Skip to main content

Highlights

Evidence So Far is the full audit: every track, honestly labeled as confirmed, demonstration, null, or non-replication. This page is shorter and more selective — the results and findings worth knowing about even if you read nothing else, in roughly descending order of how load-bearing they are for the site's central claim. Read it as a map into the deeper pages, not as a substitute for their caveats.

1. Combined evolution and learning beat every alternative​

The Ackley & Littman (1991) replication and its Trial 12 extension ran a full 500-run comparative study — 5 conditions × 100 seeds × a 1,000,000-step ceiling, matching the original paper's own scale. Combined evolution-and-learning beat evolution alone, learning alone, no adaptation, and pure random behavior, p < 0.00001 against all four (Mann-Whitney U, n = 100 per condition). This is the site's own designated Main Result — the cleanest evidence for the two-timescale claim everything else on the site builds toward.

It came after ten straight null or inconclusive trials (1–10) under a shared-policy architecture where the genome could only act on behavior indirectly. Trial 12 changed the architecture — giving each agent its own genome-conditioned network — not the trait, and that's what finally made the effect visible. The nulls weren't wasted motion; they correctly diagnosed the actual limitation.

2. Learning changes the direction a gene evolves​

In Trial 15, animals in the continuous PredPreyGrass world carry heritable body genes (top speed, view distance, field of view) and share one learning policy per species. Branched from one run, predator speed rose in all six runs where the predator policy kept learning and fell in all six where it was frozen (+0.18 on the gene's log scale, 95% CI +0.14 to +0.22, exact p = 0.005, a contrast fixed in advance), and from a second branch point the contrast was +0.19 (p = 0.005). The likely route: with speeds set by hand, extra speed paid about twice as much under the policies predators had learned. Where Trial 12 shows that evolution and learning together outperform either, Trial 15 shows the Baldwin-style link itself: learning steering which way a body trait evolves. The caveats: both branch points come from one training run, the learned policies differ from the frozen one in more than their use of speed, the learning is shared across a species rather than individual, and no arms race in bodies between predators and prey appeared.

3. A hand-designed strategy beats the project's own literal fitness reward​

Trial 13, run in the project's richer flagship predator-prey-grass ecology, is the most thoroughly diagnosed single result in the trial history. A cautious, predator-avoiding strategy beats a reckless one decisively once the two compete for the same resources (complete exclusion within ~27 generations, reversed in 0 of 10 seeds) — and beats flagship's own "+10 on reproduction" reward signal too, 27 of 30 seeds, p = 0.000008. The honest caveat that makes this a finding rather than a trophy: the winning strategy was hand-designed, not evolved. A follow-up with a stronger pooled learner let evolution discover something like it on its own, but only in 3 of 30 seeds (10%) — real and reproducible, but rare, pointing at founder-effect luck rather than a fixable bottleneck.

4. Two foundational-replication successes, hard-won​

  • Leibo et al. (2017): independent DQN learners in Gathering and Wolfpack produce policies whose induced payoffs satisfy the classical dilemma inequalities — reproduced at reduced scale (not yet the paper's own 40M-step training budget).
  • Foerster et al. (2018): tracking down the authors' own reference implementation surfaced the paper's real Table 4 delta (0.5), closing half of an asymmetric-pairing gap that a from-scratch build alone had left open — alongside a genuine bug caught in the REINFORCE baseline along the way.

5. Boyd & Richerson (2009) — confirmed, with an independent refinement​

Both replications — Culture and Voting With Your Feet — reproduce their source paper's claims: cultural group selection survives a far lower migration/selection ratio than genetics, and payoff-biased migration spreads group-beneficial behavior. Culture also adds an independently confirmed refinement to the paper's own homogenization threshold; Voting With Your Feet adds a welfare check the original paper doesn't run, showing that spreading the behavior label doesn't reliably raise welfare.

6. Two honest non-replications​

Hughes et al. (2018)'s headline Cleanup claim — that inequity aversion raises cooperation — does not reproduce: baseline beat the guilt-shaped condition in 6 of 6 seed comparisons, the opposite of the paper's own figure. Under-training, miscalibration, and seed noise were each checked directly and ruled out. McKee et al. (2023)'s reputation claim can't even be tested here — training collapses to near-zero cleaning before reputation dynamics could matter, in both conditions, including with the mechanism disabled entirely. Both papers share the Cleanup environment, which is itself a pattern worth taking seriously.

7. A decisively closed-out null result​

The reward-density comparison tested sparse vs. dense/shaped reward across five sibling PredPreyGrass modules. Sparse reward won outright on reproduction rate, population balance, and extinction avoidance against all four denser variants — a rare case of a null-hypothesis-style question getting a clean, fully replicated, ranked answer instead of an ambiguous one. The same investigation also surfaced and fixed two silent RLlib data-integrity bugs that had been present since Trial 1.

8. Reusable infrastructure: the drift-vs-selection tool​

Built to resolve one ambiguous Trial 13 result, the project's Hunt (2006)/Lande (1976) drift-vs-selection model-fitting tool turned into standing infrastructure — now wired into six modules. It confirms prior null verdicts with a formal statistical test rather than an eyeballed trend, and along the way caught a real directory-matching bug and flagged an unresolved anomaly in the nuptial-gift module.

Where this leaves the argument​

None of this adds up to a finished human-specific model of cooperation and competition — see Evidence So Far for what's still open. What it does show is a two-timescale mechanism (evolution across generations, learning within lifetimes) that reliably outperforms either half alone once the architecture gives the genome a real channel into behavior, plus a research process willing to spend a dozen trials chasing down why a result was null instead of settling for the null. Trial 15 adds the other direction of the link: learning, here one policy shared by a species and trained across generations rather than within lifetimes, changing the course of evolution, from two starting points of one training run, consistent with learned behavior making a faster body pay.