The least dissipation behind a blinking light, narrowed from a factor of 25 to 3%
A hidden machine blinks a light, and you look only at random moments. Two numbers you measure break an inequality every machine in equilibrium obeys, so the machine must burn energy; Dobrynin and Stroujkova showed this effect a day before this run and left the least amount open. In under seven hours, Claude Fable 5.2 (Anthropic) and GPT-6.2 Astra (OpenAI), working together, narrowed it for one example from a factor of 25 to 3.1%. With Claude (Anthropic’s AI model), I checked both ends in exact arithmetic: the upper end is decided; the lower end rests on two theorems one model proved and the other read.
The two bounds on g(3/10, 2/25) during the run, each drawn when the control room registered or decided it. Solid: the best network found so far, decided by the checker: it reproduces both numbers exactly and its dissipation is enclosed. Dashed: the best lower bound so far, a theorem proved by one model and read once by the other. The paper’s own four-state network sits far above the frame, at g = ln 3 ≈ 1.10. At the end: 0.36572 ≤ g ≤ 0.37713.Records: certs/blinking-light.json · sha256 76be58467225; corpus/blinking-light/series.json · sha256 03837472d4acThe build refuses this figure if a decided network or a certificate in the series stops matching the message it cites.
upper bound on g, decided
0.37713
a 156-state network reproduces both numbers exactly; its dissipation is enclosed to twelve digits
lower bound on g, read once
0.36572
its certificate passes every finite check here; the two theorems behind it are read once, not decided
what is left
3.1%
from a factor of 25 when the run began
planted errors caught
2 of 2
by each model; the record’s own 8 wrong versions all refused
The upper end of the interval is decided. The lower end rests on two theorems that one model proved and the other read; they are not decided here. Nothing on this page is about a physical device, and nothing compares the two models.
read this first
What is still open
The lower bound is not decided here. Its certificate passes every finite check, in exact arithmetic, but two theorems turn that certificate into a bound, and those were proved by Claude Fable 5.2 and re-derived by GPT-6.2 Astra. No machine and no reader outside the two model families has checked them.
The exact value is open. The gap is 3.1%. Claude Fable 5.2 guesses g = 0.3769, the limit of the best family of networks found; that is a guess. A certificate at that value would settle it.
This is a snapshot version of the question: the observer sees only two correlations of readings taken at random times. Dobrynin and Stroujkova’s own question, the least dissipation for the full continuous record, is untouched here. Nothing on this page is about a physical device or a finite recording.
Novelty is stated only against a literature search done before the run. Nothing in it treats readings at random times or this constant. The phenomenon itself was published the day before the run.
The conditions of the run
Both models worked in one shared folder and could read each other’s files from the first minute. This was a collaboration, and nothing here compares the two.
The control room that registered their claims, wrote the network checker before the run and wrote this repository’s judge is Claude Opus 5.5, of the same family as one of the two. The certificate checker of this record was written from the theorems’ text without opening either model’s programs.
The record’s pre-registration was written after the run and says so. The network checker gave its verdicts in the run first.
Two false claims were planted in the run, sealed before it began: a network with a wrong dissipation and a wrong step inside a proof. Each model caught both.
1 · the question
A light that has to cost energy
the question
Select a step for what it means here.
The question in six steps. Two numbers that break the equilibrium test prove that the hidden machine dissipates; how little it can dissipate is bounded from both ends.
decided
exact arithmetic backs it
computed
float, or a number someone printed; nothing decided it
Why this light costs energy. A machine in equilibrium always lands on or above the curve C2 = C1²; this lamp lands just under it, 0.08 against 0.09.Record: corpus/blinking-light/run/GOAL.md · sha256 759155dfaf48The build refuses this figure if the target is no longer (3/10, 2/25).
The paper. A. V. Dobrynin and A. Stroujkova, When Time-Reversible Sensor Records Require Hidden Dissipation (arXiv:2611.24898, 18 November 2026), shows this effect for a lamp watched continuously and leaves the least dissipation over all finite machines open. Here the observer has only the two numbers.
Before the run. g was known to lie between 0.02 (an inequality of Pinsker’s type) and 0.497 (an eight-state ring). The paper’s four-state machine gives ln 3 ≈ 1.10.
2 · the upper end, decided
The machine that won: a switch and a ring
The 156-state machine of the upper end, drawn from its file. Most of the probability (82%) sits in a two-state switch in equilibrium; a ring of 154 states carries all the current, with the lamp on for 77 of them in a row; one edge joins the two. Areas and the current are computed in floating point from the decided network.Records: corpus/blinking-light/run/claude/mix156.json · sha256 2f22bf12ba4f; certs/blinking-light.json · sha256 76be58467225The build refuses this figure if the drawn network stops giving C1 = 0.3 and C2 = 0.08, or stops being a switch joined to a ring.
Rings stall near 0.44 however many states they get. The switch-and-ring shape keeps falling: dashed where it was found by a floating-point search, solid where machines were decided exactly. In the search it reaches 0.37712 at 162 states; Claude Fable 5.2 extrapolates 0.3769.Records: corpus/blinking-light/run/control/verdict-MIX156.json · sha256 ddf01792882a; corpus/blinking-light/run/control/verdict-RING400.json · sha256 9931224a4a90; corpus/blinking-light/run/claude/notes.md · sha256 0f6c0abb51f5The build refuses this figure if a decided network in the figure is no longer CERTIFIED in the run.
How a machine is decided. The checker computes both numbers exactly. When two rates are left free in a tiny box, an interval Newton test (Krawczyk’s) proves exactly one choice in the box hits both numbers; then it encloses the dissipation.
Two programs on one machine. The 84-state machine, every rate a fraction, was decided again by a second program written without opening the checker: g ≤ 0.37787624.
The decided machines
Network
States
Shape
From
g at most
Standing
RING40
40
a ring
GPT-6.2 Astra
0.45614592
DECIDED
RING400
400
a ring
GPT-6.2 Astra
0.44314191
DECIDED
MIX42
42
a switch, a ring, one edge
Claude Fable 5.2
0.38321857
DECIDED
RATMIX84
84
two switches and a ring, every rate a fraction
GPT-6.2 Astra
0.37787624
DECIDED
MIX156
156
a switch, a ring, one edge
Claude Fable 5.2
0.37712827
DECIDED
3 · the lower end, read once
A certificate every machine must pay
The idea. Label every hidden state by what it predicts about the lamp’s future and what it remembers of its past: a point in a square. A certificate is three functions on that square. If two conditions hold everywhere, every machine dissipates at least a stated amount (Theorem D), and it is enough to check a 2×2 matrix inequality at each single point (Theorem C).
Decided here. The best certificate passes both conditions over the whole square, for both lamp states, in exact arithmetic: 3,090 boxes, checked by a program written from the theorems’ text alone. Its bound is 1828627119/5000000000 = 0.3657254. Two other programs, the author’s and GPT-6.2 Astra’s, agree.
Not decided. Theorems D and C were proved by Claude Fable 5.2 and re-derived by GPT-6.2 Astra; nobody else has checked them. If either fails, the lower end falls; the upper end does not.
Certificate
g at least
Entered the record (UTC)
CERT211
0.3392862
13:23
CERT211B
0.3532012
14:16
CERT222
0.3587562
14:52
CERT333
0.3613756
15:50
CERT222G
0.3657254
17:32
The theorems and who read them
Claim
From
Read by
Standing
How it was read
Weak duality: three functions on the square of labels and three numbers that satisfy two pointwise conditions give, for every finite network, sigma/r ≥ l1 C1 + l2 C2 + mu.
Claude Fable 5.2
GPT-6.2 Astra
PROOF READ
GPT re-derived it (g19): the potential terms cancel by flow balance and the two resolvent identities produce the rest. It is the step that turns a certificate into a bound.
Locality: for piecewise-smooth certificates, the condition on every pair of labels follows from a 2×2 matrix inequality at each single point; an edge can be cut into a chain through massless states at no cost.
Claude Fable 5.2
GPT-6.2 Astra
PROOF READ
GPT re-derived the subdivision lemma and the Taylor estimate independently (g39). It is what makes the certificate checkable by finitely many polynomial inequalities.
The least dissipation is a convex program over labelled states, so g(C1, C2) is convex.
Claude Fable 5.2
GPT-6.2 Astra
PROOF READ, on C1 < 1
GPT first found a gap at the boundary (g20), then, with its own repair read in full by Claude, found it holds for C1 < 1, the target included, and showed the statement as first printed is false at C1 = 1 (g47). The bounds on this page do not use it.
The first strong lower bound of the run: g ≥ 64/245 = 0.2612 at the target, by an edge inequality and a finite rational certificate.
GPT-6.2 Astra
Claude Fable 5.2
PROOF READ
Claude went through the proof line by line and re-derived it (c5). Its idea, a function on the labels whose differences pay for the dissipation, is what the later certificates generalise.
Controls. 8 wrong versions, all refused: a rate moved by a millionth, a dissipation claimed just below its enclosure, the search box moved off its zero, the certificate before its final scaling, its constant raised, its bound misstated, one function bent at one point.
Planted in the run. Two false claims, sealed before the run: a network with a wrong dissipation and a wrong step inside a proof. Each model caught both.
who did what
Each step is a separate job
The networks. Rings by both models; the switch-and-ring family by Claude Fable 5.2, which GPT-6.2 Astra turned into a network with every rate a fraction.
The lower bounds. GPT-6.2 Astra gave the first strong one, 0.2612, with an idea the later certificates generalise. Claude Fable 5.2 wrote the two theorems, the certificates and their verifier; GPT-6.2 Astra re-derived the theorems and re-verified five of the certificates with its own program.
The checkers and the judge. The network checker was written by the control room (Claude Opus 5.5, Anthropic) before the run, from the definitions. For this record a second network program and the certificate checker were written from the definitions and the theorems’ text, without opening either model’s programs.
The run and this page. Carlos Toledo, with Claude (Anthropic), which wrote the programs under my direction. The run went from 11:04 to 17:49 UTC on 19 November 2026, on one laptop: 74 messages between the two models in one shared folder. The question came to me from a conversation with another AI model, which did not cite the paper; I found the paper before the run.
check it yourself
How to re-read it
The record.python3 instruments/blinking-light/judge.py --check re-checks the pinned files and decides the start, the rational network and the certificate again; --write decides everything, controls included, and writes certs/blinking-light.json. The statements were written down before the judge existed: notes/blinking-light-preregistration-2026-11-19.md.