[router] Phase 0 Track 3: programmatic baseline (the teacher)

[dp-web4]

Track

Phase 0, Track 3 — programmatic baseline (per shared-context/arc-agi-3/phase2/brain-arch/router-sprint-1-phase-0.md)

What this ships

• sage/cognition/router/baseline.py — programmatic_decide(RouterInput, plugin_registry) -> RouterOutput, a pure-function reconstruction of the existing _select_attention_targets + _get_plugins_for_modality logic in sage_consciousness.py
• Helper functions (all spec-required, all exported): _decide_action_type, _select_plugin, _should_use_habit, _compute_rationale
• Constants mirroring dispatcher: MODALITY_MAP, NOOP_METABOLIC_STATES, HABIT_CONFIDENCE_THRESHOLD, SNARC_HIGH_NOVELTY_THRESHOLD
• sage/cognition/router/tests/test_baseline.py — 50 tests (branch coverage, rationale vocabulary coverage, edge cases, purity, determinism, dispatcher parity, latency)

What this does NOT ship (deferred)

• Feature extraction (Track 2) — RouterInput construction from live kernel state
• Consciousness-loop shadow-mode integration (Track 5) — wiring the baseline into sage_consciousness.py at step 5
• Outcome tracking (Track 6) — the baseline is decision-only; outcome backfill is later

Tests added

50 new tests, covering:

• Top-level branches: dream -> noop, wake+vision -> invoke, habit branch (above/at/below threshold), metacog blocking (partial + total), every modality in MODALITY_MAP
• Rationale vocabulary: every value in VALID_RATIONALE_CODES reachable (low_atp_rest, habit_match, metacog_blocked, high_novelty, goal_driven, escalate_frontal, federate_peer, reflex, default)
• Output validity over a 120-combination SNARC × habit × metacog × metabolic cross
• Purity: router_input and plugin_registry unchanged after call (deep copy check)
• Determinism: same input -> same output
• Edge cases: empty WM, zero SNARC, ATP=0, empty registry, registry missing tier, unknown modality
• Latency: p50=2.6µs, p99=9.9µs (budget is <1ms)
• Dispatcher parity: every modality mapping reproduced

Test results

115 passed in 5.47s  (50 new Track 3 + 65 existing Track 1 + Track 4 — combined suite green)

Benchmark (50k iterations, varied corpus):

p50=2.64us   p90=5.50us   p99=9.90us   p99.9=31.13us   max=113us

Acceptance criteria from sprint doc

• Baseline reproduces existing dispatcher behavior on golden inputs (7 test_dispatcher_parity_* tests)
• <1ms per decision (p99 ~10µs, ~100x headroom)
• Pure function — no side effects on the real consciousness loop (2 purity tests verifying no mutation of input or registry)
• Combined router test suite remains green (115/115)

Reviewer notes

Branch → rationale_code map

Dispatcher branch that fired	rationale_code
metabolic_state == 'dream' (max_active_plugins=0)	low_atp_rest
Candidates exist but all in metacog block_list	metacog_blocked
No candidates at all (empty modalities / all unknown / time)	default
habit_available & habit_confidence >= 0.85	habit_match
plugin.tier == frontal_lobe	escalate_frontal
plugin.tier == federate	federate_peer
plugin.tier == reflex	reflex
Routine/specialized + snarc_novelty >= 0.6 AND snarc_arousal >= 0.6	high_novelty
Routine/specialized + wm_goal_active + perception modality	goal_driven
Fallthrough invoke	default

Gaps where dispatcher doesn't cleanly map to RouterOutput (for Track 5)

1. List → single decision: the existing dispatcher selects up to max_active_plugins targets per tick (e.g. 3 in WAKE). RouterOutput is one decision. The baseline takes the top-salience first-plugin — consistent with "what would the dispatcher do first" but loses the "plus these others" tail. Track 5 will need to decide whether the router dispatches once per tick (as the PRD §2.10 suggests) or the harness calls it multiple times.

2. Habit branch is new: the existing _select_attention_targets has no habit short-circuit — cerebellum lookup lives elsewhere in the loop. RouterInput.habit_available / habit_confidence (populated by Track 2) let the baseline honor it, so this is a forward-looking addition. Track 5 should confirm the arbitration order is acceptable (PRD §14 proposes habit wins at confidence ≥ 0.85; that's what we use).

3. habit_id placeholder: RouterInput schema §3.1 doesn't carry a habit_id field. The baseline emits wm_state_key as the habit_id (deterministic, unique per WM state, matches what cerebellum.lookup(wm.stable_key(goal_id)) uses). Track 5 should either (a) confirm this proxy is acceptable or (b) add habit_id to RouterInput in a schema-version bump.

4. rest / crisis do NOT short-circuit to noop in the baseline — they allow max_active_plugins=1 per metabolic_controller.py. Only dream forces noop. This is deliberate and matches the dispatcher; Track 7's per-machine canary should verify that rest-state invokes behave as expected.

Missing rationale codes

None needed — VALID_RATIONALE_CODES already covers every branch the baseline uses. All 9 PRD-exemplar codes plus the two fallback codes (default, coerced_noop) are sufficient.

Performance benchmark

50k iterations, varied-corpus: p50=2.6µs, p99=9.9µs, p99.9=31µs, max=113µs. 100x headroom vs sprint's 1ms budget. Enables Track 5 shadow-mode to call the baseline on every tick without measurable regression.

Invariants preserved

• No torch dependency.
• All outputs JSON-serializable via RouterOutput.to_dict().
• No exceptions escape; PRD §3.3 rule-5 coercion is the pure side of the contract — the caller (Track 5) emits the warn event.
• Determinism: no wall-clock reads, no RNG, no dict-ordering dependencies.

Dependencies

• Track 1 (schemas) — merged in main as a1b7c1db3.
• Track 4 (dataset writer) — merged in main as 224429a89 (parallel, not a hard dependency).

Track 3 does NOT depend on Track 2 (feature extraction) — the baseline consumes RouterInput regardless of how it was built.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com

🤖 Generated with Claude Code