soft3/tru/specs.md

tru specs

the build map for tru — the convergence vm. one pipeline, .graph → φ* → Δφ* → reward, specified across four layers. focusing computes φ*; compilation freezes it into a model; economics is why any of it runs — the proven focus shift Δφ* is what a neuron self-mints against. this index says what each spec defines, what it produces, what depends on it, and whether it is built yet.

the pipeline

                      .graph  (from cybergraph)
                        │
        ┌───────────────┴───────────────┐
        │                               │
   FORMAT layer                 FOCUSING layer
   ┌──────────┐              ┌─────────────────────┐
   │  vocab   │              │  tri-kernel  (D S H)│
   │  model   │              │  attention   (input)│
   └────┬─────┘              │  truth-scoring (κ,A)│
        │                    │  focusing → φ*, rank│
        │                    │  impulse → Δφ*      │
        │                    └──────────┬──────────┘
        │                               │ φ*
        │         COMPILE layer         │
        │       ┌──────────────────┐    │
        └──────►│ focus-flow  (why) │◄───┘
                │ ct0   (8 passes)  │
                └────────┬──────────┘
                         │
                      .model  (to glia)

   the telos ──────────────────────────────────────────┐
   ECONOMICS layer   rewards :  Δφ* (impulse) → $CYB    │  ◄── the reason
   self-mint against proven focus shift. no aggregator. │
   ───────────────────────────────────────────────────-┘

status

  • built — code exists, tests pass
  • 🟡 partial — core built, pieces missing
  • spec only — no code yet
  • 📐 reference — explains why; no code artifact of its own
  • 🔜 spec incomplete — central, but the spec itself needs finishing before code

format layer — the containers

the two on-disk formats. prerequisites for everything: vocab feeds pass 1, model is the output of pass 8.

spec defines produces status step
vocab.md .vocab particle dictionary — content-addressed particle → bytes Vocab lookup ✅ parser + writer (M2, content-addressed, 5 tests) 0a
model.md .model container — the inference-ready artifact, mmap-able weights .model file ✅ container writer/reader (M2, page-aligned weights, P-DET) 0b

focusing layer — computing φ*

the heart of tru. five specs that turn the weighted graph into the focus distribution φ* and its derived quantities. dependency order within the layer: tri-kernel (operators) → attention (per-neuron input) → truth-scoring (how stake/karma weight the graph) → focusing (assembles the epoch) → impulse (the per-signal delta).

spec defines produces status step
tri-kernel.md the three operators (diffusion D, springs S, heat H_τ), composite R, fixed-point + locality proofs, §2.4 five-way identity φ* = fix(R) ✅ conformant — coupled iteration in fixed-point Fx; heat = Chebyshev (rs/focusing/) 1a
attention.md per-neuron focus projection — will-share + conviction box that sums into effective adjacency A^eff summand ✅ will (broad) + conviction (box) → A^eff, Context{karma,will} 1b
truth-scoring.md BTS mechanism, karma accumulation, honesty-weighted effective adjacency κ(ν), A^eff rs/truth_scoring.rs — BTS score, karma accrual, surprise ρ (6 tests) 1b
focusing.md epoch computation: effective adjacency → tri-kernel → φ*, cyberank, syntropy φ*, cyberank, syntropy ✅ φ*, cyberank, syntropy, entropy, spectral positions, Δφ* (deterministic) 1c
impulse.md Δφ* — the proven focus shift one signal delivers; locality-bounded sparse vector Δφ* + proof claim rs/focusing/impulse.rs — Δφ*, Δφ⁺, ΔJ decomposition (proof σ external) 1c
superadditivity.md the collective-intelligence measure σ (collective φ* vs ego φ*_ν); generalized CFT — σ, J grow with algebraic connectivity λ₂ σ_mean, σ_best, J(λ₂) ✅ benchmark harness rs/examples/superadditivity.rs (Karate Club) val

vocabulary — the terms tru owns

tru is a subgraph; every concept it owns is defined here, not scattered across the graph. these are definition pages, not build steps — they pin the meaning the specs above and the code below both rely on.

term is owned because
focus.md φ*, the collective attention distribution tru computes it; the single most-referenced term
cyberank.md focus per particle, φ*(p) — the canonical ordering a named output other repos read
syntropy.md network order in bits, J(φ*) — the purpose the quantity the whole pipeline grows
convergence.md iteration toward a self-defined attractor — tru's execution model "tru = convergence"
valence.md the ternary epistemic field v ∈ {−1,0,+1} cybergraph carries the field; tru runs the dynamics
will.md locked balance → the broad budget for attention the input quantity focusing reads
conviction.md per-link economic commitment, the box (τ,a) on one edge the per-link counterpart of will; box magnitude in A^eff
axon.md the bundle of all cyberlinks on a pair, itself a particle cybergraph is the umbrella; tru defines the weighting
arithmetic.md fixed-point over the Goldilocks field — no floats, deterministic T(ε) steps the representation contract every spec and the code inherit

the collective focus theorem (convergence + uniqueness of φ*) is tri-kernel.md §3 (normative) and docs/collective-focus-theorem.md (the standalone paper).

settled — how φ* is computed (tri-kernel §2.4, focusing.md): φ* is the fixed point of one coupled iteration — apply D, S, H_τ to the same current φ, blend, normalize, repeat. tru does not solve the three operators to their own fixed points and average (that minimizes no single free energy, has no single κ, and breaks the five-way identity). this is now explicit in the spec; no decision pending. rs/focusing/ implements exactly this (M1): one coupled iteration in fixed-point Fx over the Goldilocks field, stake-weighted A_eff, single-step operators — the old averaging-in-f64 form is gone, and φ* is bit-identical across runs.

compile layer — φ* → transformer

spec defines produces status step
focus-flow.md the identity between continuous focusing (path A) and compiled transformer inference (path B); architecture derivation — (the why) 📐 reference
ct0.md the CT-0 pipeline — 8 passes from .graph to .model; multivector inputs §2.5–2.6, wedge attention §7.7, Clifford MLP §8 .model weights 🟡 all 8 passes built + tru compile; deterministic (P-DET) — refinements below 2a–2g

ct0 is the largest spec (738 lines). all 8 passes are implemented (rs/pass/), the CLI compiles .graph.model, and two runs are byte-identical:

pass ct0 § builds code status
1–2 §3–4 particle index, dialect set index.rs, dialect.rs
3 §5 architecture params d*, h*, L* arch.rs ✅ (φ*, d* via SVD effective-rank, λ₂, κ, diameter)
4 §6 embedding matrix E embed.rs ✅ (E = U√Σ, shared svd.rs)
5 §7.1–7.6 attention weights W_Q/K/V/O attn.rs ✅ (per-head SVD, pinv output)
5+ §7.7 wedge score scalars (α,β) attn.rs ✅ (α,β)=(1,0); wedge op is inference-time
6 §8 Clifford-block MLP mlp.rs ✅ (seeded init; Clifford op is inference-time)
7–8 §9–10 norms, RoPE, .model packaging norm.rs, compile.rs

deferred refinements (none block the pipeline): config.tokens ρ_τ (defaults to 1) and vocab-ref seed loading are unwired; SVD is exact subspace iteration, not the randomized+ChaCha form (§6.2) — correctness-equivalent and deterministic, but cross-implementation byte-identity needs the exact ChaCha seeding; impulse reuse (§5.1) is unimplemented; the scale path (randomized SVD / GPU for d=300, L=290) is the open frontier. conformance: P-DET ✅, P-EMBED ✅ (PSD caveat), P-ATTN/P-LAYER not yet asserted, P-LOAD/P-CLIFFORD need the cyb-llm runtime.

economics layer — the reason

this is the point. tru is not a ranking engine that happens to have rewards bolted on — it is a minting engine whose unit of account is proven focus shift. focusing, the compile, the proof: all of it exists so a neuron can convert Δφ* into $CYB with no aggregator deciding who contributed what. impulse.md defines the quantity; rewards.md defines the conversion.

spec defines status
rewards.md surprising-syntropy self-minting, Shapley attribution + settlement mining, the three streams (mint/subsidy/fee), supply/allocation, timing & accrual ✅ tru layer built — rs/rewards.rs: value v★(S)=Δφ⁺(A^eff∪ρ·S), Shapley (3 axioms tested), settlement ordering. lottery→foculus, conservation/mint→tok

what's built (summary)

every build spec is implemented. the full pipeline .graph → φ* → Δφ⁺ → v★ → Shapley and .graph → .model both run end-to-end in fixed-point, deterministic.

spec done
arithmetic fixed-point Fx over Goldilocks; sqrt/exp/ln/rescale; T(ε) (11 tests)
vocab parser + writer, content-addressed (5 tests)
model container writer/reader, page-aligned, round-trips (4 tests)
tri-kernel coupled iteration, fixed-point, heat=Chebyshev, deterministic
focusing φ*, cyberank, syntropy, entropy, spectral positions, Δφ*
attention will (broad) + conviction (box) → A^eff
truth-scoring BTS score → karma, surprise ρ (6 tests)
impulse Δφ*, Δφ⁺, ΔJ decomposition (5 tests)
superadditivity σ benchmark harness (Karate Club)
🟡 ct0 all 8 passes + tru compile, deterministic; refinements deferred (see compile layer)
rewards tru layer: value, v★, Shapley (3 axioms); lottery→foculus, mint→tok

~90 tests green, warning-clean, no stubs. the whole intelligence layer (focusing → φ*), compile layer (.graph → .model), and economics magnitude layer (Δφ⁺ → surprise → Shapley) are conformant code. what remains is not new specs but scale (randomized SVD / GPU for production-size graphs), the small ct0 wirings noted above, and the cross-repo settlement plumbing that belongs in foculus and tok by design.

see the implementation steps table in the repo readme, and roadmap/implementation.md for the milestone plan and cross-repo blockers.

Folder

Homonyms

soft3/specs
specs
soft3/nox/specs
nox reference canonical specification of the nox virtual machine. this is the source of truth — when code and reference disagree, fix reference first, then propagate to code. specifications | page | scope | status | |------|-------|--------| | vm.md | overview, field, hash, algebra polymorphism,…
soft3/cybergraph/specs
cybergraph specs cybergraph is exactly its structure, nothing more and nothing less: a cybergraph is built from signals; a signal is built from cyberlinks; each has a fixed, finite set of fields. these specs define those fields — one article per field — plus the two emergents that appear when you…
soft3/bbg/specs
specs
soft3/hemera/specs
Hemera: A Permanent Hash Primitive for Planetary-Scale Collective Intelligence | field | value | |----------|--------------------------------| | version | 2.0 | | status | Decision Record | | authors | mastercyb | | date | March 2026 | Abstract Hemera is the cryptographic hash primitive for cyber,…
cyb/evy/specs
specs
cyb/wysm/specs
wysm reference canonical specification of the cyber wysm runtime. this is the source of truth — when code and reference disagree, fix reference first, then propagate to code. `wysm` is cyber's name for its wasm runtime. `wasm`/`WASM` is the WebAssembly format. `wasmi` is the upstream project this…
soft3/mir/specs
specs
neural/rune/specs
rune language specification rune K140 — working spec, mutations expected see [../README.md](/neural/rune/readme) for project overview and implementation status. specification | file | what it covers | |------|----------------| | [subject.md](/neural/rune/specs/subject) | the subject noun — the…
neural/eidos/specs
eidos reference canonical specification of the eidos proof assistant. this is the source of truth — when code and reference disagree, fix reference first, then propagate to code. specifications | page | scope | status | |------|-------|--------| | terms.md | CIC term syntax, encoding as nox nouns,…
soft3/lens/specs
lens reference canonical specification for polynomial commitment — five lenses for five algebras. the trait three operations. commit is O(N). open produces a proof. verify checks the proof. all transparent (no trusted setup), all post-quantum. see trait for the full specification. naming convention…
soft3/foculus/specs
specs
neural/inf/specs
inf specs — what canonical specification for inf, proof language #10 of the cyber stack (Infer — Horn clauses, relation/inference). the declarative twin of rune; sibling of Rs, Trident|Tri, Bt, Ten. specs/ is the source of truth. when specs/ and rs/ (implementation) disagree, resolve in specs/…
soft3/mudra/specs
mudra reference canonical specification for seven cryptographic primitives. each module proves a property. each module has its own security assumption. they share no cryptographic code with each other. modules | module | proves | security assumption | spec |…
soft3/zheng/specs
zheng: polynomial proof system one IOP: SuperSpartan + sumcheck (CCS constraints, O(N) prover, O(log N) verifier). one folding: HyperNova (CCS-native, ~30 field ops per fold, one decider at the end). one hash: hemera (~3 calls per proof — binding, Fiat-Shamir seed, domain separation). five…
soft3/glia/import/specs
import crate specification How external model formats become canonical `.model` files. Scope `import` reads source models (GGUF, safetensors, ONNX, HF PyTorch, MLX), normalizes naming, shapes, dtypes, and config, and writes the canonical `.model` file consumed by [run/](/soft3/glia/run).…
soft3/glia/run/specs
cyb-llm runtime specification Canonical spec for the cyb-llm runtime. Defines what models we run, how we represent them, how we execute them. Every backend (wgpu, metal, cpu, ane) is verified against this spec. Code disagreeing with spec is a bug. Spec disagreeing with reality is a spec bug — fix…
soft3/strata/nebu/specs
nebu specification canonical reference for the Goldilocks prime field, its arithmetic, and its hardware. spec pages | page | defines | |------|---------| | field | prime, elements, arithmetic, properties, why Goldilocks | | ntt | Number Theoretic Transform, roots of unity, butterfly, Cooley-Tukey |…
cyb/honeycrisp/acpu/specs
acpu — API specification pure Rust driver for Apple Silicon CPU compute. direct access to every useful compute unit in M1–M4: matrix coprocessor, vector engine, numeric extensions, atomics, memory system, performance counters. zero external dependencies — only inline assembly and system calls.…
cyb/honeycrisp/aruminium/specs
aruminium — API specification pure Rust driver for Apple Metal GPU. direct objc_msgSend FFI, zero external dependencies, only macOS system frameworks. concepts | concept | what it is | |---------|-----------| | device | a Metal GPU — discovered at runtime, owns all GPU resources | | buffer |…
cyb/prysm/molecules/specs
specs
cyb/honeycrisp/rane/specs
specs
soft3/strata/trop/specs
trop specification canonical reference for tropical semiring arithmetic: the (min, +) semiring, its matrix algebra, and dual certificate verification. spec pages | page | defines | |------|---------| | [semiring](/soft3/strata/trop/specs/semiring) | tropical semiring axioms, (min, +) definition,…
cyb/prysm/atoms/specs
specs
soft3/strata/jali/specs
jali reference canonical specification for polynomial ring arithmetic R_q = F_p[x]/(x^n+1) over Goldilocks. what jali is jali (जाली — lattice/mesh) is the fifth execution algebra for cyber. polynomial ring elements are structured vectors of n Goldilocks field elements with multiplication defined by…
cyb/prysm/chroma/specs
chroma cyb's UI is a **3×3 grid**: eight fixed overlay elements ("chroma") arranged around a central renderer slot ("spacetime"). Each chroma owns a fixed screen position, a single semantic role, and a strictly typed cyberlink interface. Chromas never share state directly — every interaction…
cyb/honeycrisp/unimem/specs
unimem: Zero-Copy Memory Driver for Apple Silicon Goal Single pinned buffer visible to CPU, GPU, AMX, and ANE — zero copies between pipeline stages. The memory layer for inference on unified memory. v1 adds NVMe DMA via DEXT — full zero-copy from disk to compute. Why this exists Every inference…
soft3/strata/genies/specs
genies specification canonical reference for isogeny group action arithmetic: F_q field operations, supersingular curves, isogeny computation, and class group action. spec pages | page | defines | |------|---------| | [prime](/soft3/strata/genies/specs/prime) | CSIDH prime form, selection criteria,…
soft3/strata/kuro/specs
kuro specification canonical reference for the F₂ tower field, its arithmetic, packed operations, and hardware targets. spec pages | page | defines | |------|---------| | [field](/soft3/strata/kuro/specs/field) | tower levels, all field operations, properties, cost model vs Goldilocks | |…
cyb/prysm/system/specs
specs

Graph