Skip to content

EXPERIMENTS.md

Latest commit

 

History

History
51 lines (42 loc) · 2.65 KB

EXPERIMENTS.md

File metadata and controls

51 lines (42 loc) · 2.65 KB
Raw

Formal experiment registry

ID Paper purpose Central claim Main output
P00 integration and alpha=0 sanity none validation rows
P01 nonlinear S1 response and finite-size scaling C1, C2 alpha curves, alpha_c(N), K_c(N), width
P02 fixed-K social/liquidity decomposition C2 size effect under q=0, 0.5, 1 and NoSocial
P03 S1--S4 scope scope of C1/C2 resolved/censored scenario summary
P04 game-theoretic coupling phase diagram C3 coupling proxy, response width, alpha_c
P05 private-to-social information cascade C3 CMI, social dominance, transfer entropy
P06 role robustness and behavioral diversity audit reviewer validity role/action MI, behavior separation, diversity-vs-size controls
P07 LLM allocation robustness implementation validity matched quota response curves
P08 compact defense utility demonstration benchmark utility risk reduction, cost, false positives
P09 depth-specific dense scaling audit reviewer W1 observed (\widehat\nu) under q=0, 0.5, 1
P10 LLM-fraction dense scaling audit reviewer W3 observed (\widehat\nu) under 0x, 1x, 2x quotas
P11 Watts threshold null reviewer W5 null-model (\alpha_c(N)) and (\widehat\nu)

Planned follow-up experiments

  • P09--P11 are reviewer-response audits. P09 repeats the P01 dense alpha_c(N) sweep under fixed-depth (q=0), baseline-depth (q=0.5), and per-capita-depth (q=1). P10 repeats the dense sweep under behavioral-only, standard, and double LLM quotas. P11 runs a Watts-style threshold null on matched graph sizes and alpha grids.

Paper order

The main text should present P01, P02, and P04/P05 in that order. P03 defines scope. P06/P07 belong in robustness or the appendix. P08 is not a fourth scientific contribution.

Freeze rules

  1. A paper profile can run only after its pilot grid brackets the intended response region.
  2. Never report alpha_c when 50% failure is not bracketed.
  3. Never combine different benchmark_version values.
  4. Always report alpha=0 sanity, critical fraction, and critical harmful count.
  5. Bootstrap episode seeds, not agent-day events.
  6. Generate P04 theory predictions before inspecting paper-scale P04 outcomes.
  7. P05 contrasts and signs are fixed in analysis/information.py.
  8. P06 must report behavioral diversity as a validity/control variable; role names alone are not evidence of heterogeneity.

Status

  • Runners rewritten: P00--P08.
  • Analyses rewritten: P01, P04, P05, P06.
  • v3 protocol: exploratory until new pilot runs finish.
  • Old v1/v2 real outputs: historical only; not valid evidence for the rewritten paper.