AI8 experimental branch · 2026-08-16

Preparation / Genesis Plane

AI8 currently selects what to investigate and which mechanisms to use. This branch makes another object explicit and testable: how a fresh worker—or a temporary research team—is prepared before research begins.

EXPERIMENTAL · not adoptedv0.4.1 R2 fail-closed execution validatorFrozen ArenaLoop v0.2.2 unchanged

← CRP hub · ArenaLoop / Bounded-Ridge · AI8 Architecture · Components

What changes

verified AI8 state / task contract ↓ Preparation Compiler / Team Architect ↓ frozen, audited preparation artifact ↓ fresh isolated worker(s) ↓ OUTER / INNER research and cheapest differentiating tests ↓ Arena result + cost + leakage audit ↓ META learns which preparation policy is worth using
Important: this is a Preparation / Genesis plane, not a fourth top-level loop. It spans AIM³/RHP, OUTER and META.

Modes

D0

Prepared Single · General

Parent prepares a researcher before task identity is revealed.

Stage 0

B0 / B1 / C1 / D1

Cold, raw-context, human-prepared and AI-prepared task-aware singles are compared under the same held-out contract. C1 is now explicitly aligned across protocol, schema and validator.

Team stage

D2 / P2_MATCHED

D2 coordinates a functionally diverse prepared team. P2_MATCHED uses the same prepared workers without communication and redirects equivalent budget into independent work.

What v0.4.1 R2 hardened

Complete immutable chain

Contract → commitment → request/trace/capsule/readiness → runtime/blueprint → sealed Round 1 → review/tension/synthesis → proposal/result → bridge adjudication → cost → blind evaluation → reveal → final run record.

Mandatory schemas

Every claim-bearing JSON is checked against an SHA3-pinned Draft 2020-12 schema. Missing jsonschema is a hard failure.

Direct evidence and hashes

Exact-byte sidecars, internal canonical self-hashes, task commitments, capsule/output bytes, raw evidence and evaluator rules are recomputed.

Causal decomposition

Preparation, diversity and coordination remain separate. Both D2 and P2_MATCHED must pass complete validation before comparison.

Team-only emergence

CROSS_MEMBER_BRIDGE requires sealed multi-source lineage, downstream testing and independent adjudication.

Real fail-closed suite

Three packaged valid fixtures and 20 integration tests invoke the actual claim validator; sixteen deliberate corruptions must be rejected.

Why this matters for Bounded-Ridge

A prepared worker may merely reach a better point on the same abstraction ridge—or may become more likely to challenge and rebuild the representation. D2 adds a second hypothesis: differently prepared workers may cover different ridges and create a valid bridge only after controlled interaction.

No solution is assumed. If Prepared Single, Prepared Team and Team Genesis fail matched held-out representation-shift tests, the Bounded-Ridge objection survives.

Public claim boundary

This page documents an experiment-ready candidate mechanism, not a validated capability. It does not establish AI self-improvement, team superiority, autonomous representation invention, AGI, consciousness, or a CLV result.

The canonical CLV v0.4.2 documents remain unchanged. Preparation / Genesis is linked here because it may become a useful experimental instrument for testing representation reopening and worker/team configuration under matched budgets.

Promotion boundary

Current status: validator-hardened and experiment-ready, not evidence-backed.

Promotion into an ArenaLoop candidate architecture should require replicated, cost-adjusted downstream gain and independently adjudicated team-only bridge evidence on more than one held-out task family. Until then, Architecture Spec v0.2.2 remains immutable.

Source package:

AI8_ArenaLoop_D_RRP_TEAM_GENESIS_v0_4_1_EXECUTION_HARDENED_R2.zip

Source package SHA3-256:

a0f3137a526c6e71bcd6c994f1b5fc894c20720ce9bed6f8c8e0a6db6592d6b9