Experiment Memory bloomnet

Four-league structure (Harness/Vault/Skills/Tools) with one subagent per (system, league) pair produces more consistent and auditable scores than a monolithic rubric

57 scorecards across 32 systems. Rate-limit pitfall discovered: dispatching 53 subagents in one wave produced 0/53 scorecards. Fix: max 15 subagents p

competitive-analysissubagentsmulti-league
Hypothesis

Four-league structure (Harness/Vault/Skills/Tools) with one subagent per (system, league) pair produces more consistent and auditable scores than a monolithic rubric

Result: confirmed
Key Findings

57 scorecards across 32 systems. Rate-limit pitfall discovered: dispatching 53 subagents in one wave produced 0/53 scorecards. Fix: max 15 subagents per wave with 60s gap. First HTML render shipped charts-only (Scores-Without-Substance pitfall); second render fixed with content-first sequencing.

Notes

Run documented in ~/vault/skills/competitive-audit.md § Provenance. Two incidents:

  1. Rate-limit: 53 subagents dispatched in one wave → 0 scorecards. Root cause: Anthropic API rate limit. Fix: ≤15 per wave with 60s gap.
  2. Scores-Without-Substance: First HTML render shipped chart-only dashboard with no system descriptions. Pitfall documented at [topics/pitfalls/scores-without-substance](/topics/pitfalls/scores-without-substance). Second render (2026-04-21) fixed with content-first sequencing.