What changes
Modes
Prepared Single · General
Parent prepares a researcher before task identity is revealed.
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.
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.
Bridge to Arena Foundry
Preparation / Genesis asks how workers are prepared before research begins. Arena Foundry asks how an entire research seed is preserved, built, tested, criticized and mutated across generations. The two are complementary experimental layers; neither changes the frozen ArenaLoop v0.2.2 architecture.
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:
Source package SHA3-256: