I have owned this through discovery and the decision about what was worth building. This is the point where it stops being mine. What follows is what a founding team picks up, what is already settled so nobody argues against it, what is still open and who owns each one — and the conditions under which we stop.
What exists · what does not · stated in that order
A tested idea, a plan, and nothing built yet.
Nothing here is built. That is not an apology — it is the honest description of a pre-build handoff, and the whole point of doing the discovery first is that the team that builds it does not also have to find out what to build. What is transmissible is the argument, the specification, the open queue and the owners.
A typed data model — DiffLine / Refusal / DiffRecord. Thirteen acceptance criteria. Eight edge cases. A refusal enumeration frozen at five reasons. Six standard car coverages as what version one lines up against, with everything outside them refused by name rather than dropped.
Nine stages, each with an evidenced exit condition rather than an asserted one — the gate passes because a test passed, not because someone said it did.
Eight gaps are listed explicitly in the architecture, four of them owned by someone other than an engineer. The legal question. The filing constraints. The manual-keying cost, which is recorded as pending rather than invented. Do not read the spec as complete; read §9 of it first.
And the largest one is not technical: nothing in the record establishes that supplying interpretation changes anyone's decision. That is the load-bearing assumption and the first thing the build is pointed at.
Four seats · what you own, what you cannot do, and your first two weeks
Gitwit assembles a founding team and keeps the studio underneath it. Harbor's specification carries six obligations, and a founding trio does not have six people — so they are mapped, and the mapping is the argument. Choose a seat and read what lands on your desk.
How many times the screen told someone to stay where they were, by week. It is the only number that cannot be argued with. If it reaches zero, the product has quietly become a sales page.
Nobody has been told to stay put for a week. Either the quotes genuinely got better, or someone tuned the screen. You are the only person who can find out, because you are the only person a conversion review is not leaning on.
Yours alone. Not the product lead’s — they are the one under pressure. Inert in this demonstration.The person who runs the diff is the person a conversion review puts under pressure. Authority to weaken their own surface under that pressure is the failure mode, not the control. No scoreboard rewards honesty here — including Harbor's own. This is the one structural decision in the handoff, and it is the one I would defend hardest.
Your first two weeks: answer the traffic question with real numbers, not a hope. Commit to the renewal diff in writing now, while Harbor has no book and the promise is free. Decide what you will tell a board when the best month looks like a conversion problem — because it will, and deciding it afterwards is deciding it under pressure.
Each one passes because a test passed, not because someone said so. You hold the go / no-go on every line.
Stage 6 is the one people skip. A “stay where you are” answer that was never run end to end before launch is an answer that never runs at all.
Your first two weeks: the schema and the blocked construction, then the null path through production before any traffic touches it. The order matters — a null path retrofitted after launch is a null path that never runs. Publish the denominator weekly from day one, while it is still embarrassing.
Seven checks. They write a result whether they pass or fail, so a silent skip is impossible. Nothing ships on a green that nobody looked at.
The one at 8.1% is the one to watch. It is people keying documents by hand because the photo could not be read. At 10% it throttles — on purpose, so the cost surfaces early instead of at scale.
Your first two weeks: ingest and the confirmation step, then the lint fail-closed in the request path. Wire the gates before the features. The first thing to measure is not whether the diff works — it is whether anyone will hand over a declarations page at all. Instrument that before anything else, because if the answer is no, the rest does not matter.
Words that appeared on someone’s real policy that we could not match to any of the six things we compare. They are shown to the customer by name, never quietly dropped. Your job is to rule on each one.
Thirty-one people were handed a document that changes their cover and names it only as “A264”. We cannot read it either — but we can say so, which is more than the document does.
Why these two sit with the studio and not the trio: they are specialist, low-volume and continuous — exactly the functions a three-person founding team cannot justify hiring for in month one and cannot safely skip either. This is the concrete answer to what studio support is for. It is not encouragement. It is holding two obligations the team would otherwise drop.
The first list is not for reopening · the second has your name on it
A handoff that hands over every question hands over nothing. These seven are decided. Each was argued, each has a reason on the record, and the reason is available — but the decision is not a conversation any more. Reopen one only with evidence that did not exist when it was made.
| Settled | Because |
|---|---|
| The moment is bind, not quote request | The quote request is the link every competitor is also optimizing. The inversion happens after it. |
| The surface may return “stay where you are” | A comparison that has never advised against its owner is a sales tool. This is the product, not a concession. |
| Refusal by printed name, never a silent drop | Naming what cannot be compared is interpretation. Pretending it was compared is the industry's current behavior. |
| Six coverages, one state, matched by hand first | The first cohort is done by hand so the mapping table is built from real documents rather than assumptions. |
| We check whether people will hand over the document at all, before anything else | The whole idea depends on people doing something first — finding a document. Test that before testing anything else. |
| No recommendation, ever — consequence only | Recommending a coverage level to a named person is a licensed activity. The line is hard and it is not a design preference. |
| Non-binding records are ephemeral | Holding a competitor's coverage detail on someone who did not buy is a liability with no upside. It is destroyed, and a canary checks that it was. |
An open decision is not a gap. It is a decision whose owner has not made it yet, and the owner is named.
One call to an insurance lawyer. It is off the shipping path — version one runs on the policy summary alone and consumes no interview output — so it gates a future surface rather than this one. It is still owed before Harbor builds anything that asks a buyer nine questions.
Every output is a statement of consequence, never of preference — that is resolved by construction. The residual is adjacency: an interpretation surface sitting next to a recommendation. The conservative reading is in force and ships regardless of the answer. The answer can only loosen it, never tighten it.
Below about sixty people, the test reports not enough people yet — neither pass nor fail. An insurer with no customers may simply not produce that traffic, and the correct response is to report it rather than to lower the bar until a number appears. Did not run is a real and informative outcome about Harbor's traffic, and it should be presented as one.
Recorded as pending rather than invented. The 10% cap and the throttle exist so the answer is discovered under control instead of at scale. The first cohort produces the number; nobody should estimate it before then.
In order · the cheapest disconfirmation runs first
Measured before anything else. If shoppers will not produce the document, the intervention is dead regardless of how good the comparison is. This is the cheapest thing to learn and it is learned first.
Among shoppers shown a material difference, against a no-diff control. Denominator, share and interval — never a bare number. If they overlap, the mechanism did not operate.
Not a product failure — an organizational one. It means the surface stopped being allowed to lose a sale. The canary exists because this failure is silent by nature.
This product succeeds by losing sales, and nothing in its design protects it. The two defences on the table — a public written commitment made before Harbor has a book, and a null count visible in production — are tripwires, not mechanisms. They make the killing detectable. They do not prevent it. I am handing this over unresolved and marked, because handing it over quietly is how it gets killed in month seven by someone who never knew it was load-bearing.
One condition
When I am not required.
Not when the documents are read. Not when the meeting happens. The handoff is complete the first time this team makes a decision I would have made, without asking me, and gets it right for the reason on the record rather than by coincidence.
Everything above is built backwards from that. The settled list exists so nobody has to ask me what was decided. The open list has owners so nobody has to ask me who decides. The stop conditions are numeric so nobody has to ask me whether it is working. If any of those three sends someone back to me, that part of the page has failed and I would rather fix it now than be the dependency.
The one thing I would say out loud rather than write: the load-bearing assumption of all of this is untested, and the build is pointed at it deliberately. If it fails, the record says where to go next — the claim, not the purchase. Both people in the record who moved of their own accord moved over how a claim was handled. That is a different company, and it is the honest second door.