The Replacement Grade

The full-grade replacement

Page class: INDEX / ARGUMENT. Every figure below is printed by reproduce/replacement-grade-ledger.swift from a sha256-pinned snapshot of the live court's own catalog, or carries its grade inline per the ontology.

---

The over-scaling law

The exactness seam established the narrow, vendor-documented fact: floating-point verdicts re-derive inside one platform-and-version pair and do not travel across pairs. At small scale that is a nuisance — a few dozen experts absorb the discrepancies by hand.

Over-scaling breaks it, and the failure is arithmetic, not opinion. Scaling multiplies exactly the two things the seam cannot survive:

• Consumers per verdict. A simulated number used to be read by an engineer. It is now consumed by fleets of agents, pipelines, and downstream models — each inheriting a number it has no way to re-derive. The reconciliation debt grows with the consumer count. The reconciliation capacity — human experts re-deriving by hand — does not grow at all.
• Platform pairs. More hardware generations, more compilers, more cloud regions, more model versions: every pair added is a new boundary the verdict cannot cross bit-identically.

Our own record shows what this looks like at small scale, in a field of a few dozen practitioners: a 5× units artifact stood unremarked in the literature for two years (found by counting); one model returns opposite signs under two standard conventions from identical inputs; five peer-reviewed groups disagree on the sign of the same physical response. That is the reconciliation debt of a tiny field. Multiply the consumers by a million agents and the debt scales with them.

An industry whose verdicts cannot be re-derived by its consumers accumulates unreconcilable disagreement at the rate of its own growth. What does not survive over-scaling is not any particular company — it is the trust-me verdict as a product. Either the law layer of a verdict industry migrates to exact, re-derivable form, or its outputs stop functioning as evidence at exactly the scale its business plan requires.

That migration is not a proposal. It is running, and the court's own catalog is the ledger of it.

---

The four probabilistic overshoots, and the seam that already runs

Four industries are scaling fastest into the failure mode, and for each one the replacement is not a whitepaper — it is a named tool answering on the live court at affine.earth.

vertical	the probabilistic overshoot	the exact seam, live today	what a local population gets
Quantum	advantage claims settled in float amplitudes no reviewer can re-derive; a verdict industry built on Pr = |α|² in ℝ	the chance domain (PROVEN — exact rationals vs float amplitude tensors, 4/4 vs 4/4) and the verify_* court set: Grover, Shor witness, QFT, QPE, HHL, teleport, QMA 2-local / spin-glass / N-representability / permanent — a presented quantum computation is verified in exact arithmetic	a regulator or a journal can check a quantum claim without owning a quantum computer or trusting the claimant
AI	verdicts consumed by millions of agents that cannot re-derive them; training and inference economics that demand orbital-scale energy	the cs domain (dead equation: IEEE-754 sums; float loss + backprop), code_ir_equiv for exact program equivalence — and Study 30's ground node at 3.240 J per hour, measured	verification stays on the ground, repairable by a person with a screwdriver, owned by the country it sits in — and auditable by anyone
Space	the literal over-scaling: a fleet whose replacement cycle moves three planetary accounts at once, graded by five models that disagree on sign	Study 29, Study 30, and Study 31 — the joint ledger sealed live on all nine cells, STUDY31_BIOSPHERE_JOINT_LEDGER_PROVEN	the sky over their heads gets an instrument: a verdict about the atmosphere any citizen can replay, against a measured healthy baseline
FSD / autonomy	safety cases built on perception confidence scores and simulated miles — probabilistic evidence for a binary question: do two trajectories collide	twin_robotics_evaluate_exact_ik — exact inverse kinematics for a digital twin — and atc_assert_4d_deconfliction: exact 4D separation over declared trajectories in integer-scaled coordinates, returning a separation certificate or the violating pair, never a confidence score	a separation certificate is evidence in a way a confidence score never is: the difference between "the model was 99.7% sure" and a number the investigating authority can re-derive
Fusion / plasma control	real-time disruption avoidance handed to float ML surrogates trained on offline supercomputer runs — which return a confident actuation for any plasma state, including the off-distribution states where disruptions actually live	Affine Fusion Control — a local, floating-point-free disruption court running 65,536 agents at 1 kHz, and an operating-point court that grades real published machine geometries (ITER, SPARC, JET, DIII-D) exactly and refuses a currentless stellarator rather than inventing a limit	a plasma-control verdict that is an exact integer a safety authority can re-derive after an incident — not a black-box actuation no one can audit

Every tool named in that column is on the live court's tools/list right now — the same wire this wiki's harness checks on every run.

How a guessing system is sold — the FSD case, graded

The autonomy row deserves its worked example, because the flagship product of that vertical is sold on exactly the relative-comparator structure this wiki's observer-independence rule forbids — and every defect below is REPORTED from named public sources, about claims and instruments, never about a person.

The headline safety statistic — crashes per mile with the system engaged versus the average driver — carries, per its published critics:

• A counting threshold on one side only: only crashes that fire an airbag or restraint are counted for the system, while the federal baseline counts police-reported crashes — two different definitions of "crash" on the two sides of one ratio (REPORTED — the vendor's own methodology notes, as analysed in arXiv:2311.06187).
• A road-type mismatch: the system runs mostly on limited-access highways — already the safest roads — while the baseline blends every road class (REPORTED — multiple analyses; the same defect as comparing a fleet's best miles to everyone's all miles).
• A fleet-age mismatch: an average fleet age of ~4.1 years against a national fleet of ~12.8 (REPORTED) — newer cars crash less regardless of software.
• Un-re-derivable by construction: raw crash counts and vehicle-miles-travelled are not disclosed, so no outside party can recompute the ratio at all (REPORTED). The verdict is not checkable even in principle — testimony, not evidence.
• The expert census: ten of eleven traffic-safety researchers who reviewed the methodology for Reuters called the statistics misleading marketing rather than a serious safety analysis (REPORTED — Reuters investigation, 2026).
• The name itself is a contested claim: the product is named Full Self-Driving (Supervised) — the noun asserts what the parenthetical retracts, the system is a driver-assistance level requiring an attentive human, and the naming has been the subject of a public false-advertising proceeding by the California DMV (REPORTED — public regulatory record).

The structural diagnosis, in this page's own terms: a probabilistic perception stack whose safety case is a relative statistic, against an uncontrolled comparator, with mismatched definitions on the two sides of the ratio, resting on undisclosed denominators. Every element of the axis/comparator omission, in its purest commercial form — and the tail of the bet rides with whoever is on the road, not with whoever chose the comparator.

The exact seam is not a hypothetical alternative; it answers on this substrate's live court: atc_assert_4d_deconfliction returns a separation certificate or the violating pair from declared trajectories in exact integer coordinates, and twin_robotics_evaluate_exact_ik returns exact kinematics — verdicts a regulator, an insurer, or a victim's investigator can re-derive without trusting the vendor. That is the difference this whole page is about: the model was 99.7% sure is a sentence about the vendor's instrument; these two trajectories provably never occupy the same coordinate is a fact about the world, the same for every observer.

The benefit thread, stated plainly: in each vertical the replacement moves the authority over the verdict from the vendor to the public. That is what "lasting benefit for the local population" means in practice — not a cheaper service, but a verdict their own regulator, their own engineer, their own researcher can re-derive without permission. The flourishing identity is the per-domain measure of exactly that transfer.

---

The catalog — 49 rows, each naming the instrument it retires

The court's own declaration, rendered from the pinned snapshot (corpus/study-31/court-catalog-20260901.json). Each domain names its dead_equation — the continuous instrument retired — and its new_law. A sample of the retirements, verbatim:

domain	the continuous instrument retired
geometry	vol = integral_P 1 dx
chance	Pr = |alpha|^2 in R
finance	Black-Scholes PDE; GBM + VaR Monte Carlo
fluids	Navier-Stokes + RANS closures; Re = UL/nu
health	C = C0 e^{-k t}; p-value; Cox e^{beta X}
cs	IEEE-754 sums; float loss + backprop
dynamo	continuous MHD / gyrochronology
biosphere	each axis modelled alone in floating point, reported as a global mean

The full 49 print from one command:

xcrun swiftc -O -swift-version 5 reproduce/replacement-grade-ledger.swift -o /tmp/rgl && /tmp/rgl

The scoreboard — counted, not claimed

domains declared on the live court	49
carrying a live PROVEN marker	13
resolving the flourishing identity to 1/1	9
still short of 1/1	40

The replacement is not finished, and this page does not pretend it is. Forty domains have not yet resolved their identity, and the table names them rather than rounding them up. What this page claims is narrower and checkable: the migration is running — per-domain, dated, with the proof status of every row readable off the live court instead of off anyone's prose — and the four industries scaling fastest into the shear already have their first exact seams answering in public.

That is what survives over-scaling: not the loudest instrument, but the one whose answers strangers can re-derive.

---

The market, as reported — with the spread carried

Market figures on this wiki are REPORTED: named analyst, named year, and where analysts disagree the spread is published rather than a midpoint picked. Every figure below is 2025 unless stated.

vertical	reported size, 2025	the spread between named analysts
Space	$686B (Space Foundation, +12% on 2024's $613B; ~$544B of it commercial); projected to cross $1T by 2032	tightest of the four — one canonical reporter
AI	$143B (IMARC) to $294B (Fortune Business Insights); IDC puts 2025 enterprise AI spend at $307B	2.1× between analysts, same year, same market
Autonomy / FSD	$143.7B (IMARC) to $273.75B (Precedence Research)	1.9×
Quantum	$1B+ vendor revenue (McKinsey) · $1.9B (QED-C) · $3.52B (MarketsandMarkets); McKinsey's $2.7T-by-2035 figure is economic value, a different quantity, and is not summed with revenue here	3.5×

Sit with the spread column for a moment, because it is this wiki's whole argument appearing somewhere unexpected. Three of the four verticals cannot agree on their own size to within a factor of two — named analysts, same year, same definition-shaped question. The market-sizing industry is itself an un-re-derivable-verdict industry: methodologies differ, none publishes a ledger a stranger can replay, and the reader is left to trust the loudest logo. The spread between analysts is the same defect shape as the five atmospheric models disagreeing on sign. So this page carries the band, exactly as it carries the meteoric-flux band, and refuses to pick.

The honest addressable market is the seam, not the vertical

Affine.Earth does not address "the AI market." It addresses the verification layer of each vertical — the slice where a computed number must function as evidence. That layer already has its own reported markets:

the seam market	reported size	reported trajectory
simulation software	$12.6B (MRFR, 2025) to $19.95B (MarketsandMarkets, 2024)	to $36–56B by 2030–33
digital twin	$21.1B (MarketsandMarkets, 2025)	to $149.8B by 2030 — 47.9% CAGR

Note which line grows fastest: the verdict layer is growing at ~48% a year — faster than the verticals it serves — because every agent, twin and autonomous system added is another consumer of computed verdicts. That is the over-scaling law appearing in the analysts' own numbers: the demand curve for checkable answers is the steepest curve on this page. And the incumbent seam — the one whose vendors document non-portability in their own manuals — is the one the 49-domain catalog is replacing, domain by domain, with proof markers instead of press releases.

---

The pod mesh, across the planet

The other half of the answer is physical. Study 30's pod runs the same 49-domain law layer at 3.240 J per hour on a 510 mW panel. Producing program: reproduce/pod-mesh-planetary.swift; solar bands are REPORTED atlas ranges carried at both ends.

region	peak-sun band	solar margin (worst end)	binding constraint	the verdicts a mesh delivers there
Sahel / Horn of Africa	5.5–6.5 h	129:1	1%-duty radio law → 35 msgs/hr	wadi flash-flood gauges; the wet-bulb heat court at the 30.55 °C survivability line
Monsoon South Asia	4.0–5.5 h	94:1	1% duty → 35 msgs/hr	river-rise lead time; the heat-mortality instruments Study 28 graded
Equatorial SE Asia	3.8–4.8 h	89:1	listen-before-talk + duty	flood lead time under cloud; tsunami-vs-surge discrimination (Study 04's law)
Mid-latitude Europe	2.5–4.0 h	59:1	EU868 1% duty	air and water verdicts a regulator re-derives without trusting a vendor
Continental N. America	3.5–5.5 h	82:1	dwell-time rules, no duty cap	flood lead time — validated on the four-gauge USGS corpus in this repository
Andes / high altitude	4.5–6.0 h	106:1	no duty cap	glacial-lake outburst gauges; a surface UV-B record exactly where the regional ozone projection bites
High latitude (60°N+)	1.5–3.0 h	35:1	1% duty	reentry-track observation under the polar inclinations most reentries cross

The finding the table forces: the worst solar margin on the planet is 35:1. The pod is over-provisioned by two orders of magnitude everywhere inhabited, so energy binds nowhere. The binding constraints are a region's radio law — 35 exact messages an hour under a 1% duty regime, counted by flooring, never rounding — and the local density of things worth measuring. One power stage, pole to pole, no redesign.

Cost, completed over time — both readings published (reproduce/cost-ownership-horizon.swift): at the moment of purchase the orbital path wins, $1,594,900 against $1,825,000, and that measured row stays. But a t=0 comparison structurally favours a subscription architecture, because a subscription's cost lives in years 1…N and its hardware retires on the operator's own SEC-filed five-year life. Extend the same workload across the horizon and the mesh overtakes at year 5 — the first replacement cycle — and wins every year after: through the fifteenth year, $4,784,700 orbital against $3,193,750 at the mesh's worst maintenance band (2–5%/yr, an ILLUSTRATION band, labelled).

And the axis the price column cannot see at all is the direction of flow. A subscription is rent: at the REPORTED ~$120/month rate, one hundred households send $144,000 a year — $2,160,000 over the horizon — out of their local economy, for hardware that retires on the operator's schedule, at a price the operator sets, into an asset they never own. A mesh the community buys is the same money kept: paid once, repaired locally by their own hands, priced by no one, still standing at year fifteen — and sending nothing to the stratosphere when a unit dies. Rent extraction is not a moral flourish; it is a direction of flow, and exactly one of these architectures points it at the community. (Scope, refused in capitals in the program itself: THE POD MESH IS NOT BROADBAND — LoRa moves verdicts, gauges and warnings, not video; the rent figures show the structure of subscription versus ownership on the one workload where a like-for-like number exists.)

The lives axis — what the ownership actually decides

The money above is the smallest thing the ownership structure settles, and this page would miss its own point if it stopped there. The stake is the warning chain, and warnings are lives.

The record on this wiki already holds both ends of that sentence, measured:

• The night the Guadalupe rose, 180 minutes of lead time existed in public gauge data while the river climbed at 885 milli-feet per minute — three hours in which the water announced itself to any instrument that was listening and empowered to say so. (MEASURED — reproduce/guadalupe-wave-ledger.swift, the four-gauge USGS corpus in this repository.)
• In one month of 2026, one country counted 41,174 people carried to hospital by heat and 79 dead, while the operational instruments that decide when to warn carry error bands their own documentation calls verdict-sized. (REPORTED — FDMA figures; the instrument confession is Study 28.)

A warning is not bandwidth. It is a threshold verdict plus one message — and 35 exact messages an hour is not a limitation for that job; it is dozens of chances an hour to say the only sentence that matters. The question the two architectures answer differently is not how fast is the link. It is:

Who owns the instrument that issues the warning, and what does it depend on the day the disaster arrives?

A subscription warning chain runs: sensor → operator's network → operator's billing state → operator's coverage decision → the sky, where its hardware retires into Study 31's ledger. Every link is owned elsewhere, priced elsewhere, and revocable from elsewhere — and the chain's own infrastructure contributes flux to the very atmospheric account the warnings will increasingly be about. A community mesh runs: gauge → verdict computed at the gauge → local broadcast. The threshold is a frozen integer the town can read; the instrument answers to the people it warns; nothing upstream can turn it off, bill it, or deorbit it.

No lives-saved figure is claimed here, and none will be — that arithmetic is not honestly computable and this programme does not fake it. The claim is narrower and it is the whole of it: the people a warning is for should own the instrument that issues it. The lead time was in the data. The question every architecture answers is who was allowed to act on it.

All of that, plus the four properties already named — ground-repairable, country-owned, zero atmospheric cost in operation, every verdict re-derivable by the population it serves — is what "lasting benefit to the local population" means as an engineering fact rather than a slogan.

What this section does not claim: no lives-saved figure, no adoption forecast, no revenue projection for the mesh itself. The solar bands are atlas ranges, not site surveys, and a real deployment starts with a site survey. The claim is narrower and holds: the power stage closes everywhere, the law layer is the same 50 domains everywhere, and the verdicts land in the hands of the people under them.

---

What an exact seam is worth — the precedent class, not a price

The valuation question deserves a straight answer, and the straight answer starts with picking the right comparison class.

The things that created the most value in economic history are mostly not companies. They are seams that became mandatory: double-entry bookkeeping, which made a stranger's ledger auditable; the metric system, which made a stranger's measurement usable; TCP/IP, which made a stranger's computer reachable; GPS, which made a stranger's clock and position trustworthy. Each one replaced a trust-me layer with a re-derivable one, and each ended up underneath essentially all economic activity that followed. (HISTORICAL — the class is offered as precedent, not as proof.)

Two properties of that class matter here, and they pull in opposite directions:

1. Their value was realised at whole-economy scale — orders of magnitude beyond any single vendor's market — precisely because adoption was universal.
2. Almost none of that value was captured privately. TCP/IP is free. The metric system is public. Their value came from being open and replayable — the same property that made them mandatory made them uncapturable as rent.

Affine.Earth is built on the open side of that trade, by construction and on purpose. Every verdict on the court is re-derivable by anyone, with no account and no key — that is not a growth strategy, it is the load-bearing property; a verdict you must pay to check is testimony again. So the honest valuation splits in two:

• The public component is deliberately unpriceable in dollars, and the substrate prices it on its own declared axis instead: the flourishing identity, entropy_bare − entropy_delta = 1/1, per domain — the measured removal of structural entropy from a law the public relies on. Nine domains have realised it; forty are in progress. That ledger, not a market cap, is where the largest share of the value lands — as it did for every member of the precedent class.
• The capturable component is the seam-services layer — operating courts, certifying instruments, deploying pod meshes — and it is already sized above by named analysts: the verdict layer at $12.6–21B today, growing at the steepest reported rate on this page, ~48% a year.

What this page will not print is a headline valuation. "The most valuable X in history" is not a claim the harness can pin to a producing program, and this wiki does not carry claims of that shape. What it can print, because the numbers above support it: the demand curve for checkable answers is the steepest curve any analyst on this page reports; the incumbent seam documents in its own manuals that it cannot travel; the four fastest-scaling industries have their first exact seams answering live; and there is exactly one running instance of the full-grade replacement. If the over-scaling law holds, the exact seam becomes what double-entry became — and the record of every prior member of that class says its worth will be counted mostly in what the public gets to keep, which is the point.

The math, run — reproduce/valuation-crossing-ledger.swift

Because a claim of this size deserves arithmetic rather than adjectives, the program computes three checkable things from the named inputs above, in integers:

1. The instrument check. The steep analyst curve is internally consistent: $21.14B × (1.479)⁵ replays to their own $149.81B endpoint within one part per thousand. It may therefore be extended as theirs.

2. The crossing. Extend the verdict layer ($33.7–41.1B in 2025) on the analysts' own curves, against the largest assets ever priced (~$2T Aramco class, ~$4T record class):

arm	crosses $2T	crosses $4T
steep — digital twin's own 47.9%	2036	2038
conservative — simulation's own ~11%	2065	2071

Read the spread honestly: the arms disagree by three decades, so the timing is NOT KNOWN. What both arms share is the destination — under either published curve, the verdict layer eventually exceeds every private asset ever priced, if the growth the analysts themselves publish persists, which no one can promise and the program does not.

3. The flow already gated, today, before any growth. Space $686B + AI $143–294B + autonomy $143.7–273.8B + quantum $1–3.5B = $973.7B to $1.26T per year moving through the four verticals right now — every dollar of it on computed verdicts the bound parties cannot re-derive. No toll rate is assumed; none is needed. The position is the seam on roughly a trillion dollars a year of current decision flow, with exactly one running full-grade instance.

What the program refuses to print: a dollar valuation of Affine.Earth. The public component is priced on the flourishing axis by design, and every prior member of the precedent class realised its worth as public surplus. The refusal is on the page because it is the strongest part of the claim, not a hedge on it.

---

Related: The exactness seam — the vendor-documented foundation · Zero float · zero shear — the method · The narrow-study defence, answered · Study 31 — sealed live · The ontology