# Stage 74: Deterministic Evaluators (aq74_register_and_run_deterministic_evaluator / DB-UNIT-06)

Persists and executes content-addressed 0-call deterministic rules, conformance checks, SLO measurements, and Kaggle tabular metrics (RMSE, MAE, LogLoss, AUC, F1, NaN/constant guards).

- Atomic micro-queue: `aq74_register_and_run_deterministic_evaluator` (visibility: `public_commons`)
- Backing database unit: `DB-UNIT-06` (`Multi-Pass Enrichment, Packaging & 6+ Locale Core`)
- Database cell: `entities-db.teleon.dev` (`teleon_core.multipass_card_projections`, `teleon_core.harness_adapter_manifests`, `teleon_core.locale_terminology_packs`)
- Claim row hydration: `single_self_contained_flat_task_row` (100% context inline, normalized writeback)
- Track 2 Sister queue (`SISTER-06`): `sq06-improve-packaging-queue.teleon.dev` (Compress L0 token budgets without losing selection accuracy and onboard new languages and harness formats.)
- Lanes: Rules & Schema Checks, Kaggle CSV Metrics, SWHID Eval Records

## Execution engines & AI Agent freedom contract

- **Deterministic (0 model calls)**: `ast_signature_card_compiler_v1` (Zero-shell JSON/CSV/text/SLO/port metric runner & cryptographic attestation signer)
- **Hybrid (<=1 model call)**: `multilingual_card_summarizer_v1`
- **AI Agent & Bot Fleet (1..6 calls)**: `cross_harness_skill_adapter_v1` (Domain metric author & adversarial holdout fixture builder)
  - Contract: **Instructions + Resources + Freedom to search for new resources and new ways not bound by instructions**

## Database invariants

- Exact source bytes, semantic interpretation, and execution authority stay strictly separated.

## External & internal bot fleet actions

- Register reusable deterministic rules, checks, and Kaggle metric functions with SWHID pins
- Reject NaN, Inf, row-count mismatch, ID misalignment, and constant predictions in 0 model calls
- Cache deterministic evaluation records by (evaluator_swhid, candidate_swhid, holdout_swhid)

## Sister self-tuning areas

- Onboard additional languages and technical terminology packs without English bias
- A/B test L0 card schemas (100 vs 150 vs 220 tokens) across small and frontier models
- Add deterministic compilers for emerging agent harness manifest formats