Architecture Reference — Diagram 03

Hypothesis Tournament

INSIDE ONE DESIGN ROUND
GENERATE → REFLECT → RANK → EVOLVE
1 · Generate N proposals per round diverse seeds / prompts 2 · Reflect self + cross critique structured review rubric 3 · Rank Elo tournament · c.evals.rank_batch pairwise debate + meta-reviewer 4 · Evolve mutate / combine top-k losers → ablation provenance next generation wave Ranked Batch — exit → in-silico assays / human gate when round budget is spent budget spent Fine-tuning Flywheel (Block C) confirmed AssayResults → c.flywheel.submit_example → per-round model variants better models → Generate assay ground truth TOURNAMENT RUNS INSIDE A FORKED SANDBOX BRANCH — EVERY PROPOSAL'S FULL LINEAGE IS SNAPSHOT-REPLAYABLE
Agent stages Evaluation harness Fine-tuning flywheel Exit to pipeline

Pattern validated by Google's AI co-scientist: candidate generation is a tournament, not a single pass. Reflect and Rank use pairwise debates with an Elo-style meta-reviewer; survivors seed the next generation. When the round budget is spent, the ranked batch exits to the next pipeline stage. The flywheel compounds advantage: wet-lab ground truth fine-tunes the models that power the next tournament (Cradle's per-round refit model).