The Method

Constraints do not eliminate candidates. They reprice them, and only physical law bars outright. The framework is stratified: a deductive floor at Tier 0, an inductive middle across Tiers 1 to 5, and an abductive top across whatever survives. That is why every verdict carries a probability instead of a proof.

What the maxim licenses

When you have eliminated the impossible, whatever remains, however improbable, must be the truth.

The maxim is deductive, and it is valid only where the candidate set is exhaustive and each elimination is a genuine impossibility. Neither condition holds for most questions about the AI economy. Candidate sets over markets are not provably exhaustive, and almost nothing in economics is impossible.

A framework claiming uniform deductive force will fail the way every constraint-based forecast has failed, by treating an improbability as an impossibility. Malthus did it. The Ehrlich side of the Simon-Ehrlich wager did it. So this framework is stratified by inferential type.

The probabilities were always the tell. A framework that declares candidates eliminated and then attaches survivor probabilities of 0.55 to 0.78 is describing itself wrongly: a survivor at 0.60 leaves forty percent of the mass with candidates it called eliminated. That is incoherent under a deductive reading and coherent under a stratified one.

Deriving the six

The constraint set is open to a circularity charge if it is chosen with the conclusion in view. So it is derived from a question that never mentions AI: what does it take to obtain one more unit of a limiting input?

Three placements follow from the derivation rather than from preference. Data is a fossil resource, an accumulated stock with no meaningful replenishment, which places it above Matter despite being informational. Absorption is Tier 1 because it is biological: the capacity of firms to reorganize work turns over with people and management generations. And energy is not a primitive; it is an output of Matter and Law, floored by Thermodynamics.

That last one matters. Treating energy as one constraint conflated a physical invariant with two orders of magnitude of headroom, an industrial stock with none, and a permitting regime that moves in months. On this horizon delivered electricity binds through transformer lead times and interconnection queues, never through physics.

The engine

Every test reduces to one comparison: time required against time available. For candidate a and constraint c, the slack ratio is the quantity permitted over the quantity required, σ = Permitted / Required, and the shortfall is g = 1/σ. The time the constraint needs to deliver that expansion at its observed rate r is

T_req  =  ln(g) / ln(1 + r)

and survival is a logistic in the difference between time available and time required, with S = 0 whenever the constraint is Tier 0. That zero is the deductive floor, and it is the only place a zero appears.¹

The parameters are observable rather than chosen, which is the substantive improvement over a penalty-coefficient approach. Matter is decomposed by sub-constraint because its expansion rates differ by an order of magnitude: advanced packaging has expanded at 60 to 110 percent annually while grid equipment has expanded at 10 to 20. That ratio alone explains why the binding constraint migrated from packaging in 2024 to grid equipment in 2025 and 2026, with no additional assumption required.

Where this does not apply

The engine needs a candidate whose requirement expresses as an expansion factor and a constraint with an observable rate. Both exist for forward capacity questions. Neither exists for structural questions (does control concentrate, who bears liability) nor current-state ones (are cash flows sufficient today). Eight of thirty verdicts are computed. The rest are disciplined judgment, and each one says which it is.

Where time enters

Time is not a constraint. It has no independent permitted quantity, and one cannot compute time permitted over time required without first naming the process being timed. But three genuine time-dimension mechanisms exist, and none is captured by a single slack ratio.

Allocation is not stock. Every constraint has two clocks that differ by one to two orders of magnitude. A firm reserving more than half of a foundry's packaging expansion shortens no lag; it establishes queue position. An allocation mechanism never relaxes a constraint at the system level, only at the actor level, so system questions and firm questions are tested against different denominators.

Synchronization. Electricity is not storable, so a site needs power at the coincident peak rather than on average. The relaxation direction is the interesting one: a candidate that accepts interruption converts a synchronization requirement into an energy requirement, and curtailable interconnection clears in months where firm service clears in years. Flexible load is the single largest available relaxation lever on the buildout.

Serial path. A candidate can fail with every individual slack ratio above one, because the chain of conversions it requires is longer than the horizon. A greenfield site with no interconnection agreement needs six to seven years of serialized conversions against 4.4 remaining to 2030. That is a timing verdict, not a possibility verdict, and it must say so. Conflating the two is how constraint forecasting earns its reputation for crying impossible when it meant late.

The parameters, in full

These are the numbers every verdict runs on. Change one and the verdicts change, which is the sense in which this forecast is auditable rather than merely sourced.

Find the binding constraint

The operational instruction is not "check all six." It is three steps, and it is portable to any claim you encounter.

Screen. Order-of-magnitude every constraint. Most resolve in one line. Compute. Take the two or three lowest and do them properly, with sourced numerators and denominators, as intervals rather than points. Attribute. Name the binding constraint, state its tier, state the gap, and state the discontinuity that would move it.

A verdict that does not name a single binding constraint has not been resolved. The reason this matters beyond bookkeeping is a result from linear programming: a non-binding constraint has a shadow price of zero. An argument about a non-binding constraint cannot change the answer. Most public forecasting of the AI economy argues about capital, which on this horizon almost never binds.

The Elasticity GapThe parameter the whole field is betting on

Every forecast of AI's contribution to output is a bet on the elasticity between deployed compute and attributable output. Nobody names it, nobody has measured it, and its observed values ran 1.70, then 1.17, then 1.37 across 2023 to 2026. At the low end of that range this paper's own GDP band is eliminated and the sole survivor is the field's most conservative estimate. At the high end even the most bullish survives.

The parameter nobody names

Every forecast of AI's contribution to output implicitly assumes a relationship between deployed compute and attributable output. Write it down:

O  =  k · C^α

where O is AI-attributable output, C is deployed compute, and α is the elasticity of the first with respect to the second. At α = 1 output scales proportionally with compute. Below one, each additional gigawatt yields less than the last. Above one, deployed compute yields increasing returns, as it would if a model trained once serves many users or if value accrues through channels consuming little marginal inference.

No published forecast of the AI economy states its α. None reports it, none defends it, and it has never been measured. Yet every such forecast is a bet on its value, because the forecast asserts an output figure, the constraints permit a quantity of compute, and only α connects them.

Every forecast, run through the engine

Anchoring on 2026 at roughly $110B of attributable output on 29.6 GW of AI data-center capacity, taking permitted incremental capacity to 2030 as roughly 200 GW, and applying the standard efficiency correction of about 40 percent per year, the slack ratio for each published forecast is a function of α alone.¹

The observed values of α, computed from the attributable-output and capacity series, are unstable and sit in the increasing-returns region:

What the table says

At α = 1.0, this paper's own GDP band is eliminated. Both bounds fail, at σ = 0.27 and σ = 0.41, alongside Goldman and PwC. The single survivor is Acemoglu, the most conservative published estimate in the field.

At α = 1.4, the midpoint of the observed range, this paper survives and so does Goldman. At α = 1.7, the value observed across 2023 to 2024, even PwC survives.

So the central quantitative dispute in this field is a disagreement about α that no participant has named. PwC's $15.7T is a bet that α ≥ 1.7. Goldman's 7 percent is a bet on roughly 1.4. Acemoglu's 0.66 percent is a bet that α ≤ 1.0. The verdicts in this paper are a bet on roughly 1.3 to 1.4. None of them has stated the wager it is making.

The observed range spans values that reverse every forecast in the table. That is not a narrow uncertainty around a central estimate. It is the difference between a technology that adds one percent of output and one that adds ten.

What it costs this paper

The exposure is concentrated rather than general. Verdicts about output magnitude are bets on α. Verdicts about physical delivery, lead times, market structure, liability and reliability are not, because they never require the compute-to-output conversion. Each is marked where it applies.

What would settle it

A panel of AI-attributable output against deployed compute at firm or sector level, with output measured as realized revenue plus estimated surplus rather than as capital expenditure. That series does not exist. Building it is the single most valuable unbuilt dataset in the economics of AI, and it would settle a dispute currently conducted entirely through unstated assumptions.

The Roots and the Economy

The majority-of-GDP claim is dead: it needs $77T of new output by 2030 and the most bullish credible estimate permits $16T. What survives is 1.0 to 1.5 points of annual growth. Realized gains today are 0.1 to 0.2 points against a 1.5 point electricity precedent. Absorption binds, and it relaxes on a twenty to forty year clock.

The five roots

Beneath the hundred questions lie five root uncertainties. Every branch question inherits its verdict from one or more of them.

How much GDP

The majority-of-GDP claim requires AI-attributable output above half of a $154T 2030 economy. That is $77T of new output in four years. The most bullish credible estimate in the field, PwC's $15.7T cumulative, permits $16T. A ninefold gap, and no constraint in the set closes it in four years.

What survives is 1.0 to 1.5 points per year, $7 to $10.5T cumulative by 2030. That is inside the electricity and information-technology precedents, not above them. The bulls are not wrong about the technology. They are wrong about the denominator.

Two qualifications, both of which narrow the claim. It is α-dependent and provisional. And it forecasts measured GDP: value delivered as consumer surplus on goods priced at zero sits outside that quantity by construction, and US consumer surplus alone is running at roughly $172B annually and growing 54 percent. A reader concluding that AI creates little value from a modest GDP verdict has misread it. The verdict is about capture and measurement, not welfare.

Why it arrives slowly

An instant productivity surge is not a live option. It requires 69.9 years of absorption expansion against 4.4 available. Internet-speed diffusion is not one either: the binding constraint is deployment rather than distribution, and the precedent is two to four decades. Electric motors were under five percent of factory mechanical drive in 1900. The productivity gains landed in the 1920s.

The distinction that matters, and that headline adoption figures obscure: consumer uptake and organizational absorption run on different clocks. Generative AI reached 53 percent population adoption in three years, faster than the PC or the internet. Over the same period AI agent deployment stayed in single digits across nearly all business functions. Both facts are from the same report.¹

The strongest contrary datum

The rest of the domain

The Binding ConstraintsCompute, data, models, labor, agents, software, the physical world

Fifteen questions, one binding constraint each. Models commoditize and the moat moves to everything around them. The data wall is weaker than the field believes. Grid equipment is stronger. The chokepoint on the buildout is a transformer with a 128-week lead time, not a chip and not a dollar.

Compute and its financing

Data

This is the verdict the paper is least confident about, and the reason is in Piece V: the framework predicted the data wall would bind by 2026, at maximum confidence, and was wrong. Tier 2 is where substitution defeats stock-based reasoning.

Models and scale

Q15. Do models commoditize?

One model winning is eliminated: no frontier lead has held for more than months. Zero margin everywhere is eliminated: the newest model always commands a premium. What survives is that the floor commoditizes fast and the frontier holds a brief, eroding lead. The model is a commodity and the moat is everything around it. BOND reaches the same premise and calls the consequence a riddle. It is not a riddle. The floor falls 11.8x to 75.3x per year while the frontier falls 1.34x to 2.25x, and that divergence is the whole answer.

binding: Capital, Tier 5 · P ≈ 0.75

Labor and value

Agents

Q20. Can agents run whole workflows unsupervised?

No, and the limit is arithmetic rather than sentiment. Success compounds multiplicatively. At 95 percent per step, a twenty-step workflow succeeds 36 percent of the time. Long unsupervised chains need 99 percent-plus per step. Production is at 56.

empirical, not constraint · P ≈ 0.72

Software and the physical world

The Call and the Record

One dated bet with five adjudication rules: the 80 percent reliability horizon stays under eight hours on 31 December 2027, at P = 0.85. Plus the backtest, including a prediction this framework got wrong at maximum confidence, on the exact constraint tier it had already flagged as vulnerable.

The Call

The widely cited METR figure is the 50 percent horizon, and it has risen several-fold since late 2025. The 80 percent horizon, the level at which work can actually be delegated, has been approximately flat at 27 to 32 minutes across two frontier release cycles.¹

Falsification therefore requires roughly a sixteen-fold rise, four doublings, in under two years, in a series that has not moved. The mechanism is the argument rather than the extrapolation: capability gains are arriving as intercept rather than slope. Models are becoming better at succeeding sometimes on long tasks without becoming better at succeeding reliably.

Adjudication rules

A bet that cannot be settled is not a bet.

1. Series. METR Time Horizon 1.1 or its successor, as published at metr.org/time-horizons.
2. Methodology change. If METR replaces the series, the Call adjudicates on the successor. If the successor reports no 80 percent horizon, the Call resolves void rather than in either party's favour.
3. Generally available. Any member of the public can obtain access by payment, without individual negotiation, waitlist approval or preview enrollment.
4. Unmeasured models. Adjudicated on the highest published 80 percent horizon among generally available models. If the strongest has no published horizon, the Call resolves on those that do and the omission is recorded.
5. Adjudication date. 31 March 2028, to allow for publication lag.

Two hazards acknowledged: METR has noted that measurements above 16 hours are unreliable with the current task suite, and its coverage is explicitly not comprehensive. Both make the Call harder to settle rather than easier to win.

The backtest

A framework that has never been scored is falsifiable in principle and not in practice. Five questions were pre-registered with parameters fixed at their historical as-of dates, before outcomes were consulted. One was chosen because the framework predicts its own failure there.

The miss

Three corrections followed. The Tier 2 expansion rate was revised by an order of magnitude, from 0.02 to 0.05 up to 0.30 to 0.60, because the parameter must be the rate at which the effective stock expands including substitution. Three verdicts moved. And the Brier score is now reported by tier rather than in aggregate: the framework is excellent on Tiers 3 and 5, where constraints are industrial and rates are observable, and it was worthless on Tier 2.

What the backtest does not establish: five questions is not a validation set, and four of five are physical-infrastructure questions, the domain where the method should perform best. It is untested on institutional, labor and behavioural questions, which is most of the economic domain, and is probably weaker there.

The objection that is granted

Six objections are answered in full in the paper. One is granted outright.

The register, in full

No participant in this field maintains a track record. Forecasts are published, cited, and quietly superseded. This is the scoreboard, and this paper's own entries are listed first and scored on the same basis as everyone else's.

Resolved: the pre-registered backtest

Open: this paper's dated forecasts

Open: the field's forecasts, scored on the same basis

What checking the citations changed

Twenty sources checked against source, sixteen corrections. Every substantive figure held up. What failed was the pointer. Four of these changed a conclusion rather than a reference.

Limits

The closed-world problem is the binding constraint on the method itself. In the worked verdict it cost a factor of two in the leading posterior before the completeness check caught it. Enumerate until two consecutive additions leave the leader unchanged, and report how many were required.

The constraint set and the conclusion overlap. The six are derived from replenishment mechanisms without reference to AI, which answers the charge structurally. What it does not do is produce a counterexample: in a hundred applications there is no question where the binding constraint is one of the six and value nonetheless accrues to the intelligence layer. Until such a case is found and published, the circularity charge is not fully answered, and this is recorded as open rather than closed.

The parameters are estimates from thin evidence. They are better than chosen coefficients because they are in principle measurable, but they are not yet well measured. Signpost S16 exists to detect when they are wrong.