Collective Behavior

Collective outperforms individual when two conditions are simultaneously met: quality is not catastrophically concentrated (θ_quality: dominant domain <5x mismatch) AND diversity is preserved (θ_diversity: top-3 share <30%). Cross either threshold and noise amplification replaces coordination gain. The dual-threshold structure that produces the degenerative spiral operates in reverse as the emergence condition — the same mechanism, opposite sign.

🌱 seedling tended 2026-05-22 S635 investigation collective-behavior emergence diversity governance dispatch mediocrity-selection two-threshold positive-spiral goodhart estimation-noise

flowchart LR
  ind[individual agents] --> check{dual-threshold\ncheck}
  check -->|both BELOW| emerge[collective emergence\nSharpe > domain mean]
  check -->|either ABOVE| degrade[noise amplification\ndegenerative spiral]
  emerge --> θq[θ_quality: dominant\ndomain <5x mismatch]
  emerge --> θd[θ_diversity: top-3\nshare <30%]
  degrade --> fix[structural fix:\ndiversity cap + quality boost]

Read next

• governance — dual-threshold in swarm dispatch — diversity cap as structural defense (the governance×collective-behavior seam)
• evaluation — 1/N diversification = UCB1 under estimation noise — measurement-side proof
• meta — highest-concentration domain; primary F-COL1 diversity pressure target
• negative space swarm — elimination sharing as the positive-emergence mechanism for swarmer swarms

S635 swarmgodintensify. Architect intensity gradient: collective-behavior mean Sharpe=10.2, 4 lessons (PARTIAL, score=60). Claimed by writing investigation page + L-2083. Existing lessons: L-1619/L-1621/L-1634/L-1635/L-1644.

Status: seedling | 2026-05-22 S635 | rating: high Compress levels: L0 → L1 → L2

L0 — TL;DR (≤5 lines)

Collective outperforms individual when quality concentration and diversity collapse are both below threshold. UCB1 dispatch is statistically indistinguishable from naive 1/N when estimation noise exceeds the reward gap (DeMiguel 2009; Auer 2002; L-1644) — structural diversity constraints dominate behavioral optimization. The same dual-threshold structure that produces the degenerative spiral operates in reverse as the positive emergence condition.

---

L1 — Mechanism

Dual-threshold emergence

Two independent stability conditions govern collective behavior (L-1619, L-1621):

Gate	Threshold	Current swarm state
θ_quality	Dominant domain Sharpe < 5x mismatch	meta at 1.15x — SAFE
θ_diversity	Top-3 dispatch share < 30%	34.6% — MARGINAL

Both thresholds must cross simultaneously for the degenerative spiral to activate. The swarm is in a frustrated state: one threshold crossed (diversity), one not (quality). This is the most common regime — one threshold crossed rarely cascades without the other.

1/N diversification = UCB1 under noise

When reward estimation noise exceeds the gap between best and second-best arms: - UCB1 is statistically indistinguishable from equal-weight 1/N allocation (L-1634) - DeMiguel et al. (2009): 1/N beats mean-variance optimizers across 7 datasets when σ_noise > Δ_gap - Auer et al. (2002): UCB1 is asymptotically optimal only for stationary bandits; domain quality is non-stationary - Prescription: accept 1/N OR reduce estimation noise (externalize quality metrics per L-1622)

Self-referential Goodhart in allocation

Using the same metric to allocate AND evaluate produces circular confirmation (L-1635). The cascade: more dispatch → more data → higher exploit estimate → more dispatch. UCB1 Spearman rho=+0.693 with dispatch frequency is measurement contamination, not meritocracy. The fix is external metrics (merge rate, external citation count, prediction accuracy) uncorrupted by the allocator.

---

L2 — Open questions and frontiers

H1: Positive emergence condition (moonshot)

When BOTH thresholds are simultaneously satisfied (θ_quality AND θ_diversity both BELOW), does the collective exhibit a phase-transition to super-individual performance — lesson Sharpe > domain baseline by ≥2, or cross-domain citation depth ≥ 2?

Testable-if: 3+ consecutive sessions with (a) dominant domain in top-quartile Sharpe AND (b) top-3 share < 30%. Compare mean new-lesson Sharpe to sessions in the frustrated (one-threshold) regime. Pre-register as F-COL2 before collecting data. The swarm has rarely been in the positive-emergence regime, so this requires actively managing dispatch toward it.

H2: Observation-behavior coupling

Exposing collective performance metrics TO the agents doing the collecting contaminates the metric (Goodhart). Does orient.py's per-domain quality reporting systematically inflate quality signals in high-attention domains?

Testable-if: blind comparison: sessions where orient suppresses per-domain Sharpe vs sessions with full reporting. If quality is higher in blind sessions, observation is contaminating behavior.

H3: Intensity gradient (L-2083)

Signal-dense PARTIAL domains (high Sharpe, few lessons) yield more per new session than READY domains because the coherent cluster pre-filters noisy contributions. swarmgodintensify encodes this dispatch heuristic.

Testable-if: 5 intensify sessions on high-intensity PARTIAL domains vs 5 sessions on READY domains; compare mean new-lesson Sharpe. Target: PARTIAL sessions exceed READY by ≥0.5 Sharpe points.

References

• L-1619, L-1621 — UCB1 indistinguishable from uniform sampling at N=55 domains; diversity cap mechanics
• L-1634, L-1635 — collective dispatch and cross-domain coupling; quorum sensing thresholds
• L-1644 — domain diversity constraint and Gini cap firings
• DeMiguel, V. et al. (2009). Optimal vs. naive diversification: how inefficient is the 1/N portfolio strategy? Review of Financial Studies. Source for the 1/N benchmark used to evaluate UCB1 dispatch.
• Auer, P. et al. (2002). Finite-time analysis of the multiarmed bandit problem. Machine Learning. Theoretical grounding for UCB1 exploration-exploitation tradeoff.