AGENT SKILL

forge-self-consistency

Governs N perspective-diverse, mutually-blind reasoning traces with structural-agreement judge-merge for gated high-stakes FORGE decisions (G2/TRANCE). Use instead of forge-deliberation when Deep depth AND a stakes flag both hold (Framing score 8–9 with irreversible/system-wide blast radius), or on explicit opt-in. Do NOT use for standard-tier decisions — self-consistency is never the default.

License 许可
Apache-2.0
所属仓库
Rynaro/magicite
最近更新
2026-10-14
Stars
★ 2
Forks
0

01 前言 Front Matter

name
forge-self-consistency
metadata
{ "phase": "R-variant", "methodology": "FORGE" }
description
Governs N perspective-diverse, mutually-blind reasoning traces with structural-agreement judge-merge for gated high-stakes FORGE decisions (G2/TRANCE). Use instead of forge-deliberation when Deep depth AND a stakes flag both hold (Framing score 8–9 with irreversible/system-wide blast radius), or on explicit opt-in. Do NOT use for standard-tier decisions — self-consistency is never the default.

03 SKILL.md 正文

Self-Consistency Mode

Loaded instead of skills/deliberation/SKILL.md when a G2 / TRANCE self-consistency escalation fires. Governs how the Reasoner runs N perspective-diverse, mutually-blind reasoning traces over a frozen Frame+Observe inventory and merges them on structural agreement, not on verbalized confidence.

When to use

Load this skill instead of forge-deliberation when BOTH a complexity flag AND a stakes flag hold: Deep depth (Framing Step-4 total 8–9) with an irreversible or system-wide blast radius. Also loads on explicit opt-in ("run self-consistency", "G2", "N=5 on this"). Use N=3 for standard high-stakes runs and N=5 for the worst-case quadrant (irreversible AND system-wide). Do not enter this mode for standard-depth decisions; do not combine with forge-deliberation in the same Reason phase.

This mode is gated, never default (see "Gate" below). Standard-tier FORGE stays single-trace (skills/deliberation/SKILL.md, 1–3 passes). Self-consistency is the TRANCE form the trance-matrix grants FORGE: "N=3 (or N=5 high-stakes) sampled traces with majority-vote / judge-merge (G2)" — operationalized here.


Gate — when this mode fires (and when it must not)

Self-consistency fires ONLY when both a complexity flag and a stakes flag hold, OR on explicit opt-in, OR when the roster declares a weak-host degraded mode (see item 3 below):

  1. Deep depth — Framing Step-4 total 8–9 (high Ambiguity AND Irreversibility OR system-wide Blast radius). A Deep score on its own is necessary but the stakes flag (irreversible / system-wide) must also be live. This is the trance-matrix C6 auto-trigger requirement: a complexity flag AND a stakes flag.
  2. Explicit opt-in — the user or the parent orchestrator names the mode ("run self-consistency", "G2", "N=5 on this").
  3. Weak or undeclared host — the roster declares FORGE degraded_mode: sample-select (roster/routing.yaml, routing-1.1): on a weak or undeclared host this mode REPLACES single-trace self-red-teaming (the Pass-3 Deep red-team in skills/deliberation/SKILL.md) rather than only gating on Deep + stakes — the roster data itself is the trigger, because sampling+selection is more reliable than unbounded single-trace self-correction on a weak host.

If none of the above hold, do not enter this mode — run standard single-trace FORGE. Forcing G2 universally violates the opt-in contract (FORGE is never a mandatory critique gate). This mode adds parallelism, not a fresh budget: it composes with — does not extend — the existing 3-pass / 1-REFORGE cap.

No tool access is gained at TRANCE. FORGE is reasoning-only at every tier. The independent-verifier hook below is a handoff recommendation, not an executed call.


Step 1 — Freeze the shared inventory (fan-out precondition)

Before fan-out, the Frame and Observe outputs are frozen: the decision question, decision type, constraint table, success criteria, and the evidence inventory (with H/M/L reliability tiers). Every trace reasons from this same frozen inventory. Traces differ in reasoning stance, never in the evidence they are given — a trace that invents new evidence is invalid and is discarded.

Pick N:

N When
N=3 Standard high-stakes (Deep depth + stakes flag).
N=5 Irreversible AND system-wide blast radius (the worst-case quadrant).

N is fixed before fan-out and bounded — exactly N traces, no re-sampling beyond N. If the merge is below floor, you emit [DISPUTED]; you do not draw more traces.


Step 2 — Fan-out: N perspective-diverse, mutually-blind traces

Diversity is across reasoning stance, not model (FORGE has no model control; intra-model sample diversity already beats naive cross-model debate). Each trace is assigned a distinct adversarial lens from this fixed persona table, in order, until N is reached:

# Persona Lens — the question this trace asks first
1 Evidence-maximizing Which hypothesis is best supported by the H-reliability evidence, weighing alignment above all?
2 Pre-mortem / failure-first Assume each hypothesis was chosen and failed — which fails least catastrophically?
3 Constraint-relaxation If exactly one soft constraint could be relaxed, which hypothesis becomes clearly best, and is that relaxation acceptable?
4 Inversion / steelman-the-rejected Build the strongest case for the hypothesis the obvious reading rejects — does it survive?
5 Simplest-viable-default What is the lowest-complexity, most-reversible option that still satisfies all hard constraints?

For N=3 use personas 1–3. For N=5 use 1–5.

Mutual blindness is mandatory. No trace may read another trace's verdict, score, or selected hypothesis before the merge. This prevents self-conditioning (a trace anchoring on an earlier trace's answer collapses the diversity the mode exists to create — the lazy-agent failure).

Two equivalent execution forms:

  • Parent-dispatched subagents (mechanically stronger). The orchestrator dispatches N clean-context FORGE subagents, each given the frozen inventory plus one persona, with no shared scratchpad. Independence is enforced by separate contexts. Prefer this form when available.
  • Sequential-blind in one context (fallback). Emit N reasoning passes one after another, each opening from the frozen inventory under its assigned persona, and do not reference any prior pass's verdict while reasoning the current one. Only the merge step (Step 3) may read all N.

Each trace produces the standard hypothesis-level work (≥3 hypotheses, the four stress tests, a selected hypothesis, and any [RISK] / [REVERSAL-CONDITION] markers it surfaces). A trace MAY state a per-trace confidence, but that number is not used by the merge.


Step 3 — Aggregation: structural-agreement judge-merge

The merge is a single deterministic tally, not another round of deliberation. There is exactly one merge pass — no merge loop.

  1. Tally hypothesis selection. For each trace, record which hypothesis it selected. Identify the modal hypothesis (the one selected by the most traces).

  2. Compute the structural-agreement score:

    structural_agreement = (traces selecting the modal hypothesis) / N
    
  3. Cross-reference recurring markers. Collect every [REVERSAL-CONDITION] and [RISK] that ≥2 traces surfaced independently. Independently-recurring reversal conditions are high-trust signals — they were found by differently-biased reasoners, so they are unlikely to be persona artefacts.

  4. Decide against the 60% consensus floor (trance-matrix FORGE consensus floor):

    structural_agreement Result
    ≥ 60% PASS — emit one merged [VERDICT] on the modal hypothesis.
    < 60% FAIL — emit [DISPUTED] (see Step 4).

Merged confidence anchors on structural agreement, NOT on averaged verbalized confidence. The merged verdict's confidence IS the structural_agreement score (optionally tightened upward only by the count of independently-recurring reversal conditions). Do not average the per-trace confidence numbers: verbalized confidence is systematically overconfident and a poor diagnostic of correctness, so it is excluded from the trust anchor. The mechanical signal — how many independently-biased traces converged — is the calibrated one.

A PASS verdict lists: the modal hypothesis, the structural-agreement score as its confidence, and the convergent (≥2-trace) reversal conditions as the verdict's [REVERSAL-CONDITION] set.


Step 4 — Disagreement: [DISPUTED] below the floor

If structural_agreement < 60%, never force-merge into a single winner. Emit a [DISPUTED] verdict that:

  • enumerates the live positions (each hypothesis that ≥1 trace selected, with which persona(s) backed it),
  • states a reduced confidence equal to the structural-agreement score,
  • lists which reversal conditions recurred and which were trace-local, and
  • routes a handoff: → human (genuine judgment split) or, if a tiebreak is feasible, → independent verifier (see Step 5).

This mirrors the single-trace tied-score rule: when the top positions are genuinely close, the verdict acknowledges ambiguity rather than manufacturing a winner.


Step 5 — Independent-verifier hook (opt-in ceiling-breaker)

Self-consistency on a single reasoner saturates around a known accuracy ceiling; the lever that pushes past it is an independent / cross-model verifier. FORGE is tool-less (P0) and cannot dispatch a second model itself, so this is a documented handoff, not an executed call:

  • When to offer it. A near-floor PASS (structural_agreement just above 60%), a [DISPUTED] FAIL on a high-stakes/irreversible decision, or explicit opt-in.
  • What to emit. An [ACTION] / handoff recommending that the parent orchestrator (or nexus eidolons run) re-derive the verdict via an independent FORGE instance or a cross-model verifier, seeded with the same frozen inventory but not shown this run's merged verdict. Name it explicitly as → independent-verifier in the handoff section.
  • What NOT to do. Do not call a tool. Do not start a debate loop between traces. The verifier is a separate, blind re-derivation orchestrated outside FORGE — that independence is exactly what makes it a ceiling-breaker.

Bounds & honesty

  • Exactly N traces, one merge pass. No re-sampling, no merge loop. Composes with the 3-pass / 1-REFORGE cap; does not extend it.
  • Diversity is tuned, not naive. Fixed adversarial personas are the tuned form of multi-trace reasoning; naive same-prompt sampling or unbounded debate is excluded (see DESIGN-RATIONALE.md §9).
  • Confidence is structural. It reflects cross-trace agreement, never an averaged self-grade. This is a methodology-layer estimate, not a benchmarked accuracy figure.

Reasoner — Self-Consistency Skill

04 同仓库更多 Skills

Rynaro
Rynaro/magicite

wine-dxvk-cache-clear

Clear a stale DXVK shader cache that is causing graphical corruption in a Wine or Proton game. Use when: a game shows corrupted textures or shader flicker that started after a driver or DXVK version change.

★ 2
Rynaro
Rynaro/magicite

steam-download-region-fix

Fix a Steam client stuck at very low download speed by changing its download region. Use when: Steam downloads are far below the local internet speed for every game, not just one.

★ 2
Rynaro
Rynaro/magicite

proton-battleye-eac-toggle

Toggle BattlEye or Easy Anti-Cheat runtime support for a Steam Proton game. Use when: a multiplayer game refuses to launch with an anti-cheat error under Proton.

★ 2
Rynaro
Rynaro/magicite

vivi-verify-incoming

Load when reading any upstream artefact handed off to Vivi that carries a sibling .envelope.json. BLOCKING per ECL §6.2.2: the orchestrator MUST have verified the envelope's SHA-256 (eidolons verify-envelope --block / eidolons run --verify) and recorded a verify_pass before dispatch. If no verify_pass exists for the message_id, or a verify_fail is present, REFUSE to process the payload and hand back to the orchestrator. Symmetric receiver gate — every Eidolon enforces it identically.

★ 2
Rynaro
Rynaro/magicite

vivi-parallel-tracks

Load ONLY under TRANCE authorization when the Plan phase yields N independent, non-overlapping implementation tracks (disjoint file sets). Operationalizes the TRANCE G4 form — parallel multi-track implementation in mandatory git-worktree isolation, max 5 clean-context tracks, a per-track verifier cascade, a non-fungible per-track ≤3 reflection budget, explicit stop conditions, and a single-threaded merge/aggregation step. Absent TRANCE gating, Vivi runs exactly as today: single-track A→P→I→V→Δ/R.

★ 2
Rynaro
Rynaro/magicite

vivi-methodology

Full Vivi cycle reference — A(nalyze) → P(lan) → I(mplement) → V(erify) → Δ(Delta)/R(eflect). Use when the task is a non-trivial feature implementation in a brownfield codebase and you need the complete methodology reference: complexity routing, evidence-grounded planning rules, test-anchoring requirements, and failure-escalation thresholds.

★ 2