# North star — v3, the self-funding loop, one step at a time

*(a waypoint: drafted with agents, accepted only by the human, versioned, provenance recorded — this document is the normative layer)*

**Harbour keeps human intent in command of AI execution: it meets a developer where they are, completes verified backlog work at a cost and cadence a solo operator can sustain, funds itself doing it — and proves every word.**

**Verified beats claimed.** Work counts only when its evidence chain closes: CI green, merged, ledger discharged. done is a claim; the artifact is the fact.

**Cost per verified task, visible and falling — in the money actually spent.** Every session attributes its spend to the lane that billed it; cash the operator feels is the headline, API-equivalent keeps it honest. Every cycle forecasts the next and scores the last against actuals. Routing, plan-review, and re-grounding economy are judged by this number alone — and pricing is policy, never judgement: no agent reasons about price; defaults are config, changed only behind an eval, reverted in one line. Frontier price buys judgement only: work-shaped legs run on the cheapest tier that passes the verifier; judgement legs may run on the frontier. The subscription is a lane, not a ceiling: the exchange rate at which the API lane costs less than the plan is a published number.

**Operator minutes and sessions are the scarce resources.** Silent failures and detection gaps outrank feature work — a halt the system didn't report costs more than the halt. Follow-on tasks and wakes per verified task are tracked taxes; work that shrinks them ranks high. No autonomous run starts without a declared task budget, enforced at the seam — a run that cannot finish inside its bound hands back; it does not sprawl. False-escalation rate is a headline KPI: every surface that asks a human is judged by how often the answer was "why was I asked this?".

**Gates buy evidence, never delay.** Every phase gate is a measured artifact, not a declared intention — and not a calendar habit: shorten a gate by densifying evidence, never by waiving it. Time to any milestone is a policy choice and cost is a scope choice; neither excuses the other.

**Finish transitions before starting capabilities.** A substrate 90% migrated is a liability, not progress. The longest blocked chain and the oldest open compat lane are standing priorities.

**Meet a developer at the saved prompt, and make the next step safe before it is visible.** People come to trust agents in steps: they ask and copy; they drive a session by hand; they save the prompts that work; they hand over one task; they ratify a run of several; they feed tasks into a standing loop. Each step is gated by trust and budget. Harbour meets them at the saved prompt, where the grounded next prompt is the whole product: log in, connect a tracker and a repo, click a task, copy, go. Every feature names the step it serves; a person sees their step and the next one; handing over the first task means one task, one bounded run, and a human click before merge. Login → connected → first evidence-verified merge, without talking to the operator, is that step's bar. Secure defaults assume hostile input; payments land only behind the hardening gate; the free tier is a published, ring-fenced line item. The measure is people per step and where they stop.

**Every claim reachable from its evidence; every contract legible in one screen.** Reports cite dispatches, KPIs cite outcomes, relaxations name their adversarial follow-up. Complexity that breaks legibility is a defect even when locally justified.

*Provenance: v3 drafted with a Claude session on 19 September 2026 from a planning conversation with John Kershaw, under the authorship rule adopted on LIN-2926; accepted by John Kershaw on 19 September 2026; the diff against v2 and the sentence of his behind each change are in `docs/north-star-v3-proposal.md`.*