HoloTrade — Compute Execution & Settlement

Exchange

The ask side is the pricing engine: every listed node posts an offer at its own clearing price, so every line in the book decomposes back into six auditable multipliers and a named machine. There is no market rate you cannot take apart.

SIMULATION seeded prices, fleet and order flow EXACT finite W(3,3) graph and adjacency rules PROTOTYPE no live feeds or physical substrate

Instrument

Order book

Depth

Value frontier

every listed node · $/hr against delivered TFLOPs

A node-hour is not comparable across hardware classes — a CPU node at $4 and an accelerator at $60 are not competing offers. Delivered throughput per dollar is comparable. The highlighted line is the Pareto frontier: the nodes that nothing else beats on both price and capacity at once. Everything above it is dominated, and the exchange will tell a buyer so.

Tape

Time	Side	Node	Addr	HW	Site	Qty	Price

Ticket

Workload class Quantity (node-hours) Limit ($/node-hour, blank = market) Anchor address (your data's position, optional)

Best quote, decomposed

P = base × E×G×D×H×Q×L

Your orders

ID	Inst	Qty	Filled	Status

Fleet

Every node is four things at once and all four are priced: what it is (hardware), where it sits (fabric address), what it has learned (genome), and what it has spent (health). Click any cell.

Fleet map

Listed nodes

Node	Addr	HW	Site	$/node-hr	$/TF-hr	Util	Fitness	Wear	Best at

Fabric

The modeled substrate treats topology as part of capacity, so a basket is more than a node count. Exact induced-graph facts make the difference between connected and scattered holdings visible without pretending they are physical-network SLAs.

W(3,3) — one cell

40 points · 240 edges

occupied vacant selected + neighbours

Routing — address is route

one symplectic inner product per hop

From To

Recursive instrument sketch

DESIGN SKETCH · shared quote interface

The prototype applies one quote interface — address, modeled specialization, health, utilisation and throughput — to leaves and simulated composites. That makes a self-similar instrument ladder testable in software. It does not establish a deployed recursive fabric, atomic composite reservation, or a campus that can already settle as one node.

Level	Objects	Leaves	Coherence	Model diameter bound	Quote $/hr

Cells

model treats a complete cell as one composite

Cell	Site	Nodes	Occupancy	Induced edges	Coherence

Instrument ladder H_n

40ⁿ leaves · model bound 16n−14

n	Leaves	Model diameter	Components	$/node-hr	Seats

Defragmentation — the swap book

Repeated spot fills can leave the seeded portfolio as confetti across many cells, reducing its induced-edge coherence even though every individual position looks fine. A modelled swap exchanges orphans between cells. A swap is only offered when both sides gain coherence; physical delivery and interconnect performance remain unmodelled.

Swap	Give	Get	Your coherence	Their gain	Cash adj.

Optimal reservation shapes

exact · every witness attains its spectral bound

The spectral bounds say how dense a reservation of a given size could be. A scheduler cannot reserve a bound — it reserves a set. These are the sets that actually attain the bounds, found by exhaustive search and re-checked independently by eigenspace projection. A densest shape exists at every multiple of four and nowhere else; the four-point ones are exactly the 40 totally isotropic lines.

“Survives” is a worst-case guarantee, not an average. It is one less than the minimum number of blocked points that defeats every placement of the shape — so with that many nodes busy a placement is certain, whichever nodes they are. Proved two independent ways: branch-and-bound, and SAT with every smaller size discharged as UNSAT. The four-node figure of 10 is exactly the no-ovoid result in disguise: ten points could only block all 40 lines by being an ovoid, and there is none.

Busy nodes (blocked)

0 busy · 40 free

Size m	max e(T)	min b(T)	Orbit	Survives	Placeable now	Placement

Anti-affinity ceiling

Perfectly spread set

Exact finite mathematics about the level-1 point graph, verified in tests/shape-catalogue.test.js. Placement transports a shape by an automorphism, which preserves induced edges and boundary exactly. Healthy and failure-aware recovery policies are tested separately below; link-state routing, congestion-aware forwarding, physical recovery and deployed multi-cell placement remain unbuilt.

Recovery closure · loading exact software and modeled fleet certificates…

E8 unitary elastic ladder

EXACT · loading GAP certificate

Two larger unitary carriers inherit the same nested line-atom scheduler, but neither has a full spread. Choose a carrier and rung to see the exact regular subgraph, spectral-minimum boundary, and the residual hole sector. The final rung is a certified ceiling, not full coverage.

Carrier Prefix rung

rung 1 / 6

Maximum-rung boundary geometry · loading independent GAP/GRAPE certificate…

Cross-prime fibre and logic switch · loading exact quotient certificate…

Backend closure · loading stable-module, affine-lift, canonical-fibre, fibre-router, protected-qutrit-control, Payne five-state atlas with one-hop repair, reversible-circuit, recursive-fabric, homology-code, bicolour, Goursat, and sentinel certificates…

Carrier	Line atoms	Vertices	Induced degree	Internal edges	Boundary	Holes now
loading certificate…

The standard Hermitian incidence model is an abstract topology plan. GAP point IDs are not live hosts; inventory binding, physical links, routing, and runtime topology attestation remain required before dispatch.

Neighbour averaging

toy scalar model · not consensus

This is a neighbour-averaging recurrence on the exact level-1 graph. It does not implement leader election, settlement, consensus, crash tolerance, Byzantine tolerance, networking energy, or a physical holonet.

Balance

The seeded demand-response model applies a premium above the target utilisation band and a discount below it. Use the toggle to compare the resulting utilisation distributions; this is a simulation outcome, not an operator savings claim.

Paired simulation certificate

SIMULATION · frozen 64-seed packet · same seeds in both arms

Balancer on

0.06458

Gini · 95% CI 0.06043–0.06874

Balancer off

0.16110

Gini · 95% CI 0.15455–0.16764

Paired reduction

60.34%

95% CI 58.90–61.77%

Seeds improved

64 / 64

220 nodes · 500 × 60s steps

Checked-in deterministic model result, not field telemetry or an operator savings claim. Reproduce with npm run experiment:balancer. Row digest: 1cdc21978a17…b8a.

Two-sided pricing on/off

Balancer ON

Watch the Gini in this run. Nothing in the loop pushes utilisation toward the band directly. A node's target is set by its price relative to the median of its hardware class, and that price came from its utilisation. Toggle the model and compare its current seeded trajectories.

Utilisation distribution

How the wear model treats dispersion

Model assumption. The seeded wear function penalises utilisation swings more than steady load at the same average. Smoother load therefore accumulates less modelled wear and spreads modelled service events. HoloTrade has not measured a hardware lifetime improvement.

The demand multiplier D

premium above band · discount below

Health & Wear

The seeded model converts use, cycling and service state into a visible health multiplier and maintenance reserve. These are policy inputs for exploring a wear-aware market, not telemetry from deployed hardware.

Weibull hazard

wear-out dominated, shape k = 2.4

Service queue

Taking a node out of service costs its owner the revenue it would have earned — which is exactly why the balancer's discount window matters. It creates the trough you schedule service into. Service recovers most of the derate but not all of it: wear is partly permanent, silicon does not go back to new.

Wear ledger

RUL = remaining useful life

Node	HW	Site	Wear	Derate	Hazard	RUL (h)	Cycles	CE	Reserve $/h	Status

Genetics

Spot node-hours are a commodity and commodities converge to marginal cost. Trained cores are not: they are heterogeneous, their quality is observable, and their value compounds with use. That is where the spread lives.

Fitness leaderboard

#	Node	HW	Score	Fitness	Gen	Jobs

Bloodlines

founder cores by descendants

Founder	HW	Descendants	Founder fit	Line mean	Line best

Breeding console

prediction stated before the cross

Crossover of two high-fitness cores onto fresh hardware — the software operation of merging two checkpoints and warm-starting from the result. The child's expected fitness is stated before the cross so it can be checked afterwards, and it comes in below the better parent on purpose: a merge is lossy and a fresh core has no track record. The exchange must not credit it with its parents' completions.

Parents	Target	Distance	Target now	Predicted	Gain

Crosses made

Target	Predicted	Before	After	Gen

Where the priors sit

fleet census by strongest class

Workload class	Nodes strongest here	Mean score	Share

Energy

Compute is bought in dollars and paid for in joules. You cannot price per-second energy if your settlement unit is an hour — which is the real reason the tradeable unit here is the node-second.

Wholesale grid price

$/MWh · seeded simulation

Sites

Site	Grid	$/MWh	gCO₂/kWh	PUE	Nodes	Load kW	E mult

Hardware energy inputs

catalogue assumptions · not measurements

The J/op values below are seeded catalogue inputs used by the pricing model. The UI does not compare them with a per-cycle thermodynamic quantity: those units and operation boundaries are not interchangeable.

Hardware class	J/op	Status

Execution

A complete single-node lifecycle simulation: the node is the asset, the plan is the contract, the node-second is the settlement unit. No real microVM or workload is launched by this browser prototype.

New execution plan

Name Workload class

Seconds Nodes (prototype lifecycle: one)

Declared artifact digest (bound into demo plan) Egress grants (blank = deny all, no network device)

Placement

ranked by total cost, not headline rate

The placement model itemises a declared cold-start distribution and modeled pin-validation work once instead of hiding it in the rate. This demo does not boot a kernel, mount a verity root, or hash artifact bytes; those are production integration requirements.

Node	Addr	HW	$/node-s	Boot	Migration	Total

Simulated microVM runs

node-second ledger · clipped final interval

VM	Plan	Node	Uptime	Spend	kWh

Plans

Plan	Digest	Class	t	Node-s	Status

Audit chain

#	Event	Detail	prev→hash

Nested state namespace

illustrative grammar count · not runtime capacity

Under the illustrative copy-on-write grammar, untouched siblings retain their digests, so a uniform root at depth n denotes the address tree from n+1 model blobs. At level 6 the arithmetic counts 4,096,000,000 leaf labels plus 105,025,641 internal labels. It does not instantiate 4.2 billion VMs, allocate their state, or establish runtime density.

Level	Leaf labels	Internal labels	Total labels	Model blobs	Distance bound

Portfolio

Positions, marked against current simulated quotes. The capacity roll-up is in effective TFLOP-hours, which is the number that matters to a buyer — not the node count.

Positions

Node	Addr	Inst	Class	Qty	Avg	Mark	Value	P&L

Your shape

coherence, not count

Fills

Time	Side	Node	Qty	Price

Assets & UOR

Each asset carries explicit delivery policies. The dashboard turns documented policy defaults into an estimated market-breadth score across forty reference regions; it is not observed liquidity or a group-theoretic co-volume.

The address

2⁶⁴ = 40 × 1,296 × payload

The forty vertices of W(3,3) are in canonical bijection with the forty Sylow-3 subgroups of Sp(4,𝔽₃), so an object reference is a pair (P, x): a Sylow subgroup and a coset representative. 40 × 1,296 = 51,840 = |Aut(W(3,3))|. An asset's identity and its position in the fabric are the same 64 bits — there is no separate registry mapping ids to locations that can go stale, and no lookup to do.

Policy-estimated market breadth

configured defaults · not tape liquidity

Each default estimates the fraction of forty reference delivery regions in which an asset remains eligible. Multiple policies use the tightest score, avoiding an invented independence assumption. These are configurable model inputs unless a venue supplies measured eligibility.

Policy	Mobility score	Eligible regions	Band

Rank-3 shell from an anchor

1 + 12 + 27 = 40

Anchor node

Every listed node falls into one of three geometric relations to the anchor: identity, intersecting, or disjoint. The 1:12:27 ratio describes the forty reference points; a listed fleet may place zero, one, or many nodes at each point, so its observed shares need not match it.

Tier	Nodes	Fleet share	Geometry ratio	Δ

Illustrative rate transforms

dimensional displays · not ceilings

Transaction rate (TPS)

The displayed rates apply documented dimensionless factors to the TPS slider. They are illustrative transforms, not measured throughput, protocol limits, projections, or service-level objectives. Full reference scan time is simply 51,840 transactions divided by the supplied transactions per second.

Receipts

Every simulated fill emits a deterministic quote sample carrying node and model provenance. No workload has run at fill time, so these records are not execution, delivery, verification, or attestation evidence.

Evidence boundary

SIMULATED_QUOTE_RECEIPT

Substrate-labelled sample. The contextual fraction is generated by the seeded quote model. It is not a photonic measurement and carries no clean/dirty verdict.

Clifford-labelled sample. The lane is catalogue classification only. No stabilizer trace is produced or checked by a fill receipt.

Simulated quote receipts

Receipt	Node	Addr	Lane	Synthetic CF	Label	Qty	Price	Gen	Model provenance

The Model

What HoloTrade sells, how it prices it, and which surfaces are exact, formal, synthesized, simulated, prototyped, or still design work.