Humanity's Last Experiment · The Saturation Clock
Can the examiner still keep pace with the examinee?
Model expert review as a communication channel. Generation grows exponentially; review grows linearly. The coverage ratio ε(t) — the fraction of new work that can still be checked — falls toward zero. This instrument estimates, per domain, the year that ratio crosses below usefulness. Move the assumptions yourself; the verdicts recompute. Every estimate is OPEN.
A companion to Humanity's Last Exam, which asks whether we can still write questions hard enough to test the models. This asks the inverse, and the harder one: whether we can still grade the answers.
v0.1 · July 8, 2026 · Status: OPEN — a forecast instrument, not a measurement of record · every parameter is adjustable and every default is defended below · errata will be logged in place
The leaderboard
Six channels, ranked by how much time is left
The result is scale-invariant: for any generation growth above review growth, ε(t) → 0. So the question is never whether a channel saturates, only when. The clock below computes the crossover year from your assumptions. The one figure that matters is not the date — it is that every value in the defensible range still crosses.
Ordering is illustrative. The bit-ceilings below are computed from published sensitivity and specificity. The per-domain baseline offsets that order this leaderboard are editorial, not derived — they are the softest numbers on this page and are tagged as such. The crossover direction is robust; the ranking is a working estimate.
Default gai = 0.67 is measured (Liang et al. 2024: AI-modified papers 17.5%→~25% in one year). It is a single one-year extrapolation and is shown here as an adjustable band, not a fixed fact — drag it and watch that the verdict class holds across the whole range. Sublinear novelty (Heaps' law, ∝ producers0.5) is folded into the effective rate, not ignored.
| Channel | Channel capacity (ceiling) | Coverage ε now | Crossover | Verdict |
|---|
What makes it an experiment
A test names what would prove it wrong — then shows that closed
A prophecy asserts. An experiment states its falsifier up front and reports whether the falsifier fired. Here is ours.
The falsifier
Any one of these observations, sustained, would show the examiner keeping pace — that ε is recovering rather than collapsing. The experiment is honestly lost only if none of them appears:
- Coverage ε(t) rising for two or more consecutive years in any domain.
- Review capacity growing super-linearly — a structural break in the expert-pipeline rate, not a one-off funding bump.
- A collapse in publishable-quality fraction q — generation getting worse, not better — large enough to reverse effective growth.
- An automated verifier achieving human-equivalent first-order coverage at a cost that scales sub-linearly with the flood.
What the data shows instead
Every observed signal points the wrong way for the falsifier:
- arXiv submission growth accelerated from ~2%/yr (2021–22) to 12–17%/yr (2023–26) after ChatGPT — a structural inflection, not noise.
- Major-conference load rose faster still: ICLR +69.7% in a single year; ICML roughly doubled 2025→2026.
- Median review times lengthened 45→83 days — the channel already stretching under load.
- Better models raise q, which increases the review burden. The one force that would help — worse output — is not occurring.
The instrument, applied to itself
This page is inside its own channel
The materials behind this clock were generated by the same pipeline the thesis describes, and reviewed by the same multi-family process. It is tempting to offer the rising review scores as evidence for the thesis. We do not. A self-referential verification chain converges to a fixed point when — and only when — its update step is contractive; a score that climbs is equally consistent with convergence to a fixed point that is confidently wrong. That is our own Ouroboros result (HOE Chains, Theorem 5), and applied honestly it means the self-score is disclosure, not proof. The clock earns belief the same way any claim here does: an external check it has not yet passed. The loop is closed only from outside.
Why "hire more reviewers" fails
The ceiling is in bits, not in headcount
Each channel has a hard information ceiling — the maximum any review of that kind can extract about ground truth, no matter how many people run it. This is why the crisis is information-theoretic and not a staffing gap: even doubling the global reviewer pool buys roughly a year, because the gap is exponential and the per-review ceiling is fixed.
What "the channel is saturated" costs, per domain
These ceilings are computed from published sensitivity, specificity, and information content — not asserted. A 95%/95% diagnostic that sounds superb carries only ~0.71 bits: it cannot, even ideally, separate more than about 1.6 equally likely states. When generation outruns a channel whose ceiling is already that low, first-order verification does not merely slow — it stops being possible.