this team, and handed over to the Orchestrator.
Questions asked 8 of 11
- Your website? — "nakheel-tech.example" → reading it settled 6 conditions
- What do you want fixed? — "people sign up and go quiet" → lifecycle over email
- Who buys? — "finance managers at small companies, they find us on LinkedIn"
- Who sells today? — "just me (the founder)" → enablement for one person
- Ads? — "budget and intent" → the seat installs, the tool gets asked for by name
- The stack? — HubSpot · GA4 · WordPress (declared; OAuth after the team is agreed)
- Competency questions? — 9 questions became the day-zero benchmark
- Times someone over-promised? — 5 incidents became the anti-claims fence
The condition ledger
| Condition | Resolution | How |
|---|---|---|
| Their website matters | yes | read from site |
| They publish content | yes | read from site |
| Signups go quiet | yes | problem question |
| Paid budget live or intended | yes — intent | ads question |
| Traffic arrives, won't convert | no | problem question |
| Sells via tenders | no | who-sells question |
| Video is a real channel | no — aspiration | read from site |
| Enough data to read | defer — month two | a time condition |
Recorded stop reason: "every condition resolved to yes, no or defer" — auditable, not a vibe.
The Hippocampus injects on every message and every sub-agent dispatch, and it is measured from hour one: the day-zero benchmark was seeded from the customer's own competency questions.
Quarantine memory/_inbox — invisible to retrieval
- "Competitor cut prices 20%" — source: sales call · pending
- "Reference customer approved being named" — source: email · pending
Agents never write truth/ directly. A fact learned mid-task sits here, invisible to retrieval, until the facts-gate owner promotes or rejects it with a dated note.
The seven organs
- Injection — every prompt + every sub-agent live
- Provenance gate — blocking pre-commit live
- Quarantine — invisible to retrieval live
- Learning — corrections become files same-session live
- Measurement — the benchmark gates every change live
- Maintenance — rot audit on demand live
- Visible pulse — this screen you're looking at it
One file per entity, every file carrying a source and a verification date. A missing fact is written TODO-verify — an honest blank beats an invented number, and the pre-commit gate blocks the rest.
truth/company.md
Who buys: finance managers at 5–50 person companies
Markets: Saudi Arabia · UAE · Languages: Arabic · English
Pricing: TODO-verify — not asked in the interview; pulled from the site after connect
truth/anti-claims.md built from real accidents
- Never — discounts above 25%: a rep once promised 40% that didn't exist
- Never — claim an unbuilt integration: a chatbot did, once
- Never — name a customer without written approval
- Never — promise a feature date before it's announced
- Never — "cheapest on the market": a relative price claim breaks in a day
Conflicts register: empty — when two sources disagree on a figure it is marked CONFLICTED = unusable, never silently resolved.
Every deliverable passes two gates in order — facts, then quality. Every verdict is logged with the producing agent named, and the numbers below are computed live by the real v2 formula from the demo verdict stream — not typed in.
Agent Strength v2 — from evidence, never self-graded
5 × (0.40 grounding × facts-gate pass rate + 0.30 stability with rework discount + 0.20 depth + 0.10 freshness). A seat with no verdicts shows "—": no evidence yet, and the blank is honest.
Verdict stream data/gate_verdicts.jsonl
Deliverables, never behavior: everything the function shipped — agents and humans, inside and outside the system — attributed per person and mirrored into the company's own PM tool. What isn't logged didn't happen.
Latest deliverables attributed
| Deliverable | Who | Gates | Sync |
|---|---|---|---|
| Pricing page — draft v2 | Content Writing | facts quality | synced |
| Activation journey — 4 emails (draft) | Lifecycle Marketing | facts quality: in review | synced |
| Q4 keyword map | SEO | facts quality | synced |
| Welcome email approved & sent | Laila (marketing) | — | synced |
| Site schema fix recovered from git | Fahad (engineering) | — | synced |
Candidate queue: "prep a partner deck?" (from Slack) · "refresh the FAQ page?" (from a meeting) — waiting for a human yes before becoming tasks. The suggestion is automatic; the decision never is.
Account settings and the company profile the founding interview built. Editing the profile happens the way it was built: another interview, or a booked meeting.
Account
Demo fields - the real product edits these.
Company profile built by the founding interview
Who buys: finance managers at 5-50 person companies
Markets: Saudi Arabia - UAE - Languages: Arabic - English
Stated problem: people sign up and then go quiet
Pricing: TODO-verify - pulled from the site after connect
Saving repaints the whole workspace — the team's outfits, the screens, the accent, even the break-room rug. Faces stay human.
the consultant - the founding interview setup