Skip to content

trance-matrix.md

Latest commit

 

History

History
104 lines (75 loc) · 7.19 KB

trance-matrix.md

File metadata and controls

104 lines (75 loc) · 7.19 KB
Raw

Cortex Deep — TRANCE Tier Matrix

Load this file when evaluating or authorizing an TRANCE escalation. See EIDOLONS.md for the always-loaded routing cortex; the Activation Gates table below was relocated here from EIDOLONS.md per R-021/R-022 (generalist-eidolon P2 cortex re-fit) — it was mislabeled "(always-loaded)" while not actually needed for single-Eidolon dispatch.

Activation Gates

TRANCE grants: parallel fan-out (max 5 branches), worktree isolation per branch, verifier-cascade wrapping, evaluator-optimizer loop (cap 3 iterations), model-tier upgrade (lead = deep, workers = light).

TRANCE is never the default. Auto-trigger requires both a complexity flag AND a stakes flag. Cost warning emitted at ≥ 5× standard-tier budget.

Gate Eidolon Condition
G1 — Discovery scatter ATLAS Surface > 25 files OR > 5 modules → scatter sub-agents per module, aggregate via Abstract phase
G2 — Hard-decision consistency FORGE ≥ 3 plausible alternatives AND (high-stakes flag OR explicit TRANCE token) → N=3 reasoning traces, majority-vote
G3 — Spec evaluator-optimizer RAMZA Complexity ≥ 7/12 AND (high-stakes OR ambiguous reqs) → generator + evaluator, max 3 iterations
G4 — Parallel implementation Vivi RAMZA emitted > 1 independent story AND budget bounded → one Vivi per track, worktree isolation
G5 — Doc parallel synthesis IDG Large source artifact set AND topological order allows parallelism → per-section parallel, CHT per section, one-revision cap preserved
G6 — Forensic counterfactuals VIGIL ≥ 2 plausible root-cause hypotheses AND bisect surface allows independent testing → parallel hypothesis tests on isolated bisects

TRANCE refusals (immutable): A refused capability does not become available at TRANCE. ATLAS still does not write. RAMZA still does not implement. IDG still does not retrieve. FORGE still does not tool-call. VIGIL still does not auto-apply patches. Per-Eidolon retry budgets remain enforced inside TRANCE. Gilgamesh (generalist, fallback-only) has no TRANCE form — it dispatches only at standard tier via Step-2(a); TRANCE escalation never applies to a Step-2(a) fallthrough.

Per-Eidolon TRANCE Capability Matrix

Eidolon TRANCE form Granted Forbidden at TRANCE
ATLAS Scatter sub-agents per module Parallel fan-out (G1); worktree isolation; Abstract-phase aggregation Writing, editing — D2 refusal stands
SPECTRA Evaluator-optimizer loop on draft spec Generator + evaluator + termination gate (G3); cap 3 iterations Implementing code; > 3 cycles
APIVR-Δ Parallel feature branches in worktrees + verifier cascade Multi-track implementation (G4); per-track verifier; reflection memory bounded ≤ 3 retries Re-attempting beyond category budget; writing in shared tree without isolation: worktree
IDG Parallel doc-section synthesis Per-section parallelism with topological respect (G5); CHT verification per section Retrieval; > 1 revision per section
FORGE Self-consistency on reasoning chains N=3 (or N=5 high-stakes) sampled traces with majority-vote / judge-merge (G2) Tool calls, retrieval, code emission; debate without heterogeneity
VIGIL Parallel hypothesis testing on isolated bisects Counterfactual fan-out on worktrees (G6); 5-intervention budget preserved Auto-apply patches; writing outside verified-patch.diff

Model Tiers

The tier ladder light < standard < deep assigns a vendor-neutral capability level to each Eidolon. Concrete model strings live exclusively in roster/model-profiles.yaml (never in this file or EIDOLONS.md — PD #162).

FORGE-derived per-class defaults:

Eidolon Suggested tier Rationale
SPECTRA deep Planner; spec quality gates require full reasoning
FORGE deep Deliberation; self-consistency via N=3 traces
VIGIL deep Forensic; counterfactual chains require depth
ATLAS standard Scout; deterministic retrieval first, LLM second
APIVR-Δ standard Coder; loop_native:false — deep is benchmark-gated
IDG light Scriber; synthesis from provided context
Kupo light Executor; localized micro-tasks

Resolution precedence (most-specific wins): per-member PIN → per-tier calibration → active profile → roster suggested_tier → class default → nexus default_profile. Use eidolons model show to inspect resolved values; eidolons model use to override.

Cost Ceiling Rules

Rule Value Rationale
C1 — Max parallel branches 5 Empirical sweet spot for orchestrator-workers
C2 — Max model-tier upgrade lead = deep, workers = light Mirrors research-system topology; D9
C3 — Max wall-clock Host-enforced; cortex emits wall_clock_budget_seconds hint Cortex never spins indefinite background jobs without user opt-in
C4 — Token budget warning Emit [DECISION] at ≥ 5× standard-tier budget 15× lift number is upper bound, not default
C5 — Surface-size threshold > 25 files OR > 5 modules = large surface Heuristic; configurable
C6 — Auto-trigger requirement Both complexity flag AND stakes flag must hold Either alone keeps cortex at standard tier

Refusal Gates (TRANCE MUST NOT override these)

  • R1 — Refused capabilities remain refused at TRANCE. ATLAS still does not write. SPECTRA still does not implement. IDG still does not retrieve. FORGE still does not tool-call. VIGIL still does not auto-apply patches.
  • R2 — FORGE parallel fan-out is reasoning-only. FORGE's TRANCE is self-consistency on reasoning chains; it does not gain tool access.
  • R3 — Bounded budgets are inviolable. Per-Eidolon retry budgets remain enforced inside TRANCE. TRANCE adds parallelism, not a fresh budget.
  • R4 — D5 unbounded reflection forbidden. TRANCE may not extend reflection loops past published caps (SPECTRA 3 cycles, IDG 1 revision, APIVR ≤ 3 same-category attempts).
  • R5 — Partial-team deployment degrades gracefully. If a needed Eidolon is not installed (eidolons.lock check), TRANCE is degraded: cortex emits a [GAP] and a fallback chain rather than spawning fan-out into a member that doesn't exist.

Confidence Thresholds

Threshold Value Source
τ_standard (min to dispatch) 0.6 Hybrid-LLM dial behavior [D §3.2]
τ_trance (min to auto-escalate) 0.8 RouteLLM preference data [D §2 #1]
FORGE consensus floor 60% Self-Consistency threshold [D §3.6]

[OQ-2] Assumption: τ values are reasonable defaults until calibration experiment shows optimal point is far off. Tune from routing-decision logs.

Open Questions Carried Forward

ID Assumption Mitigation
OQ-1 LLM-self-routing against descriptors is good enough for v1. Migrate to calibrated classifier (RouteLLM) when mis-routing > 10% in telemetry.
OQ-4 TRANCE-on-demand is the right v1 stance. Add TRANCE-default flag in eidolons.yaml; cortex still emits [DECISION] at C4 ceilings.
OQ-7 Self-consistency at N=3/N=5 is right for FORGE-TRANCE. Tune from logs; drop N back to 3 or standard if empirical evidence disagrees.