Scope, non-goals and conformance classes
Version 1 covers marketing mix models (Bayesian and frequentist), third-party response curves, geo and user-level experiments, budget scenarios and the decisions they feed. Other model families join through family profiles (section 16) without changes to the core.¶
In scope for v1¶
- MMMs from in-house frameworks, modeling-api, Meridian, Robyn, PyMC-Marketing and vendors¶
- Vendor response curves as parameters or spend tables¶
- Test plans and results for geo, user-level holdout and ghost-ad designs, including plans imported from Google’s geo planning product¶
- Scenarios, optimisations and decision records¶
- Lineage of every dataset and change feeding or leaving our own models¶
- The audience family profile, at the Sandbox and Pilot lifecycle stages only¶
Non-goals. UMP does not fit models or prescribe how to fit them. It does not replace Meridian, Robyn or vendor tools, and it is not a dashboard. It does not make decisions: machines draft, named people approve.¶
Conformance classes. A tool claims one or more classes and proves each against the conformance suite (section 21).¶
| Class | Must do | Proved by |
|---|---|---|
| Producer¶ | Write valid artifacts with full estimand, representation and provenance fields | Schema and round-trip fixtures |
| Consumer¶ | Apply the fallback a gate prescribes (or stop, in strict mode); show trust and fallback labels | Expected-error fixtures |
| Engine adapter¶ | Answer standard requests it declares, with typed errors for the rest | Score, optimise and calibrate fixtures with known answers |
| Gate engine¶ | Evaluate published policies on every operation and transition each gate’s applies_to names | Policy unit tests in the suite |
| Agent host¶ | Prefer validated artifacts, write drafts only, cite or label a best guess | Golden-question and injection fixtures (section 17) |
| Log verifier¶ | Verify records, checkpoints and inclusion and consistency proofs of the integrity log independently | Section 18 test vectors |
| Hosted endpoint¶ | Answer the standard requests a Tier X model declares over HTTP, with typed errors for the rest | Engine profile fixtures (section 8) run over HTTP |
Class requirements. Each class is bound by the requirements tagged with it in the conformance checklist published with the suite (section 21). The minimum requirements include:¶
- REQ-02-01 (Consumer): a Consumer MUST preserve every
fallbacksentry it receives and MUST re-emit it with any result it derives or forwards.¶ - REQ-02-02 (Consumer): a Consumer MUST NOT present a derived result whose input fallbacks it dropped.¶
- REQ-02-03 (Consumer with a user interface): it MUST show the trust label and every fallback label beside the number.¶
- REQ-02-04 (Gate engine): it MUST evaluate each gate on every operation and transition named in that gate’s
applies_to.¶ - REQ-02-05 (Log verifier): it MUST reproduce every section 18 test vector, check every checkpoint’s consistency proof, and raise
integrity.brokenon failure.¶
Conformance claim. A claim MUST be a machine-readable document giving the class, spec version, suite version, threshold-set version, the standard requests and model tiers covered, and the SHA-256 digest of the suite report. A partial claim MUST list what it covers. Anything unlisted is unclaimed, not conformant.¶
Design rules (rules 8 and 9 extend the original seven):¶
- When the best answer is unavailable, use the next best and say which was used.¶
- Each model declares what it can calculate; tools check before asking.¶
- Uncertainty travels with every result; summaries are labelled as summaries.¶
- Never overwrite; a new version points to its parent.¶
- Media names match through a maintained lookup, never by similarity.¶
- Each artifact states how far it can be audited and claims no more.¶
- Independent of language and storage.¶
- A number travels with its estimand; any part that is unknown is stated as an assumption and widens the range. Schema-valid numbers can still answer different questions, so meaning travels with the value (section 1).¶
- Gates degrade before they block. Missing or weaker evidence leads to a labelled best guess; work stops only where going ahead is unsafe. A blocked step pushes the work into unlabelled spreadsheets, where no gate sees it at all.¶