Most AI contracts are written so the vendor wins on day one and the client discovers it on day three hundred. The language that flips that outcome is not exotic. It is five clauses, and they fit on one page.
This post lays them out, with redline examples you can hand to your counsel and demand from any firm that wants your business. We use them ourselves. They are why our engagements have an off-ramp on the first day they have an on-ramp.
The frame: what "partnership" actually has to do
A vendor delivers an output and keeps the means of production. A partner delivers the output and the means of producing it again without them. Everything else is marketing. If a clause in a contract makes you more dependent on the supplier over time, it is a vendor clause. If it makes you more independent over time, it is a partnership clause.
The five clauses below are the smallest set we have found (Class C — configuration & integration from our own live engagements) that produces a contract you can walk away from cleanly at any point, with everything you paid for in hand.
Clause 1 — Exit-Ready Artifacts (delivered weekly, not at termination)
The standard vendor pattern is: artifacts live in the vendor's systems, and at the end of the contract you get a data dump and good luck. The partnership pattern is the inverse.
Redline language:
Supplier shall deliver, at the end of each calendar week of the engagement, an "Exit Pack" to a Client-controlled storage location. The Exit Pack includes: (a) all source artifacts produced that week (code, prompts, configs, datasets, model weights or pointers, dashboards, runbooks); (b) the SHA-256 hash of every artifact in a signed manifest; (c) a written summary in plain English of what was done, what is now in production, and what would have to be re-built if the engagement ended on Friday. Client may at any time terminate for convenience and shall owe no further fee beyond the most recent completed week.
The point is not the legalese. The point is the cadence. Weekly means the off-ramp is always one week old at worst. Termination for convenience with no penalty means there is no contractual lock-in dressed as a "transition fee." A vendor will resist this. A partner will already be doing it.
Clause 2 — Shared Ledger Access (read access from minute one)
The shared ledger is the single, append-only record of what happened, who did it, and when. Every meaningful AI engagement should have one. The vendor pattern: the ledger lives behind the supplier's login and the client gets a quarterly export. The partnership pattern:
Redline language:
Client shall have read-only credentialed access to the engagement's full operational ledger (work performed, model calls, prompts, evaluations, deployments, errors, costs) from the first business day of the engagement onward, with no roll-up, no redaction, and no rate-limiting that prevents real-time inspection. Supplier shall not maintain any "internal" log of engagement activity that is not also visible in this ledger.
Two operational tests, both falsifiable (Class F):
- The client's own staff can pull the ledger at 2 a.m. on a Tuesday without asking permission, and the data is current to the minute.
- There is no second ledger. If a supplier engineer would consult logs the client cannot see, the clause is being violated.
We publish our own ledger discipline on /transparency. If a supplier won't agree to a shared ledger, that is itself the answer to whether they intend partnership.
Clause 3 — No Dark Dependencies
A dark dependency is anything the running system needs that is not on the client's inventory. A private API key in someone's head. A scheduled job on a personal laptop. A model endpoint whose owner is one person at the supplier. Dark dependencies are how vendors stay essential.
Redline language:
Supplier warrants that no component of the delivered system depends on any credential, endpoint, scheduled job, person, vendor relationship, or piece of infrastructure that is not (a) listed in the Dependency Inventory delivered with each Exit Pack, (b) owned by Client or held in a Client-controlled vault, and (c) replaceable by Client within five business days using documentation provided by Supplier. Discovery of an undisclosed dependency is a material breach.
The "five business days using documentation" line is the teeth. Anyone can list dependencies. Few suppliers can document them well enough that a competent engineer who has never seen the system can swap them out in a week. That is the partnership bar.
Clause 4 — Published Falsifiers (every load-bearing claim has one)
A falsifier is a written statement of what would prove the claim wrong. If your supplier tells you the new system "improves outcomes by 30 percent," the partnership move is to write down, before the work starts, exactly what observation would make that claim false.
Redline language:
For every quantitative or qualitative performance claim Supplier makes about the delivered system (in proposals, status reports, dashboards, or marketing), Supplier shall publish, in the shared ledger and on a public artifact page, the corresponding falsifier: a specific, measurable observation that would, if recorded, disprove the claim. Claims for which no falsifier has been published may not be used in Client-facing communications.
This is doing more work than it looks. It kills three failure modes at once: vanity metrics (no falsifier means no claim), retroactive goalpost-shifting (the falsifier is timestamped before the result is known), and the "AI improved everything" handwave that fills industry case studies. The Themesis writeup on the actual hiring market in 2026 — "Meet Jay Kumar Chimata: JobFirst.ai and the Real AI Job Market" — is a useful adjacent read: employers, per Chimata's 22,000-user platform, are paying for verifiable shipped outcomes, not credentials. Falsifier-published contracts are how an agency proves it ships to that bar.
Clause 5 — Evidence-Class Disclosure on Every Claim
Our internal evidence taxonomy is public: A empirical in this session, B code/inspection, C configuration/integration, E expert citation, F falsifier present, U unverified. Suppliers should disclose, per claim, which class the evidence belongs to.
Redline language:
Each performance, capability, or quality claim Supplier makes about the engagement shall carry an inline evidence-class tag drawn from a published taxonomy (or equivalent), accessible to Client. Untagged claims default to class U (unverified) and may not be invoked in pricing, scope, or renewal discussions.
This is the smallest possible change to how AI services are sold and it has the largest possible effect. Every "our model achieves 94 percent accuracy" becomes either "94 percent (Class A, falsifier published)" or it goes away. (Class E for the taxonomy itself: cf. the discipline of classified evidence in Bayesian statistical practice, where prior provenance is tracked rather than hidden.)
Why these five, and why now
The macro reason is straightforward. As industry chatter about more autonomous systems grows louder — the Themesis essay on pseudo-persons, tools, and what is and isn't yet here argues persuasively that the prep window is short — clients who lock into opaque AI vendors today will be the most exposed when those vendors get acquired, pivot, or start charging rent on systems the clients can no longer leave. The five clauses above are the cheapest insurance available against that exposure, and they cost nothing to demand on day one. They cost a great deal to retrofit on day three hundred.
The micro reason is that we have used these clauses in our own engagements and they work (Class C). They do not slow down good work; they slow down extractive work, which is the point.
What to do if your current supplier won't sign these
Read their answer carefully. A supplier that pushes back on Clause 1 (Exit-Ready Artifacts) does not want you to be able to leave. A supplier that pushes back on Clause 2 (Shared Ledger) does not want you to be able to see. A supplier that pushes back on Clause 3 (No Dark Dependencies) is, by definition, planning to be the dark dependency. A supplier that pushes back on Clauses 4 and 5 wants to keep selling on the basis of unfalsifiable claims. None of these are partners. They are vendors who have correctly understood what partnership would cost them.
The list of things we promise not to claim, ever, is on /standard/what-we-do-not-claim. The full posture is on /anti-extraction-engagement-model. If you would like to walk through these five clauses against your own active contracts, that is exactly what the /workshop session is for.
— Michael