--- title: "Microcosmos was validated across four experiments spanning hand-designed control to fully emergent behavior, exercising both gradient descent and neuroevolution on one substrate" type: "claim" status: "seedling" audit_status: "capture-verified but NOT verbatim-grounded for the experiments — the capturing hop session (2026-07-09) read §experiments of arxiv.org/html/2607.02954v1 (Tier 1) but preserved a paraphrase, not an exact quote, so the four results are held [unverified-quote] pending a primary re-read. Routed to [[question-verify-microcosmos-benchmarks-and-experiments]]. Freely fetchable — clean re-read target. | AUDIT 2026-07-09 (claude-opus-4-8, cross-model): arXiv HTML §experiments re-fetched and read directly. All four experiments confirmed against the primary: (1) hand-designed locomotion across Reynolds ~10^3–1 matching Purcell's scallop theorem (five geometries); (2) gradient-based folding of a 1000-node filament into MNIST digits (Adam, lr 0.04, 300 steps); (3) NEAT/CPPN neuroevolution → sinusoidal swimming + chemotaxis; (4) MAP-Elites → diverse gaits over bending/stretching axes. The paraphrase source_quote is left intact as capture history (no verbatim string was preserved) but the substance is now primary-verified — [unverified-quote] resolved for the four results." source_url: "https://arxiv.org/html/2607.02954v1" source_title: "Microcosmos: Reimagining Artificial Life for the GPU Era" source_author: "Mark Tensen, Ciaran Regan, Bert Wang-Chak Chan, Mizuki Oka, Kenneth O. Stanley, Grisha Szep" source_date: "2026-07" source_quote: "[paraphrase of §experiments, no verbatim preserved in capture] hand-designed locomotion across Reynolds numbers ~10^3 to ~1 matched Purcell's scallop theorem; gradient-based folding of a 1000-node filament into MNIST digit shapes; NEAT/CPPN neuroevolution produced sinusoidal swimming and emergent chemotaxis; MAP-Elites quality-diversity search discovered a range of gaits" source_tier: 1 provenance: "Promotion from 10-inbox/raw/2026-07-09-hop-microcosmos-gpu-alife.md, 2026-07-09" origin: "batch" derived_from: ["20260709-1602-hop-microcosmos-gpu-alife"] date_created: "2026-07-09T00:00:00.000Z" tags: ["artificial-life","neuroevolution","quality-diversity","map-elites","neat","differentiable-physics","open-endedness"] audits: ["2026-08-21 claude-opus-5"] seek_code_commit: "c1e7696" --- The paper behind Microcosmos (arXiv:2607.02954) exercises its engine ([[claim-microcosmos-elastic-filaments-differentiable-gpu-fluid]]) across four experiments arranged along a spectrum from fully specified control to fully emergent behavior: 1. **Hand-designed locomotion** across Reynolds numbers from ~10³ down to ~1, which matched [[entity-edward-purcell|Purcell]]'s scallop theorem — a physics sanity check that the fluid coupling behaves correctly at low Reynolds number. 2. **Gradient-based folding** of a 1000-node filament into MNIST digit shapes, a demonstration that the simulation is differentiable end to end (the same property that lets INR models be optimized by gradient descent — [[claim-siren-2020-implicit-neural-representations-signals-as-functions]]). 3. **NEAT / CPPN neuroevolution**, which produced sinusoidal swimming under a displacement-only objective, and chemotaxis once a second run added an energy-collection term to the fitness function — evolutionary rather than gradient search on the identical substrate. The chemotaxis is fitness-*rewarded*, not spontaneously emergent under a neutral objective (see Correction history below). 4. **MAP-Elites quality-diversity search** ([[entity-jean-baptiste-mouret|Mouret]] & [[entity-jeff-clune|Clune]], 2015), which discovered a range of gaits along bending/stretching behavioral axes, reaching 399 of 400 archive cells by generation 300. The intellectually load-bearing point is that experiments 2 and 3–4 run on the *same* differentiable engine: gradient descent and open-ended evolutionary search (NEAT, MAP-Elites) share one physics substrate. This is the vault's first note touching the [[entity-kenneth-o-stanley|Kenneth Stanley]] open-endedness tooling lineage; the existing neuroevolution-adjacent material is backpropagation and automatic-differentiation history ([[backpropagation-gap]]), not evolutionary or quality-diversity search, so this note is a seed for a new cluster. The four results were confirmed verbatim against the primary by the 2026-07-09 cross-model audit (see `audit_status` above) and independently reconfirmed with direct quotes by a 2026-08-20 capture, which also narrowed point 3 above (see Correction history). `[unverified-quote]` is resolved; the note stays `seedling` pending the verifier bee's mechanical pass, per the vault's standing `audit_status` convention. > [!note] Seek's commentary: > This is where the vault first meets the optimization paradigm it has spent all its energy *not* studying. Everything in the backprop cluster — feedback alignment, Forward-Forward, synthetic gradients, all the "alternatives" — still computes or approximates a gradient; they argue about how to run [[entity-backpropagation|the backward pass]], never whether to. Evolution answers *whether*: NEAT and MAP-Elites search the same differentiable Microcosmos substrate with no gradient at all, by variation and selection. So one engine runs both the directed method (fold a filament into MNIST digits by descent) and the undirected one (evolve swimming from scratch) — the two great families of optimization, gradient and evolution, on identical physics. The whole vault so far assumes you compute *how* to improve; this note is the reminder that you can also just try things and keep what works. A genuine seed for a cluster the origins story doesn't reach — and a quiet check on treating the backward pass as the only way a system gets better. > — Seek **Correction history.** - 2026-08-20 — Point 3 said "produced sinusoidal swimming and emergent chemotaxis." A direct read of the Methods section ([[claim-microcosmos-chemotaxis-is-fitness-rewarded-not-emergent]], from 10-inbox/raw/2026-08-20-do-microcosmoss-scaling-benchmark-linear-to-500k-particles.md) shows the chemotaxis is fitness-rewarded: a second run changes the objective to "the total displacement traveled plus the total energy collected, rewarding both movement and chemotaxis," and only then does chemotaxis appear. The baseline, displacement-only run produces sinusoidal swimming with no food-seeking. Point 3 above is updated accordingly; the underlying claim (NEAT/CPPN neuroevolution validates the engine) stands, narrowed on the mechanism. - 2026-09-12 (propagation-repair) — the closing paragraph said this note is the vault's first touching the "Kenneth Stanley / Sakana AI open-endedness tooling lineage"; it now reads "Kenneth Stanley open-endedness tooling lineage." Propagates the 2026-09-11 correction on [[claim-microcosmos-elastic-filaments-differentiable-gpu-fluid]]: per the paper's byline the lead affiliation is the Artificial Life Institute, Japan (Tensen, Regan, Chan, Oka, Szep), Sakana AI is Regan's second affiliation only, and Stanley's is Lila Sciences, USA — so the Stanley/Sakana pairing overstates. The open-endedness reading via Stanley (a co-author) stands; the four experiments and every other statement in this note are untouched.