A partnership contract that only describes what the client has to do is a vendor contract with better marketing. Symmetry is the test: what are we on the hook for, and where can you check that we did it?
The asymmetry to watch for
Most "partnership" language in the AI services market is one-directional. It describes how the client will give access, sign off on milestones, attend the weekly call, and pay the invoice. What the vendor owes — beyond "the deliverable, eventually" — is left to good vibes and a logo on the slide deck.
We treat that asymmetry as the tell. If the only commitments in writing are the client's, the relationship is still a vendor relationship. The partnership word is doing branding work, not contract work.
What we owe, in writing
Our engagement clauses include things we have to do, on a clock, with a falsifier (Class C — these live in the signed scope and the artifact list, not in a tagline).
Promptness, with a default. We commit to a response window on every channel we use with you — typically one business day for substantive questions, same-business-day for "is the thing on fire" pings. The window is in the contract. If we miss it, that is a fact in the ledger by Friday, with the reason. We do not get to redefine "prompt" after the fact.
Evidence on every non-trivial claim. When we tell you a thing works, we tell you which class of evidence we have for it: A = we ran it in front of you, B = the code or the artifact is inspectable, C = the configuration is in the repo and you can read it, E = an outside expert source supports it (cited), F = the falsifier is named, U = we have not verified this yet. The U class is the important one. Saying "we don't know yet" honestly is part of the deliverable.
The hard truth, scheduled. Once a week, in writing, we publish a status note that includes at minimum: what shipped, what slipped, what we got wrong, and the one thing we are most worried about. If the engagement is going badly we are required to say so by the end of the week we noticed it — not the end of the quarter. The clause is there because intent without a clause is not a commitment.
Exit-readiness as a continuous state. From day one, the artifact is operable by someone outside the engagement. We test this on a schedule, not at sign-off, so the artifact never silently rots into a thing only we can run.
The receipts we publish on ourselves
The point of writing these obligations down is that you can check them. So we publish receipts.
The running ledger. Every engagement runs against a shared document the client owns. Decisions, claims, evidence classes, costs, falsifiers — all of it. The ledger is the canonical source. If it is not in the ledger, it did not happen. Sample format is on /transparency-architecture-overview.
Response-time logs. The promptness clause has a metric, and the metric has a log. At review time we hand over the actual median and tail response numbers for the period. If we missed the window twice, you see both misses, with timestamps and reasons.
A "what we got wrong" section in every status note. This is the line item people skip. We do not. A status note that does not name a thing we got wrong is, by our standard, an incomplete status note. It goes back for a rewrite before it ships.
Public version of the standard. We do not get to soften our internal standard in private. The /partnership-vs-vendor-contract page reproduces the actual clauses, including the falsifier language and the response windows. If we have not lived up to that page in your engagement, the page is the receipt you point at.
What symmetry does to the conversation
When both sides of the engagement have written obligations and a public scoreboard, two things change.
First, the weekly call gets shorter. There is less ritual reassurance, because the ledger and the status note already did that work. The call becomes a place to make decisions, not to perform alignment.
Second, the bad week gets handled. In a vendor relationship a bad week is a thing the vendor hides until it becomes a quarter. In a symmetric partnership a bad week is the moment the model proves itself, because the clauses fire and the hard-truth status note actually goes out. We would rather have an honest bad week on the record than a quiet quarter we have to explain in a deposition (Class E — this is the broad pattern observable in published post-mortems across the AI services market).
What this is not
We are not promising to be infallible. We are promising to be auditable, on a schedule, against a written standard. Those are different commitments, and confusing them is itself a tell — the vendor who promises infallibility is the vendor who will quietly miss the falsifier.
The standing offer is the same as always: come to the intake call and walk the clauses against your actual situation. If the symmetry does not hold up under your questions, we want to know — that is the engagement we want to have.