Skip to content

Latest commit

 

History

History
65 lines (54 loc) · 3.33 KB

File metadata and controls

65 lines (54 loc) · 3.33 KB

Phase 9 (designed, not yet run): equilibration at depth, the final test

Status: DESIGNED 2026-07-03, deliberately deferred. This document is the pre-commitment: when this experiment runs, its verdict on equilibration is FINAL. No post-hoc caveats. Depth is the last untested variable.

Why this experiment exists

Phase 8 (docs/phase8_prereg.md, docs/phase8_results.md) measured equilibrated standing vs vote-count standing on the graph that today's pipeline actually produces: a depth-1 forest (one node per work unit, exactly 3 judge claims each, single epoch, no disputes). Result: +0.127, real (CI excludes 0) but below the pre-registered 0.25 bar → equilibration demoted to experimental kernel.

The standing objection: equilibration's distinctive machinery (propagation through deep contested trees, cross-claim coupling) had almost no structure to work with. Phase 8 measured its floor. This experiment measures it on the graph the architecture was designed for.

Design: hand-built depth (no live adversarial motions needed)

The substrate does not care whether debate structure arrived from live agents or was constructed directly: replay determinism guarantees structural equivalence. So we BUILD the mature graph by hand:

  1. Start from the phase 8 corpus (200 units, experiments/phase8/).
  2. Synthetic debate generation (LLM-driven, cached): for each of the ~60 highest-retrieval-traffic units, generate a debate tree of depth 3 to 5: targeted CONs against specific claims (via submit_con semantics), PRO rebuttals, counter-CONs, cross-unit co-parent edges where units genuinely relate. Target ≥2000 nodes total, mixed contested/uncontested regions, generated by Opus with instructions to argue both sides honestly (not to favor either arm).
  3. Build two standings over the SAME graph:
    • C': recursive vote standing (net_score recursion over the explicit tree: posts + signed children, no geometry).
    • D': full equilibration (geometric propagation + recursion). NOTE: phase 8's arm C used flat per-unit vote counts; at depth the fair no-geometry baseline is the tree recursion, which the ledger keeps anyway. This is pre-registered here to prevent later dispute.
  4. Contest protocol identical to phase 8 (same selected questions, same contestant, same blind dual grading, same guards, same stats), arms: B (plain rag), C' (tree-recursion re-rank), D' (equilibrated re-rank). Primary contrast D'−C', same decision rule: Holm p < 0.05 AND mean diff ≥ 0.25.

Pre-commitments

  • The verdict is final in BOTH directions: D'−C' clears the bar → equilibration is re-promoted and the O(N²) scaling work is funded; it fails → equilibration is retired from the mint path permanently (may remain a research kernel, never a launch dependency).
  • The synthetic-graph generator is committed before any contest call, with its prompts; graph statistics (depth histogram, contested-node fraction, CON share) are published with the results.
  • No new caveats may be introduced after data collection. Any validity concern must be raised (and fixed) before the contest runs.

Cost estimate

Graph generation ~$15 to $25 of Opus calls (cached, one-time); world build at 2000+ nodes will be slow under current O(N²) equilibrate (hours), acceptable for a one-shot offline build; contest + grading ≈ phase 8 (~350 bridge calls).