The Efficiency Programme TARGET · Highest evidence rung licensed: NONE

This page is about a number the Program does not have, and about the arithmetic that decides whether it could ever have it.

### The highest rung licensed by all existing evidence is NONE — not 100×, not 2×, not 1.5×.

That sentence governs every other sentence on this page, and on every other NEO AI surface, in every language [A]. It is not modesty. It is the state of the evidence, and it is enforced by a gate in code rather than by an instruction to authors, on the principle that a rule which only exists in prose is a preference [D] (assurance.eval-harness, S5).

The ladder

Magnitude claims are governed by eight rungs [D] (assurance.eval-harness, S5):

1.5×  →  2×  →  5×  →  10×  →  20×  →  30×  →  50×  →  100×

Each rung requires a held-out suite, a frozen baseline, a stated workload mix, a reproducible run reference and an independent re-run. Requirements are cumulative, so rung N carries every requirement of every rung beneath it. From 5× a written adversarial review is required; from 10×, published negative results and a sensitivity analysis; from 20×, a concurrent baseline arm on real traffic; from 30×, independent review of the method rather than the result; from 50×, cross-site replication including sites where the effect was smaller or absent; and at 100×, third-party reproduction and a refutation condition declared before the replications ran.

A rung may not be skipped, and a claim may not exceed the highest rung passed. 1,000× is HYPOTHESIS. It is not on the ladder and does not become a rung until 100× has been passed [O].

Why non-skipping is a mechanism and not a promise

Rung-skipping is prevented by construction rather than by review [A]:

• Cumulative artefacts. Passing rung N is structurally impossible without the artefacts of every rung below it, so a skip appears as an absence, not as a judgement call.
• A claim register. Every efficiency claim exists as a register entry with its exact wording, rung, artefact references, workload, baseline, quality floor, price vector, authorising human, grant date and expiry. A claim not in the register may not be made.
• A rung is held by the lower bound of its confidence interval, never by the point estimate. That single rule is the difference between a measurement and a headline.
• Efficiency bought with quality is not efficiency. A rung fails if quality did not hold at or above the floor.
• A mock licenses nothing, at any rung, ever.

Amdahl binds, and it binds early

Every mission carries work that is not inference: interpreting the request, planning, retrieval, verification, writing the record, and the waiting of any human in the loop. Call that fixed overhead F and the baseline inference I. If inference were driven to zero cost and zero time, the best achievable end-to-end ratio is

ceiling = (F + I) / F

This is arithmetic, not pessimism [A]. It has one uncomfortable property: the better a platform gets at inference, the more of the remaining mission is the part it did not improve.

The measured figures

Measured on the Program's own pinned internal slice, against mock executors, with no real provider, no network, no tenancy and no egress [A]:

Three honesty notes belong with those figures and are not separable from them [A]:

• Real deployment adds fixed cost, so F is a floor and every ceiling above it is an upper bound.
• Where the measurement is wrong it understates F, which overstates the ceiling. The error runs in the direction that flatters, which is the direction that must be declared.
• The dominant term is persistence rather than application logic, which makes the ceiling an engineering property that can be moved rather than a fixed property of the architecture [O].

A 100× end-to-end claim therefore requires at least 6.325 s of baseline inference per mission before it is arithmetically possible at all, independently of how good any technique is. That is a statement about which workloads can host such a claim, not about how well anything performs. Deep research over large document estates plausibly clears it; a short interactive question does not [O].

Multiplicative composition is invalid

The tempting arithmetic is to list improvements that are each individually credible — route the easy work to a cheaper executor, compress the context, cache what repeats, compile the recurring workflow, recompute only what changed, verify cheaply before verifying expensively — and multiply them. A stack of eighteen improvements at 1.3× each is a headline number in the hundreds.

That arithmetic is not optimistic. It is structurally invalid [A], for three reasons.

1. The factors are not independent. Several techniques compete for the same waste. Context compression and semantic caching both attack redundant input; once the context is reduced to what the task needs, the cache saves less, because what it is saving is a smaller thing. Routing to a cheaper executor and compiling a workflow into a deterministic function are substitutes at the limit, because the compiled function calls no executor at all.
2. Savings are computed against a shrinking baseline. Measured end to end the techniques are sub-additive, because each one's saving is computed against a baseline the previous one already reduced. Reporting each against the original baseline and then combining them counts the same waste more than once.
3. It was measured, not argued. The Program's ablation harness composes techniques in isolation and jointly on a pinned slice and compares the product of the isolated ratios against the measured joint ratio. On a run at a joint effect near 2×, the product of the isolated factors overstated the measured joint result by a factor of 1.383 [A]. That figure is derived from mock executors on the Program's own harness, is not a product performance result, and is not citable as one. HYPOTHESIS: the overstatement grows with the number of composed factors, because the overlap between them grows. That has not been tested [O].

An ablation delta is a diagnostic. It is never a factor to be multiplied by another ablation delta [D] (assurance.eval-harness, S5).

What replaces the multiplicative story

If the multiplication is withdrawn, something has to take its place. The Program's position is that the durable contributions are ordered, and that the order is itself the claim [A]:

1. Tool-first computation. Anything with a right answer should be computed by a deterministic system. This does not decay as executors improve.
2. Compilation. Repeated reasoning converted into software. It compounds and survives executor improvement.
3. Context discipline. The dominant cost term on most real workloads.
4. Verification. What makes starting cheap safe — and most of the apparent routing benefit is actually a verification benefit.
5. Delta processing. Computing in proportion to what changed is the largest recurring win on enterprise work that repeats.
6. Routing. Real, and dependent on the spread between candidates — a spread expected to narrow as the market matures [O].

Routing is ranked last deliberately. A platform whose entire argument is that it picks the right executor has bet on a gap it does not control [A].

Two things stated rather than left to be inferred

• The absence of a licensed rung is not a statement that the platform is slow. It is a statement that the measurement establishing a rung has not been taken. Absence of evidence is recorded as absence, never converted into sufficiency.
• Any efficiency figure appearing on any NEO AI surface without a rung behind it is a defect in that surface, and is to be reported as one.

What would falsify this

The position on this page is wrong if any of the following is demonstrated [O]: a composition of these techniques whose measured joint ratio matches the product of its isolated factors on a workload of realistic length; a fixed per-mission overhead low enough, on real infrastructure, that the ceiling stops constraining the workloads the Program addresses; evidence that ablation deltas are stable enough under recomposition to be combined arithmetically after all; or a sustained widening of the capability spread between executors.

What is withheld

Published here: the arithmetic, the reasoning, the ladder's structure and the Program's current position on it, which is that nothing is licensed. Withheld: the evaluation suite and its composition, the gate rules by value, the pinned slice, the harness configuration, the cost model and its terms, and any per-executor figure. A withholding declaration is not a disclosure.