Eve library

P2-003 evidence pack — hierarchical taste models

Evidence

Revision 1

On this page

Work-package charter#

  • Task ID: P2-003
  • Owner: Shared contracts and evaluation-governance maintainers; implementation owner: Elikem Agudogo
  • Accountable product: Shared V1–V9 autonomous-content platform
  • Priority: P2
  • Target milestone: 2026-07-15 SOTA gap-closure program
  • Risk tier: High privacy, minority-preference erasure, manipulation, and hard-constraint trade-off risk
  • Dependencies: Zod, the canonical @oshun/contracts common export, product-owned consent/deletion systems, manipulation detection, and independently versioned safety/rights/factuality/accessibility policies
  • Affected schemas and content slots: New strict hierarchical-taste policy, node, and decision contracts plus common/root exports. No production preference record, learned model, user cohort, or content slot changed.
  • Rollout strategy: Publish the contract as a fail-closed governance boundary, require each adopter to populate all five hierarchy levels and exact hard-constraint references, and admit optimization only after its own privacy/consent/data-quality review
  • Rollback strategy: Revert the source commit and exports together before any adopter persists these records. Do not flatten stored child nodes into a global vector or treat a missing contract as permission to optimize.
  • Evidence-pack location: This file

Closure mapping#

Source commit 64d87fdf18c2a464f0c7486e27564fa392f52c70 implements the three checked ledger requirements:

  1. TASTE_HIERARCHY_LEVELS fixes the exact product → tenant → locale → expert-cohort → user order. Every child must reference a node at the immediately broader level, so a minority cohort or user cannot be silently collapsed into one majority vector.
  2. The policy requires bounded differential privacy, a per-subject contribution cap, independent aggregate and user sample floors, an uncertainty threshold, opt-out and deletion support, and a manipulation-detection reference. Undersampled, uncertain, opted-out, deleted, or flagged nodes are excluded from optimization and routed to broader fallback.
  3. Safety, rights, factuality, and accessibility are an exact separate constraint vocabulary with independently referenced policies. They are never preference-vector entries and cannot be traded against taste score.

This is a governance definition and evaluator, not a trained taste model or production preference corpus.

Verification record#

The current refresh ran against repository commit 47adfe1d4b4e4017a141faa785fdc0980cf18de7 on app-dev-server, Linux 6.8.0-134-generic x86_64, Node.js v24.15.0, pnpm 10.25.0, TypeScript 6.0.3, ESLint 10.4.0, and Vitest 4.1.7. No model/provider credential, database, or customer data was used.

UTC interval Exit Exact command and result
2026-07-29T11:21:38.029Z–2026-07-29T11:22:05.706Z 0 pnpm verify:contracts:hierarchical-taste-governance — targeted ESLint passed with zero warnings; the focused contract suite passed 15/15; libs/contracts/tsconfig.lib.json production typechecking passed

The implementation publication at 2026-07-24T14:51:44Z also passed the targeted contracts build and the repository's staged gates. The current task-scoped command above is the admitted refresh; it does not rely on an Nx cache or a historical pass.

Test and negative evidence#

  • Value/contract: The positive case requires the complete five-level hierarchy, exact immediate parentage, and five usable nodes.
  • Negative/adversarial: Parameterized fixtures remove each of the five levels; another binds locale directly to product; five fixtures independently exercise insufficient data, excess uncertainty, opt-out, deletion, and manipulation fallback; malformed privacy/consent controls fail schema admission; duplicate identities produce a deterministic blocker.
  • Hard-constraint non-substitution: The test requires exactly safety, rights, factuality, and accessibility outside the preference vector.
  • Integration/E2E/UI/accessibility/performance: Inapplicable to this schema-only package. No persistence adapter, API route, rendered surface, accessibility behavior, production data aggregation, training job, or latency boundary changed. Product adopters must add those avenues at their invoked boundary.

Artifact, authority, and blocker disposition#

Current artifact SHA-256
libs/contracts/src/common/hierarchical-taste-governance.ts b0c6382c722fbb525b285eaf2264e6ea18d528f628f99de870f53e4b0da11591
libs/contracts/src/common/hierarchical-taste-governance.spec.ts 1ce1ec83d6e76a230c2cf8b8a2ece5171db535fa09d117cac4d768a9d0bb77d4
libs/contracts/src/common/index.ts a557ed7f6dc5614b63cf60fc33ad6109c5831757a845125d016df710dd5e0d25
package.json c6f342a2399a4708ca7dcbe9d0f484f5da8915dc67c6a01cd10bc62884ef5c10
.github/workflows/content-slot-contract.yml 37e02a066006145bd238e8a89cd250d3d1149da372f81a38dfefc8469d550b19
  • There is no artifact release, source/canon/rubric change, production policy instance, runtime placement, screenshot, or media probe for this contract-only package.
  • No policy requires a human reviewer signature to publish the schema. No human approval is fabricated; future production aggregation/training requires its own privacy, consent, and product authority.
  • No P2-003 implementation blocker remains. Production taste data, learned parameters, subgroup outcomes, and adoption evidence are deliberately not claimed by this package and must be supplied by an adopting product.
  • Generated capability/status views do not advertise a production taste model. The completed-work-package admission verifier derives P2-003's evidence status from the checked ledger state and this task-specific pack; no generated green production claim is hand-edited.