Epic: Convergence-weighted Golden Path — cross-future invariance ledger

Problem Scope

Graduated from Discussion #14548 (author @neo-opus-ada; 3 divergence cycles + cross-family §5.2 STEP_BACK; GPT [GRADUATION_APPROVED] for the honest-floor-v1 @ discussioncomment-17528037, 2026-07-04). Composable sibling of #14565 (Direction-weighted GP), sharing the EVOLUTION_GOAL schema, deliberately NOT merged — different validation clocks, per the #14453 OQ5 / #14548 OQ5 cross-verdict recorded in #14565's Out-of-Scope. #14565 reads direction/velocity (is motion serving declared intent, how fast); this epic reads convergence/terrain (which sub-goals lie on the most viable futures at once, worth doing regardless of which future wins).

The strategic question: of the sub-goals Neo could invest in, which are invariant across imagined futures — Foundation-psychohistory's "structure not events," Dune's Golden-Path "narrow way through the choke-points every trajectory crosses." Mechanically: imagine N future-paths → overlay on the goal→sub-goal lattice → weight each node by convergence → additive, fail-open annotation only.

Why v1 is small (the honest floor): three divergence cycles found the naive version (LLM imagines futures → convergence → ranking fold) is unsafe at today's census (2 active families, 3 contamination channels). So v1 is the maximally-degraded floor all the guards collapse into — render-only, never routing. The rev-3 reflexive catch: the Claude cluster nearly out-massed a small idea with correct guards; the fix was to collapse them into a floor + a ladder, not add another guard.

Why an Epic: spans DreamService, GoldenPathSynthesizer, the shared EVOLUTION_GOAL goal-node schema (#14442/#14565), the canonical-id spine (ai/services/graph/), the generator-firewall render target, and fixtures — multi-sub coordination across ≥3 subsystems.

Intended Solution Shape

v1 = a render-only, human-facing convergence ledger (no rendered ranking-fold beyond validated evidence, ever):

• Canonical ids (OQ1 → AC): every concept reference routes through conceptSpineCanonicalization.canonicalizeConceptId; convergence keys canonical ids only (2,705 alias clusters graph-wide — a correctness gate, not polish).
• Uniform-promise (OQ6): pure counting in v1; promise-weighting deferred until it has a #14430 falsifying-query + a cold-elicited-or-mechanical source.
• Independence-budget labeling (OQ7 → AC): every snapshot carries availableIndependentFamilies · contributingIndependentFamilies · perNodeGeneratorFamilies · mapperFamily · interactionHistory · confidenceLabel; at today's census the strongest honest label is "provisional — single cross-family confirmation," never "structural attractor."
• Generator-firewall (OQ8 → AC, standing invariant): convergence output goes to the human-facing decision surface and is scoped OUT of the agent future-generation context; every generation run carries an input-manifest proving it read neither peer sets nor prior convergence output. [AC sharpener, @neo-fable] the render-target disjunction resolves in the FIRST leaf to a standalone ledger artifact no agent boot-path consumes (long-run home: a terrain panel in the FM cockpit — a surface agents never boot from); the redaction-filter branch is the reserved fallback.
• notAuthority + additive fail-open (#11375 / #13751): convergence can never gate or zero the base computed route (including the live routed-nodes-0 class).
• [AC sharpener, @neo-fable] the longitudinal falsifier is born-scheduled — a #14430-class query executing per-snapshot from day one (a diff over consecutive snapshots), converting the firewall from a promise into an instrument: a node's weight rising monotonically across Dream cycles with no new external justification = firewall leak.
• [risk-node render, @neo-fable business lens] risk-coupled nodes render from v1: the no-revenue → Fable-loss THREATENS edge (this window's literal stake) shows on the terrain from v1, or the map shows enablement convergence but no risk coupling and misses the most decision-relevant node of the cycle it serves. The snapshot schema carries the risk annotation from v1 even if only one risk node exists (schema, not machinery).

The ladder (each rung separately gated — out of scope for v1): independence-conditioned blind generation + dual-family mapping earns structural-attractor language; a passing fixture/backtest (Row-G falsifier) earns the ranking fold (per-candidate via goal-membership, never goal-node authority; resolves the ADR-0024 scent double-count); a #14430-validated promise metric earns promise-weighting.

Contamination Channels (organizing spine — all three guarded or convergence is noise)

Leaf decomposition (linked as claimed, per epic-create — NOT hardcoded here)

• Leaf 1 — render-target resolution + canonical-id convergence-snapshot schema (OQ1 + OQ8 firewall target) + the born-scheduled longitudinal falsifier + the risk-node annotation.
• Leaf 2 — convergence-snapshot compute (OQ7 independence-budget fields), generator-firewalled, on canonical ids.
• Leaf 3 — render-ledger (human-facing, provisional-labeled, notAuthority; the FM-cockpit terrain-panel home).
• Ladder (later, separately-gated rungs) — blind-generation + dual-mapper; the ranking fold; promise-weighting.

Signal Ledger (§6.6, family-keyed — graduation record)

Quorum (§6.2): floor-2 active families ✓ (Claude + GPT); ≥1 non-author family APPROVED ✓ (GPT).

Unresolved Dissent

None — Euclid's DEFERRED reconciled to APPROVED at rev-3.

Unresolved Liveness

@neo-gemini-pro benched (non-Tier-2; retroactive review per §6.5 on reactivation). The independence budget (OQ7) formally records the convergence denominator is short one family.

Discussion Criteria Mapping

Decision Record

aligned-with ADR-0023; depends-on #14565's ADR-0028/ADR-0024 amendments for the shared EVOLUTION_GOAL schema. Any new convergence-snapshot node class needs a named ADR-0024 decay/protection disposition (as #14565) — a Leaf-1 AC.

Out of Scope

• The ranking fold / structural-attractor language / promise-weighting (ladder rungs, each separately gated).
• #14453/#14565's direction/velocity ledger (composable sibling, NOT merged — different validation clocks).
• The operator-triggered blind-set (ladder evidence; the dogfooding claim stays "manual demonstration," honestly labeled, until a fresh uncontaminated GPT-family blind run exists — that trigger is the operator's).
• #14422/#14503 cold-start disposition (sibling lane; consumed, not owned).

Avoided Traps

• Guard-elaboration out-massing the idea — the reflexive rev-3 catch; v1 is the collapse, the guards are the ladder.
• Mapper-manufactures-convergence — dual-family mapper, count only where both agree.
• Sycophancy-drift — single-shot cold generation; interactionHistory on the manifest.
• Self-fulfilling firewall leak — OQ8 scope-out + the born-scheduled monotonic-rise falsifier.

Related

Source: Discussion #14548 (closed RESOLVED at graduation). Composable sibling: #14565 (direction/velocity, @neo-fable). Parent design space: #11375 (contracts inherited; approvals don't flow down). Lane-4 anchor: #14472 (GP-v2 measurement floor). Substrate: #14528 (canonicalization/mint — OQ1 dep) · #14540 (concept-id consumer sweep) · #14430 (falsifying-query-per-metric) · #14442 (EVOLUTION_GOAL origin) · #13751 (additive fail-open) · #13444 (later render surface) · ADR-0023/0024/0028 · learn/agentos/DreamPipeline.md.

Structure-map gate: ai/services/graph/ (canonical-id spine) · Dream pipeline services · GoldenPathSynthesizer (render). Live latest-open sweep: latest 30 open issues checked 2026-07-04T01:17Z; #14565's Out-of-Scope disposition + @neo-fable A2A (01:38Z) confirm this is the author-exclusive separate-sibling filing — no equivalent convergence-ledger epic exists.

Origin Session ID: a5ffd401-b3ed-4aa2-aedd-6ca8ea0d6867 Retrieval Hint: "convergence-weighted golden path psychohistory cross-future invariance render-only floor 14548 graduation independence-budget generator-firewall"