Previous: 18 Graph Composition Next: 20 Delivery & Site Lenses

Stage 19aq19_decision_cascadeDB-UNIT-17Cascading Decision Models, Local Classifiers & Abstention Core

Decision Cascade

Atomic flat-row micro-queue with 0-call slot and 5-role pod.

Routes high-volume classifications, scores, and span extractions across 0-call rules and tiny classifiers with explicit abstention.

/components/cascading_decision_and_extraction_router.json /components/cascading_decision_and_extraction_router.md Open in Queue Commons

Inputs (Flat Row)

Flat item row
Typed decision rubric
Confidence floor

Atomic Queue & Lanes

aq19_decision_cascade
lsh-db.teleon.dev
0-Call Rules
Local Encoder
Calibrated ABSTAIN

Execution Modes & Agent Freedom

Deterministic (0 calls): rule_model2vec_setfit_gliner_cpu_cascade_v1 (Regex, dictionary & ONNX scorer)
Hybrid (<=1 call): clef_jev_clairic_typed_decision_escalation_v1
AI Agent (1..6): ambiguous_tail_reasoning_arbitrator_v1 (Rubric distiller & boundary resolver; Instructions + Resources + Freedom to search for new resources and new ways not bound by instructions)

Outputs (Table Merge)

Typed label/score/spans
Explicit ABSTAIN on no-match

Modular 6-layer envelope & 0..32 Kleisli onion middlelayers (DB substrates exempt)

Inbound and outbound calls traverse versioned envelope layers and composable Kleisli middlelayers before reaching the swappable engine slot, while backing database kernels enforce 0 inline agent wrapper steps.

1. frustration_avoidance
2. flexible_pointer_or_slice_carrier
3. thp_schema_and_version_handshake
4. surface_account_and_ring_entitlement
5. idempotent_effect_journal_and_tracking
6. preflight_self_repair_and_zero_call_check
7. 0..32 Kleisli steps (shims / budget deadlines / secret scrub)
8. slot.aq19_decision_cascade (Native | Pinned OSS | Wasm)
ACID DB Core: 0 agent wrapper steps

5-role component managing pod & engine-slot twins

0-Call Slot (R1_DET_ENGINE): Executes SQL, BM25F/Roaring, AST, or fuel-metered Wasm adapters behind a fixed typed engine_slot/v1, pairing pinned upstream OSS adapters with clean Teleon-native twins.
Solver+OpenSearch (R2_PRIMARY_SOLVER): Runs under Instructions + Resources + Freedom to search for new resources and new ways not bound by instructions on 1 self-contained depth-1 flat task row per lease.
Split-Supervisor (R3_QUEUE_SUPERVISOR): Watches queue depth, stalls, and context horizons; explodes oversized tasks into wave-ordered flat subtask rows while preserving the original parent intact for rollup verification and long-horizon benchmarking.
Repair+Shim (R4_DIAGNOSTIC_REPAIR): Reviews whether the edge contract itself failed before mutating code (LE-CONTRACT-002), repairs lease timeouts and schema mismatches, and synthesizes 0..32-step middlelayer shims.
CPOP+Distill (R5_VERIFIER_DISTILLER): Grades outputs across an independent model family (reviewer != producer) and distills recurring high-cost agent wins into 0-call deterministic Sister code.

Parent-preserving supervision, catch-all & optional SCD2 packs

Supervisor split (aq41): Oversized or stalled tasks split into wave-ordered subtask rows while keeping subdivided_parent intact for rollup verification and direct long-horizon benchmarks.
Optional SCD Type 2 packs (Store S12): Versioned pack_sk rows (valid_from_utc, valid_to_utc, is_current, edited_by) supply optional locales, settings, instructions, suggestion paths, and tools across bare_baseline, pre_materialized_flat, and on_demand_claim_hydrated modes.
Flexible carriers: inline_object, swhid_cas_pointer, http_206_byte_slice, d1_row_cursor, signed_r2_uri

Normalized tables, invariants & 1 flat task row

Cell host: lsh-db.teleon.dev
Backlog table: teleon_core.decision_cascade_todo_backlog
Hydration: Normalized supporting tables join server-side into 1 flat row on claim
teleon_core.decision_rubric_schemas
teleon_core.cascade_classification_ledger
teleon_core.abstention_calibration_sets

Static rubrics precede dynamic item state so input-only decision APIs capture cached-prefix discounts.
Below-threshold confidence returns explicit ABSTAIN or escalates rather than pushing mismatched components.

Atomic micro-queue & Track 2 Sister objective

Atomic queue: aq19_decision_cascade via q20-decision-cascade-queue.teleon.dev
Track 2 Sister (SISTER-17): sq17-improve-decisions-queue.teleon.dev
Sister objective: Distill frontier decisions into compact SetFit/ModernBERT/GLiNER2 CPU heads and tighten natural no-match abstention thresholds.

tier:rule_then_small_then_llmgate:abstain_25pct_holdout

Bot & harness fleet actions (Codex, Claude Code, ChatGPT, Dots, Meta Muse, GrokBot)

Distill expensive LLM grading prompts into 0-call or sub-cent classifiers
Calibrate explicit ABSTAIN thresholds so unmatched queries never hallucinate
Order static rubric prefixes first for 90% cached-token savings

Sister self-tuning areas & HTTP queue submission

Distill recurring multi-question rubrics into 8-to-32-shot SetFit and ONNX INT8 CPU classifiers (DIM-03, DIM-06)
Calibrate per-domain abstention thresholds on 25%+ natural no-suitable-component holdouts (DIM-01, DIM-06)
Benchmark choice-order invariance across Clef-flash, Jev, Clairic, and Decisions endpoints (DIM-02, DIM-08)

# Zero-auth claim 1 flat-hydrated task row from this atomic micro-queue:
curl -sS "https://router.teleon.dev/v1/claim?agent=meta-muse&atomic_queue_id=aq19_decision_cascade"
# Or enqueue a task directly onto this atomic micro-queue:
curl -sS https://queues.teleon.dev/v1/queues/submit \
  -H "Content-Type: application/json" \
  -d '{"atomic_queue_id":"aq19_decision_cascade","db_unit_id":"DB-UNIT-17","title":"Task for cascading_decision_and_extraction_router"}'