# Cosmogonic Quantum Mechalogodrom — Technical Specification **[CQM-25 canonical capability framework](./alife/cqm-capability-profile.html)** — version **1.0.0**, adopted **2026-10-22**. The [source data and protocols](./alife/cqm-capability-profile.json) own the 25 stable axis IDs. Feature plans and capability assessments must name affected IDs, subject/scope, measures, controls, falsifiers and pinned source/run provenance, including failures. The [adoption contract](./MODULE-CONTRACTS-2026-06-26.md) governs taxonomy changes and publication. Evidence states describe implementation and bounded tests; canonical `assessment.score` fields remain null until calibrated. A single organism is eligible; ecology is one domain, and physical models are judged against declared dynamics. CQM-25 is the project standard, not an externally validated scientific certification. The separate 14 Butlin indicators remain distinct. The historical nine-axis 155-system comparison cannot supply CQM-25 scores, ranks or an aggregate verdict. [Provisional author estimates](./alife/cqm-capability-profile.html#provisional-estimates) are a separate, explicitly labeled layer with their own [rubric and source data](./alife/cqm-capability-estimates.json). They retain source provenance and limitations, confer no measured capability, and do not populate canonical `assessment.score` fields. No overall or domain mean, ranking or aggregate verdict is published. > **Current intelligence contract (2026-07-11):** `TsotchkeOrganismIntelligence` is one shared, > bounded-cadence field feeding living-system consumers through a stable signal object. The public > census contains 23 repositories; the causal/runtime ledger intentionally contains 22 entries > (`8 deep`, `7 wired`, `2 harvest`, `4 fenced`, `1 meta`; `17/21` non-meta integrated). > `homebrew-moonlab` is census-only deployment metadata and `classical-contrast` is an internal control > outside both denominators. V3 remains the predecessor fixed-family receipt. The verified V4 Phase-A > descendant publishes 1,152 rows over 64 frozen seeds: ordinary and Petri pass inference but miss the > fixed magnitude floor, the adaptive predictor loses to frozen and shuffled controls, and only Titans > pass. The only new authorized claim is bounded Titan game-policy semantic causality—no ordinary > recurrent benefit, adaptive prediction, Petri causality, neural scaling, pooled cross-family result, > numeric capability change, consciousness, or sentience uplift. > [V4 receipt and limits.](./reports/ORGANISM-INTELLIGENCE-V4-RESULTS-2026-07-11.md) > > **Implemented world coupling:** flora, waste, trophic interactions, sanctuary refuge, > titan markets and weather form declared feedback paths. Collective fields and NHI controllers > are separate mechanisms; wilderness is one-way non-writeback. Capability claims use affected > CQM-25 axis IDs and their measures/controls/falsifiers. The historical nine-axis rubric is not > the current capability specification, and implementation alone does not establish consciousness. > A complete, measured technical specification of a browser-native, deterministic, 50,000-agent > WebGL cosmic-ecosystem simulation. Every figure here was measured from the repository, not > estimated. **Live:** · **Spec page:** > `/spec` · **Architecture docs:** `/docs` **Version:** v0.23.0 · **Generated:** 2026-06-26 · **License:** Non-commercial research & play (© 0thernes; study / run / modify / share non-commercially, keep attribution, no for-profit use). **Gate:** 3,631 tests · 84.64% line / 82.21% func (canonical synced via verify-receipts 2026-06-26, receipts law enforced). **Historical standing / xeno-A-life position (2026-07-12):** readable synthesis of engineering, comparative matrix, and apex-mind design (SuperCreature ≈1.4k-param multi-faculty spine + online heads; SuperMind ~10k composite; Apex/Mechalogodrom designed scaling roadmaps)—**xeno scaffolds, not OpenWorm competitors, not sentience claims.** See [COMPREHENSIVE-STANDING report](./reports/2026-07-12-COMPREHENSIVE-STANDING-AND-XENO-ALIFE-POSITION.md) · [HTML](./reports/2026-07-12-COMPREHENSIVE-STANDING-AND-XENO-ALIFE-POSITION.html). **Mechanism and separate Butlin indicator status:** [NHSI-PROGRESS-DASHBOARD-2026-06-26.md](./NHSI-PROGRESS-DASHBOARD-2026-06-26.md) · **Tsotchke depth:** [TSOTCHKE-INTEGRATION-MAP-2026-06-26.md](./TSOTCHKE-INTEGRATION-MAP-2026-06-26.md) **Tsotchke depth ledger:** The public census contains 23 repositories. The causal/runtime ledger intentionally contains 22 entries: `8 deep`, `7 wired`, `2 harvest`, `4 fenced`, and `1 meta`, with `17/21` non-meta entries integrated. `homebrew-moonlab` is census-only deployment metadata. The internal `classical-contrast` control is operational but outside both counts. Eshkol is represented by bounded ports and an order-0-through-8 Float64 Taylor analogue pinned to v1.3.2; the QRNG path is a deterministic classical statevector adaptation pinned to v3.0.1, not physical entropy or a CSPRNG. Primordial soup/petri is a growth engine for computational digital biologics and indicators, not evidence of sentience. --- ## 0 · Identity A single-page **TypeScript** simulation — with an optional native **C++/Jolt** physics engine (ADR-0007) — that renders up to **50,000 morphogenic organisms** (10,000 at 60 fps on integrated graphics; the 50k mega tier is opt-in for beefy GPUs), governed by **deterministic, seeded** physics and a **pre-2016 classical-AI** cognition stack — finite-state machines, GOAP planning, utility AI, tiny perceptrons, Markov chains, game theory. There is **no neural-network accelerator in the loop**: the intelligence is architectural, not parametric. The whole world is **bit-reproducible from one 32-bit seed**. --- ## 1 · Codebase metrics (measured) These figures are **measured, not pinned** — line counts move every commit, so they are a dated snapshot rather than a gate-enforced receipt. Refresh them with `bun run metrics` ([`scripts/codebase-metrics.ts`](../scripts/codebase-metrics.ts), deterministic over `git ls-files`). Snapshot: **2026-07-14**. (Coverage + passing-test counts ARE SSOT-synced — see §1's last two rows.) | Metric | Value | | --------------------------------------- | --------------------------------------------------------------------- | | Total tracked authored files | **940** | | Total tracked authored lines | **295,227** | | App source (`src/`) | 131,572 lines · 327 files | | Tests (`tests/`) | 63,495 lines · 367 files | | Native C++ engine (`native/`, ADR-0007) | 1,835 lines · 12 files | | Test : source ratio | 0.48 → **84.64% line / 82.21% func** coverage (`bun test --coverage`) | | Passing tests | **3,631** (exact tracked suite; 0 failing) | (Excludes the vendored `node_modules/`, generated `dist/` · `coverage/` · `native/build/`, and nested `.claude/worktrees/` checkouts.) ### 1.1 By file type | Type | Files | Lines | Share | | -------------------- | ----: | ------: | ------: | | TypeScript | 762 | 218,578 | 74.04 % | | Markdown | 78 | 24,928 | 8.44 % | | JSON | 19 | 16,797 | 5.69 % | | HTML | 11 | 15,583 | 5.28 % | | PNG (shots) | 4 | 9,552 | 3.24 % | | CSS (Tailwind) | 2 | 2,476 | 0.84 % | | CSV | 6 | 2,070 | 0.70 % | | YML | 11 | 1,016 | 0.34 % | | C/C++ headers + impl | 8 | 1,527 | 0.52 % | | XML | 4 | 707 | 0.24 % | | PowerShell | 2 | 515 | 0.17 % | | Lockfile | 1 | 454 | 0.15 % | | No extension | 9 | 287 | 0.10 % | | Text | 5 | 238 | 0.08 % | | SVG | 14 | 335 | 0.11 % | | JavaScript | 1 | 85 | 0.03 % | | DOCX | 1 | 64 | 0.02 % | | YAML | 1 | 9 | 0.00 % | | TOML | 1 | 6 | 0.00 % | **Code (TS + C/C++ + HTML + CSS) = 238,164 lines = 80.67 %**; documentation + config + assets = 19.33 %. ### 1.2 Lines by area | Area | Files | Lines | | ------------------------------------- | ----: | ------: | | `src/` (application) | 327 | 131,572 | | `tests/` | 367 | 63,495 | | `docs/` | 95 | 32,958 | | `scripts/` | 52 | 23,990 | | `lab/` (self-contained artifact) | 5 | 14,039 | | repo root (README, LICENSE, configs) | 27 | 12,187 | | `output/` (tracked visual receipts) | 3 | 9,551 | | `legacy/` (preserved origin) | 6 | 2,072 | | `native/` (C++ engine, ADR-0007) | 12 | 1,835 | | `bench/` | 18 | 1,605 | | `.github/` | 13 | 1,080 | | `masters/` (3 governing XML personas) | 3 | 477 | | `.memory/` | 5 | 167 | | `.vscode/` | 3 | 94 | | `.githooks/` | 2 | 93 | | `.claude/` | 1 | 11 | | `public/` | 1 | 1 | Detailed `src/` subsystem counts are intentionally generated on demand by `bun run metrics`; this section records only the current high-level measured areas to avoid stale hand-maintained internals. --- ## 2 · Technology stack & versions **Runtime & frameworks (5):** Bun `1.3.14` (runtime + bundler + test) · TypeScript `7.0.2` (strict, `verbatimModuleSyntax`) · Three.js `0.185.1` (WebGL2 renderer) · Tailwind CSS `^4.3.2` · HTMX `^2.0.10`. **Production libraries (12):** three `0.185.1` · mermaid `^11.16.0` · graphology `^0.26.0` (+communities-louvain `^2.0.2`, +metrics `^2.4.0`) · d3-delaunay `^6.0.4` · simplex-noise `^4.0.3` · simple-statistics `^7.9.3` · @noble/hashes `^2.2.0` · htmx.org `^2.0.10` · @fontsource-variable/inter `^5.2.8` · @fontsource/jetbrains-mono `^5.2.8`. **Dev tooling (10):** oxlint `^1.73.0` · prettier `^3.9.5` (+tailwind plugin `^0.8.0`) · tsc `7.0.2` · mitata `^1.0.34` (bench) · bun-plugin-tailwind `^0.1.2` · @types/{bun, three, d3-delaunay}. **22 declared dependencies resolve to 153 CycloneDX components** in the current frozen install graph — a facade-isolated tree with explicit runtime/development scopes and dependency references (see [ADR 0005](adr/0005-math-stack-selection-2026-06-26.md)). --- ## 3 · System architecture A strict acyclic layering (verified across 50+ modules — no runtime import cycles): ```text math / constants (leaves: rng, scalar, spatial-hash, quantum, heap, games) ▲ sim/* (30+ behavioural systems: entities, titans, shoggoths, nhi, factions, ·) ▲ core (engine, quality) ── world.ts (composition root: builds SimContext, ticks every system) ▲ ui / render (observatory, hud, panels, viz3d, instanced-entities, postfx) ▲ server.ts (Bun.serve) ──serves──▶ index.html (/) · docs.html (/docs) · specs.html (/spec) · lab (/lab) ``` - **Composition root:** `world.ts` owns the `SimContext` dependency bag and ticks every system once per frame. Leaf modules never import the type hub at runtime. - **The aesthetic law** ([PHILOSOPHY-2026-06-26.md](PHILOSOPHY-2026-06-26.md)): _every system reads AND writes at least one other system._ The frame loop is an explicit feedback web (e.g. quantum collapse → quantum cloud; reaction-diffusion → ground emissive map; analytics → audit; NHI → entity behaviour + factions; entity connectome firing density → xenomimic swarm agitation; beings → xenomimic predation). - **Xenomimics** (`sim/xenomimics.ts` + `sim/xenomimic-brain.ts` + `sim/xenomimics-render.ts`): a self-contained ground-fauna subsystem on its own seeded rng substream (never touches `ctx.rng`, absent from every golden). 10 tessellated species live as bipolar ENTANGLED TWINS - one ~100-parameter brain per pair, run with opposite thought-curvature over a real 3-qubit quantum singlet (superposition / Born-rule teleport / entanglement), with IIT-integration, GWT-broadcast and a Free-Energy-Principle predictive loop wired into the same brain. The swarm starts as one pair, multiplies toward a cap of 1000, grazes the flora, and respawns 5s after any being grazes it. World-level couplings (all golden-safe; no test constructs the full `World`): the entity connectome's live firing density drives every twin brain's `chaos` sense, and the entity spatial grid drives bounded predation. Surfaced via the XNO spawn button, the XENOMIMIC focus + Archon-template data window, the telemetry row below Entities, and a dedicated eerie tonality bus. - **Boundary fence:** the non-deterministic LLM copilot (`server/`, `ui/copilot.ts`) imports nothing from `sim`/`world`/`types`, so it provably cannot reach the seeded RNG or simulation state. - **Center HUD readability:** `ui/center-hud.ts` owns one tall shared center slot (`clamp(300px, 56vh, 660px)` desktop; `clamp(320px, 64vh, 720px)` touch). The Architecture panel uses a responsive canvas/data body and resizes its backing canvas to the actual CSS box, so it no longer collapses into a thin visualization strip. - **9 ADRs** record binding decisions (Bun runtime, Three.js, HTMX+Tailwind, deterministic RNG, math stack, [ASI graphics stack](adr/0006-asi-graphics-and-language-stack-2026-06-26.md), the native C++ engine, the Super-Creature deep mind, and genome reproduction). --- ## 4 · Frame pipeline (`world.step`) Each frame (`dt` clamped to ≤ 50 ms, then × time-scale): 1. camera + hotkeys + chaos/entropy decay 2. audio analysis → 4-band reactivity 3. macro-agents: puppet-masters, shoggoths, titans, leviathans, singularities 4. **entities.update** — up to 50k organisms: behaviour fields, physics, spatial-hash neighbours, auto-split, temperature-modified death + sparse respawn 5. **NHI exact-target pass** (every frame while an NHI is live, hard cap 1,000 `NHI_SYSTEM_MIND_CAP`; populations <=64 keep the exact pairwise social sweep, above that it is stagger-budgeted): rebuild current positions, prune lifecycle state in O(live NHI), use spatial local-kin sensing, then percept → GOAP-biased decision → structured effect/fact outcome → ordinary minions / dominate / hunt / mimic / broadcast. The index rebuild is zero work at `M=0` and one `O(N)` pass for any `M>0`, independent of NHI count. One mind failure cannot starve later minds. 6. connectome (graphology, strided cadence), quantum circuit + cloud, reaction-diffusion (128²), constellations, analytics 7. instanced-mesh mirror + render Population-scale steady-state paths reuse typed-array scratch; low-cardinality control/error paths use bounded objects, and organism allocation occurs only on event-driven growth. At the ultra tier, organic mitosis, sparse recovery, behavior births, and NHI swarms share one 512-birth frame budget; failed reservations refund and tiers at 5,000 entities or below preserve their direct spawn stream. --- ## 5 · Determinism model - **One root seed, domain-separated generators:** `mulberry32(seed)` plus fixed derived namespaces (`math/rng.ts`). Same 32-bit root seed and event schedule → identical run, bit-for-bit. Physics, genome, UI, economy, NHI birth, NHI policy, and direct NHI action sampling have separate streams. Material NHI organism creation intentionally consumes the ordinary physics/genome streams. - **Banned in simulation logic:** `Math.random`, `Date.now`, wall-clock — enforced by review + the determinism test suite (`determinism.test.ts`, `feature-determinism.test.ts`, `nan-stability.test.ts`). - **Boot-stream neutrality:** systems that draw no RNG at construction can be placed anywhere without shifting the seeded stream. User events draw RNG and are audit-recorded. Failed NHI launch attempts consume a monotonic attempt id and any draws already made; rollback never pretends streams rewind. --- ## 6 · Render pipeline - **WebGL2** via Three.js, ACES filmic tone-mapping, `LinearSRGBColorSpace` output with `ColorManagement` disabled (reproduces the authored legacy palette exactly). - **Instancing:** up to 50,000 organisms collapse into **InstancedMesh pools** (≤ 80 pools, one draw call per geometry × transparency pair — not 50,000 draws). Per-instance channels: transform matrix, `instanceColor`, and a custom `vec4 instEmissive` (rgb = emissive·intensity, a = opacity) patched into the standard shader via `onBeforeCompile`. - **Owned shader effects (GPU, no textures):** a baseline glass-jewel fresnel + thin-film sheen on every organism; 10 render modes (solid → wire → ghost → neon → chrome → hologram → iridescent → plasma → obsidian → prismatic); a SIMULATION N(2) vertex-melt gated on a `uNightmare` uniform. - **Optional cinematic post-FX (`?fx=1`):** a procedural cosmic PMREM environment map (glass reflections) + an `EffectComposer` UnrealBloom pass — guarded (any failure falls back to the plain pipeline), OFF by default. - **Environment systems:** `cosmic-web.ts`, `gold-lattice.ts`, `quantum-lattice.ts` add depth and context (frost-fractal web, gold-line architecture, neon sacred-geometry lattice). - **Ascension temple:** `monolith-temple.ts` is now a deterministic reactive abomination, not a static arch: black-hole shadow core, singularity ring, warped impossible line cage, jagged altar spikes, portal rings, and a halo. It reads chaos, entropy, and population/capacity crowding through `setEnvironment()` and exposes a finite `snapshot()` for tests; it draws no RNG and writes no sim state. - **Central spectacle systems:** `mechalogodrom.ts` renders 10 additional bipolar titan-variant shells that converge into a central shadow-core / event-horizon / warped-mesh hybrid; `alphabet-pantheon-render.ts` renders all 100 Greek+Latin alphabet archetypes as five instanced dome-body pools. Both are boot-stream-neutral visual projections: no RNG draws, no simulation-state writes, chaos read only. --- ## 7 · Intelligence architecture (the "AI") A deterministic, allocation-free kernel of pre-transformer techniques (`sim/ai/brains.ts`): | Primitive | Technique (origin) | | ----------------------------------------------- | ------------------------------------------------------------------ | | `utilityPick` / `softmaxPick` | utility / needs AI (The Sims, 2000) | | `TinyMLP` | single-hidden-layer perceptron — an inheritable gene | | `MarkovChain` | first-order Markov (Shannon 1948) — agent "speech" | | `fsmStep` | finite-state machine (Pac-Man 1980) | | `goapPlan` | goal-oriented action planning (F.E.A.R., 2005) | | `MemoryRing` | bounded episodic memory (Halo 2 blackboard, 2004) | | `bestResponse` / `iteratedMove` / `regretMatch` | game theory (von Neumann 1928; Axelrod 1984; Hart-Mas-Colell 2000) | **The NHI super-mind** (`sim/nhi.ts`) composes all of the above into the apex agent: it models rival factions as iterated-game players, **plans a multi-step GOAP scheme toward DOMINION** (`F_DOMINANT | F_DECEIVED`), nurses memory grudges, speaks an inherited alien Markov dialect, and WRITES to the world (spawns ordinary swarmling minions, manipulates factions, perturbs nearby behaviour, broadcasts hallucinated utterances) — wired into `world.ts`, guarded, deterministic. Its inherited gene now receives the four bounded Tsotchke resource/threat/exploration/social lanes directly, with a separate neural-semantic ablation that leaves the hybrid hand-written utility routes intact. Zero semantic input preserves the former neural output and default constructor RNG position exactly. Versioned JSON-safe mind/system checkpoints restore exact neural, memory, game, GOAP, regret, and beat state. World actions separate `effectApplied` from the narrower `factSupported`, affected-body count, and energy transfer; contradictory/action-invalid evidence is rejected, and the mind reports whether its internal fact state actually transitioned before planning advances. ### 7.1 Model parameter sizes (measured) | Mind | Network | Parameters | | ------------------------------------------------- | ------------------------------------------ | ----------------------------------- | | Organism brain (× up to 50,000) | TinyMLP 6→6→4 | **70 weights** | | Faction brain (× 8 archetypes) | TinyMLP 6→6→4 | 70 weights | | NHI intuition gene | TinyMLP 9→6→7 default; H=3/6/12 tiers | **109 live**; 58/109/211 by tier | | NHI alien voice | Markov 12×12 | 144 weights | | Ecology Predictor V3 (development-only) | 101→H→1 + direct skip; H=8/16/32 | **926/1,750/3,398** | | Super Creature minds (GOAL5: 5 Archons) | 5× composite · 12 sub-nets · 5 stages × 25 | ~10,081 weights each (~50.4k total) | | Super Creature quantum minds | 5× 6-qubit statevector + Clifford reflex | 64 complex + 16q stabilizer each | | Quantum register (1 shared, puppet-master-driven) | 5-qubit statevector | 32 complex amplitudes (256 B) | **Whole-world neural mass at the 50k mega ceiling ≈ 3.5 million parameters** (≈ 50,000 × 70 organism brains, plus the 5×~10,081-weight Super Creature / Archon composites (GOAL5), the 8 faction brains, and the apex NHI mind), stored as Float32 = **≈ 14 MB of weights**, executed on **one CPU thread**. (The default ultra tier caps at 10,000 organisms ≈ 700 K params ≈ 2.8 MB.) The Predictor-V3 size is excluded from live-world neural mass because it is not production-integrated. Its 46,080-row terminal-twin temporal task failed all eight advancement criteria. The NHI's separate 41,472-row closed-loop task retained every row but authorizes only HUNT/resource and SPAWN/social development diagnostics; its paired conflict response declined. See the [Phase-B V3 falsification report](reports/PHASE-B-MECHANISM-DEVELOPMENT-V3-2026-07-11.md). Neither result authorizes adaptation, learning, scaling, consciousness, or sentience claims. The quantum minds no longer merely _study_ the Tsotchke quantum stack — **V84 ports three of its primitives into development** and wires them into the live apex creature: the **Eshkol** qubit-RNG (`src/math/eshkol-qrng.ts`) drives the Super Creature's thought-collapse Born sample; the **Quantum- Geometric-Tensor / Fubini–Study metric** (QGTL + **Moonlab** `qgt.c`, via `src/math/quantum-geometry.ts`) reads the curvature of the mind's own thought-space; and a **spin-glass instinct** (`src/sim/spin-glass.ts`, from `spin_based_neural_network`) settles into a behavioural archetype each beat. All three are seeded + deterministic, unit-tested, and MIT-attributed to © 2024–2026 tsotchke (see [NOTICE.md](../NOTICE.md) + [THIRD-PARTY-NOTICES.md](../THIRD-PARTY-NOTICES.md)); the 64-amplitude statevector simulator itself (`src/math/quantum.ts`) remains the project's own Moonlab-style implementation. See [AI-SUBSYSTEM-2026-06-26.md](AI-SUBSYSTEM-2026-06-26.md) for the full quantum-mind design. **Super Creature 1.1 (V89) — the consciousness-metrics layer.** Atop those substrates, the apex mind now measures itself against the two leading _scientific_ theories of consciousness each beat, both deterministic and unit-tested: a **Global-Workspace ignition** (GNW — a winner-take-all over the 7 plan-coalitions that, on crossing an access threshold and dominating the runner-up, is "broadcast" and gates memory consolidation) and an **Integrated-Information Φ proxy** (IIT — the participation/coherence ratio of 8 named module activations; a _tractable surrogate_, since true Φ is intractable + non-unique). Both ride on the `Consciousness` snapshot and render as the **Ignition / Φ** meters on the SuperCreature board. The real 2023–2026 research grounding (the Cogitate IIT-vs-GNW adversarial test, organoid "wet computing", active inference, quantum cognition) is catalogued with citations in [SUPER-CREATURE-RESEARCH-2026-06-26.md](SUPER-CREATURE-RESEARCH-2026-06-26.md) — framed as _"models / inspired by"_, never _"is conscious."_ **Super Creature 1.1 — the multi-pillar expansion to ~20 coupled faculties.** Beyond ignition + Φ, the apex mind now runs roughly twenty deterministic, unit-tested faculties each beat, each reading from AND writing to the others (a negotiated plan-vote, not parallel gadgets): **Active Inference** (Friston FEP — variational + expected free energy), a **Metacognitive Executive** (Higher-Order confidence → control), a **Successor Representation** predictive map (Dayan/Stachenfeld), **Doya neuromodulation** (DA/5-HT/NE/ACh), an **Empowerment** drive (Blahut–Arimoto channel capacity), **Theory of Mind**, an **echo-state Reservoir**, **Neural Criticality** (edge-of-chaos σ̂→1), a **spin-glass instinct**, a **VSA/HRR holographic memory**, and a quantum core wired into the decision loop — the genuine statevector **min-cut Φ**, **Quantum Reservoir Computing** (Fujii–Nakajima), a **Lindblad/GKSL deliberation qubit**, **Grover** amplitude amplification, and **Quantum Natural Gradient** descent on its own Fubini–Study geometry. The Aaronson–Gottesman **Clifford stabilizer tableau** (ported from Moonlab; 16 qubits) is a fourth MIT-credited primitive. **Current measured cost (2026-06-26):** `SuperMind.think()` is **1.99 ms** in the full bench suite and **9.77 ms** for the 5× `think()` batch; the old sub-millisecond figures are superseded. Full frontier assessment in [docs/reports/](reports/) and current numbers in [BENCHMARKS-2026-06-26.md](BENCHMARKS-2026-06-26.md). ### 7.2 Contrast: this world vs. large language models | Model | Parameters | × larger than this world | | ---------------------------------- | ----------- | ------------------------ | | **This entire 50k-organism world** | **≈ 3.5 M** | 1× | | BERT-base | 110 M | 31× | | Llama-7B | 7 B | 2,000× | | GPT-3 | 175 B | 50,000× | The simulation's _entire_ population mind is **1/50,000th of GPT-3**. A single organism's "brain" (70 weights) is smaller than a textbook perceptron demo. The point of this project is that **emergent intelligence is engineered, not downloaded.** --- ## 8 · Memory, performance & hardware footprint | Resource | Requirement | | ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **GPU** | Any **WebGL2** device — integrated Intel / Apple / Android included. Quality tiers auto-scale (DPR cap, shadows, entity cap, instancing) down to phones. No discrete/RTX GPU required. | | **NPU / AI-accelerator / TOPS** | **Zero.** No transformer inference. All organism-brain forward passes ≈ 10,000 × ~140 FLOP × 60 fps ≈ **~0.08 GFLOP/s** — negligible on one CPU core. (The optional copilot chat calls a **remote** free LLM; that model never runs on-device.) | | **CPU** | Any modern CPU. The live POWER runtime is intentionally single-threaded: the wilderness Web Worker lane is dormant after worker-enabled runs showed multi-second delivered-frame stalls that the nominal FPS label concealed. A `perf-budget.test.ts` enforces the CPU-step budget; worker re-entry additionally requires a production-shaped delivered-frame gate. | | **RAM** | ≈ **300–600 MB** resident at full population (Three.js scene + typed-array entity state + instanced buffers + audit ring). | | **Storage / download** | `dist/` is 109 MB on disk (the full multi-page build), but the **app's initial payload is the ~960 KB entry chunk + Three.js** (single-digit MB; ~1–1.5 MB gzipped). The bulk — **ten 3.2 MB chunks (~32 MB) of Mermaid** — loads only on `/docs`, plus ~12 MB of self-hosted-font CSS. Routes load only what they need, gzipped. | Complexity classes are catalogued in [COMPLEXITY-2026-06-26.md](COMPLEXITY-2026-06-26.md); hot-path benchmarks in [BENCHMARKS-2026-06-26.md](BENCHMARKS-2026-06-26.md) (run via `bun run bench`, mitata). --- ## 9 · Security & compliance - **Read-only copilot sandbox** (`server/ai-sandbox.ts`): default-deny, repo-confined; blocks `.env*` / `.git*` / `legacy` / `node_modules` / `dist`; allow-listed binaries; deny-listed tokens (incl. `find -delete` / `-exec`); Git diff requires confined literal paths after `--` and rejects repository root, revisions, pathspec magic, and globs; shell-metacharacter filter; secret-free subprocess env; turn cancellation is threaded through Git-grep walks. The copilot routes are **OFF by default in every environment** and require explicit `COPILOT_ENABLED=1`. - **Server:** HTML-escaped HTMX swaps (no stored XSS), body size limits, a fixed server-side LLM provider allow-list (no client-controlled SSRF), a 200-entry in-memory audit ring. - **Supply chain:** private package under the repository's custom non-commercial license; automated Dependabot PRs disabled by binding policy; push-diff dependency review, full-tree audit, CodeQL `security-extended`, cross-platform CI, CycloneDX SBOM, checksums, and release provenance. License: **Non-commercial research & play (© 0thernes, patent-pending, commercial rights reserved); study / run / modify / share non-commercially with attribution, no for-profit use**. --- ## 10 · Build & deployment - **Build:** `bun scripts/build.ts` → `Bun.build` of three HTML entrypoints (`index.html`, `docs.html`, `specs.html`), minified, Tailwind-compiled → `dist/`. Bundled asset references are relative, so the output is subpath-portable. - **Pages assembler:** `bun scripts/build-pages.ts` copies `dist/` → `site/`, adds the `/lab` artifact, and rewrites the absolute nav links (`/docs`, `/spec`, `/lab`) to subpath-relative for GitHub _project_ Pages. - **CI/CD:** GitHub Actions — `ci.yml` runs the full gate on every push; `pages.yml` builds + deploys the real app to GitHub Pages (source = GitHub Actions); `release.yml` packages a tarball + SBOM on `v*` tags. --- ## 11 · Quality engineering The single gate — `bun run check` — must pass before every commit: `prettier --check` → `tsc --noEmit` (strict) → `oxlint` → `bun test` (**3,631 tests, 0 fail**) → `bun scripts/build.ts`. Receipt regression guard: coverage must stay within the `verify:receipts` tolerance from the canonical **84.64% line / 82.21% func** floor; local Windows receipt runs may measure higher. Three governing "master" personas (`masters/*.xml`) encode the discipline: **the Executor** (finish everything, full gates), **the Architect** (contracts before code, exclusive ownership), **the Physicist** (determinism, measurement, frame budgets, provenance). --- ## 12 · Per-file line counts A full per-file roster is generated, not hand-kept — run `bun run metrics` ([`scripts/codebase-metrics.ts`](../scripts/codebase-metrics.ts)) for the current per-area / per-type breakdown, or `bun run filemap` for the file tree. Below is the **2026-07-14** measured snapshot of the heaviest files (where the weight sits); refresh it with the same commands when source lines move. ### `src/` — heaviest files (327 files · 131,572 lines total; top of the list) | Lines | File | Lines | File | | ----: | --------------------------- | ----: | -------------------- | | 7076 | world.ts (composition root) | 2077 | sim/super-mind.ts | | 3036 | sim/crystal-ecosystem.ts | 1863 | sim/titans.ts | | 2442 | styles/app.css | 1855 | sim/big-tree-zone.ts | | 2430 | ui/observatory.ts | 1847 | ui/super-neural.ts | | 2259 | sim/alien-flora.ts | 1605 | sim/entities.ts | | 2249 | sim/apex-brain.ts | 1516 | sim/super-body.ts | ### `tests/` — heaviest files (367 files · 63,495 lines total) `big-tree-visitors 988 · worker-pool 899 · big-tree-fauna-visitors 841 · singularities 837 · titans 786 · nhi 769 · crystal-ecosystem 745 · big-tree-zone 682 · observatory 670 · xenomimics 669 · shoggoths 617 · alien-flora 569 · operational-organism-intelligence 568 · phase-b-predictor-development 527 · organism-intelligence-v4-benchmark 510` · … (full list via `bun run metrics`). --- ## 13 · Positioning This is a ~131.6k-line app-source (+ optional C++/Jolt native engine) browser-native simulation that renders **up to 50,000 agents** and targets 10,000 at 60 fps on supported laptop iGPUs with zero AI accelerator. The final headless SwiftShader smoke establishes liveness, not that native-GPU target. It is **bit-reproducible from one seed**, ships through a **full CI/CD gate** (3,631 tests, 84.64% line / 82.21% func coverage), and whose entire emergent intelligence weighs **≈ 14 MB — 1/50,000th of GPT-3** at the mega ceiling. It demonstrates that depth comes from **architecture, determinism, and engineering discipline**, not parameter count or hardware. **Frontier assessment:** [VERIFICATION-ANALYTICAL-DATA](./VERIFICATION-ANALYTICAL-DATA.md) · [Super Creature Research](./SUPER-CREATURE-RESEARCH-2026-06-26.md).