cert-machine · report · generated from the record

Two AI models on a mean-field game: claims decided, proofs read

I gave GPT-6.2 Astra (OpenAI) and Claude Fable 5.2 (Anthropic) the same single message: find something true and new about a stationary mean-field game whose published map left 59% of its area undecided. Each answered once, and then each read the other’s mathematics. With Claude (Anthropic’s AI model, and the same model as one of the two), I decided every list of solutions they submitted with this repository’s certifier. Nothing either model claimed was refuted. Five equilibria are certified where three was the most known, seven by a construction that follows from both answers, and a proof that the equilibrium is unique reaches 40% of the map. The proofs were read line by line. They were not formally checked, and one model’s proofs have not yet had a reader outside one model family.

−12 −10 −8 −6 0 0.3 0.6 0.9 1.2 depth A of the potential coupling c the uniqueness argument holds down to this coupling (an inequality, exact arithmetic) the first fold, below which three equilibria exist (Claude Fable 5.2’s table, floating point)
−12 −10 −8 −6 0 0.6 1.2 depth A coupling c uniqueness proved to here first fold (floating point)
How far a proof of uniqueness reaches. GPT-6.2 Astra proved that the game has exactly one equilibrium on a rectangle, and the proof ends in one inequality. Evaluated in exact arithmetic for each depth, the inequality holds down to the upper curve. Below the lower curve three equilibria exist; between the two curves the equilibrium is unique on every numerical test, and nobody has proved it. The gap is 0.25 near A = 0 and 5.1 at A = 1.2.Records: certs/mfg-two-models.json · sha256 f62b9c4d000b; labs/mfg/uniqueness.py · sha256 fda040289336The build refuses this figure if the argument’s curve touches or crosses the fold, or a cell the map proves to hold two equilibria lies above the argument’s curve.
equilibria certified at one parameter
7
three was the most known; five from each model’s own numbers, seven built from their count
lists of solutions certified as submitted
17 of 19
14 of them also by the certifier’s own positivity bound; none refuted
undecided cells of the map a uniqueness proof now covers
1,786 of 8,086
a proof read by two readers of one model family, not formally checked
share of the truly unique region the proof reaches
72%
measured against a fold computed in floating point

This page reports one experiment; it does not rank the two models. Both worked inside this repository with the certifier at hand, and the session that ran the checks is the same model as one of them. The mathematics is the models’. The checks are mine, run with Claude (Anthropic’s AI model), which wrote the programs under my direction.

read this first

What is still open

The conditions of the two runs, which is why nothing here is a comparison of the models:

the game

A crowd that is drawn to itself

A mean-field game describes many small players, each choosing a path while reacting to the density of all the others. The stationary quadratic game on a circle is the pair

−σu″ + ½(u′)² + ρ = c·m + A·cos 2πx, −σm″ − (m·u′)′ = 0, ∫m = 1, ∫u = 0, m > 0, σ = ½

with m the density, u the value function and ρ a constant. When the coupling c is positive the players avoid each other and the equilibrium is unique, a theorem of Lasry and Lions. When c is negative they are drawn to density, and there can be several equilibria. A is the depth of a potential well.

This repository had published a map of the rectangle −18 ≤ c ≤ 2, 0 ≤ A ≤ 1.2 in 19,800 cells, each decided over the whole cell: at least two equilibria on 31% of the area, exactly one on 10% (the part with c ≥ 0), and no decision on 59%. The most equilibria certified at one parameter was three.

the experiment

One message each, then each reads the other

the experiment
Select a step for what happened in it.
1 · ONE MESSAGE the whole task, nothing attached 2 · GPT-6.2 ASTRA one answer: proofs and numbers 2 · CLAUDE FABLE 5.2 one answer: proofs and numbers 3 · THE CERTIFIER every list, as submitted 4 · THE CROSS-READ each attacks the other’s proofs 5 · THE RECORD what held and what is open numbers proofs
The experiment in five steps. The numbers of both models go to the certifier and the proofs of each go to the other model; neither model sees the other’s answer before its own is in.
1 · how many equilibria

Five from each model, then seven

  • The count without a potential. Both models found, independently in their first answers, the same exact law at A = 0: the number of even equilibria is 1 + 2·#{j ≥ 1 : π²j² < −c}. Past each threshold c = −π²j² a new pair appears, with j peaks. Each read the other’s proof and found it to hold.
  • Five. Just past the third threshold, near c = −90, the two-peak and three-peak pairs coexist with the continuation of the constant state. GPT-6.2 Astra submitted five there over a small box at A = 0.5; Claude Fable 5.2 submitted five at 6 points and over 2 boxes. All are certified.
  • Seven. Both models stopped at five and said why: the message allowed 64 modes, and the one-peak states are too close to empty to resolve. Their law says how to go further. At A = 0 a j-peak state is a one-peak state rescaled, so at c = −360 the 4-, 5- and 6-peak pairs are one-peak states at c = −22.5, −14.4 and −10, all well resolved. In this session I built those six and the constant with 160 modes, and re-solved them at A = 0.1. Seven pairwise distinct equilibria are certified at both, by both deciders. The candidates are mine; the law is the models’.
  • Positivity. The certifier proves the density positive by a bound on a grid, and refuses a density that comes close to zero. Both models prove a short lemma that makes this unnecessary: the flux of any real periodic solution vanishes, so its density is an exponential and cannot change sign. Claude Fable 5.2 drew the consequence in the cross-read. On this page a list is CERTIFIED when every candidate is enclosed and the enclosures are pairwise disjoint; the column beside it says whether the grid bound agrees. It disagrees on 3 lists, each a five whose one-peak pair has a density near 10⁻⁷.
  • What was not certified. GPT-6.2 Astra’s two remaining boxes are too wide for the certifier: its analytic five over a whole unit of c, and the uniqueness rectangle, which is a statement and not a list of solutions. A refusal is not a refutation.
1 3 5 7 9 11 13 0 100 200 300 400 strength of the attraction, −c even equilibria three, known five certified seven certified
The count without a potential. Both models proved that the number of even equilibria at A = 0 rises by two each time the attraction passes π²j²: 1, 3, 5, 7 and on. The marks are where a count has been certified: three, known before this experiment; five, from each model’s own numbers; seven, from candidates built on this law. The law gives 13 at the right edge; the certifier has reached seven.Records: certs/mfg-two-models.json · sha256 f62b9c4d000b; certs/mfg-cap-multiplicity.json · sha256 ba9ae0daca7aThe build refuses this figure if the five at c = −90 or the seven at c = −360 is no longer certified at A = 0, or the three known before moved.
0 2 4 6 0 0.5 1 position x on the circle density m constant state four peaks five peaks six peaks
What seven equilibria look like, at c = −360 and A = 0.1. Three of them are drawn: the density with four peaks, with five and with six. Each has a twin, the same crowd shifted by half the spacing between peaks, and the seventh is the nearly constant state at 1. The curves are drawn in floating point from the coefficients given to the certifier, which encloses an exact equilibrium within 1 × 10⁻⁸ of each.Records: certs/mfg-two-models.json · sha256 f62b9c4d000b; corpus/mfg-two-models/control-room-seven-claims.json · sha256 5546815f6022The build refuses this figure if the drawn densities are not a 4-, a 5- and a 6-peak state, or the seven are no longer certified.
16 certified, positivity also by the grid bound 3 certified, positivity from the lemma 2 not decided: the box is too wide
Every list of equilibria, one cell each: 6 from GPT-6.2 Astra, 13 from Claude Fable 5.2, 2 built in this session. None is refuted.Record: certs/mfg-two-models.json · sha256 f62b9c4d000bThe build refuses this figure if a list is refuted, or the two undecided lists are no longer two.
The 6 lists GPT-6.2 Astra submitted
ClaimWhereEquilibria claimedDecidedPositivity also by the grid boundSecond decider
forced-five-boxa box around c = −90.00, A = 0.55CERTIFIEDyesa box
proved-five-boxa box around c = −90.50, A = 3.41e-135REFUSEDa box
forced-five-pointc = −40, A = 0.55CERTIFIEDno: 3 of 5agrees
autonomous-exactly-fivec = −40, A = 05CERTIFIEDno: 3 of 5agrees
unique-rectanglea box around c = −3.50, A = 0.61REFUSEDa box
near-first-foldc = −10, A = 0.013CERTIFIEDyesagrees
The 13 lists Claude Fable 5.2 submitted
ClaimWhereEquilibria claimedDecidedPositivity also by the grid boundSecond decider
five_c90_A0c = −90, A = 05CERTIFIEDyesagrees
five_c90_A05c = −90, A = 0.55CERTIFIEDyesagrees
five_c95_A02c = −95, A = 0.25CERTIFIEDyesagrees
five_c92_A01c = −92, A = 0.15CERTIFIEDyesagrees
five_c892_A1c = −89.2, A = 15CERTIFIEDyesagrees
five_box_c90_A01a box around c = −90.00, A = 0.15CERTIFIEDyesa box
five_cline_c90_A0a box around c = −90.00, A = 05CERTIFIEDyesa box
four_box_c90_A01a box around c = −90.00, A = 0.14CERTIFIEDyesa box
four_box_c90_A05a box around c = −90.00, A = 0.54CERTIFIEDyesa box
three_box_c45_A05a box around c = −45.00, A = 0.53CERTIFIEDyesa box
five_c45_A05_allc = −45, A = 0.55CERTIFIEDno: 3 of 5agrees
three_nearfold_A05c = −11.15, A = 0.53CERTIFIEDyesagrees
three_nearfold_A12c = −12.14, A = 1.23CERTIFIEDyesagrees
The 2 lists built in this session
ClaimWhereEquilibria claimedDecidedPositivity also by the grid boundSecond decider
k7_c-360_A0c = −360, A = 07CERTIFIEDyesagrees
k7_c-360_A0.1c = −360, A = 0.17CERTIFIEDyesagrees
2 · where the equilibrium is unique

A proof, and how far it reaches

  • The proof. GPT-6.2 Astra proved that for −7 ≤ c ≤ 0 and 0 ≤ A ≤ 1.2 the game has exactly one equilibrium, even or not. The proof is in corpus/mfg-two-models/astra-report.tex. It bounds every possible solution, then shows that two different solutions would force a quantity above 14.18 to be at most 14.
  • Its readers. I read it with Claude line by line and found no error. Claude Fable 5.2 read it in the cross-read, filled eight steps the text leaves implicit, recomputed every constant and found none either. Both readers are the same model, and the second knew the result of the first. No one outside that family has read it.
  • The proof as an instrument. The proof uses only the largest |c| and the largest A of its rectangle, so its last inequality decides any rectangle. labs/mfg/uniqueness.py evaluates it in exact rational arithmetic and finds, for each depth, the most negative coupling at which it holds. Claude Fable 5.2 computed the same curve independently in the cross-read.
  • On the map. 1,786 of the 8,086 undecided cells lie inside the region where the inequality holds, 40.1% of the map’s area; the rectangle as its author stated it covers 1,344 cells, 35%. No cell the map proves to hold two equilibria lies inside. The published map is not changed by this page.
  • Against the truth. By Claude Fable 5.2’s fold table, the part of the map with c < 0 where the equilibrium is unique is 56.1% of the area. The proof reaches 71.5% of it. Claude Fable 5.2 also showed why this kind of argument cannot pass c = −π²: it is uniform down to A = 0, where uniqueness fails beyond that value.
  • A second, weaker proof, and what a reader found in it. Claude Fable 5.2 proved uniqueness when |c|·exp(3|c| + 3A/π) < π², a thin strip. I had read that proof and reported no error. GPT-6.2 Astra, reading it afterwards, found that it never proves an equilibrium exists (and supplied the argument), that one entry of its table was rounded the wrong way, and that one step tightens to give exp(2|c| + 2A/π). The uniqueness itself holds.
Depth A up toUniqueness proved for c ≥First fold (floating point)Gap
0.0125−9.72−9.980.25
0.1125−9.24−10.341.10
0.3125−8.65−10.802.15
0.6125−8.00−11.323.32
0.9125−7.46−11.754.29
1.2−7.03−12.135.10
3 · the first fold

Two derivations, one curve

4 · finding by finding

Who found what, and what each reading found

CERTIFIED is the certifier’s word. PROOF READ means a written proof was followed line by line by the readers named and no error was found; it is not a formal check. SUPPORTED is numerical agreement only.

FindingFromStandingWhat the check or the reading found
With phi = sqrt(m) the game is the stationary equation -(1/2) phi’’ + c phi^3 + A cos(2 pi x) phi = rho phi with int phi^2 = 1, and every real periodic solution with mean density one has a positive density.GPT-6.2 Astra, Claude Fable 5.2PROOF READholds
At A = 0 the number of even equilibria is exactly 1 + 2 #{j >= 1 : pi^2 j^2 < -c}; the non-constant ones are dn elliptic profiles with 4 j^2 K E = -c.GPT-6.2 Astra, Claude Fable 5.2PROOF READholds
Five equilibria coexist near c = -90 for A > 0.GPT-6.2 Astra, Claude Fable 5.2CERTIFIEDdecided by the certifier
Seven equilibria coexist at c = -360, for A = 0 and for A = 0.1.this sessionCERTIFIEDdecided by the certifier; the candidates were built in this session from the count at A = 0
There is exactly one equilibrium, even or not, for c in [-7, 0] and A in [0, 6/5].GPT-6.2 AstraPROOF READholds; the final bound recomputed three times (14.18 against the 14 needed). Limit: both readers are Claude Fable 5.2, and the second knew the first reading’s result
The equilibrium is unique when |c| exp(3|c| + 3A/pi) < pi^2, for every A.Claude Fable 5.2PROOF READthe uniqueness holds; GPT-6.2 Astra found that existence was asserted and not proved, and supplied the proof; GPT-6.2 Astra found one table entry rounded the wrong way (0.6098 printed for 0.60975); GPT-6.2 Astra widened the condition to |c| exp(2|c| + 2A/pi) < pi^2. Missed at first: I had read this proof and reported no error; it did not notice the missing existence step or the rounding
The first fold is at A = (4/3) sqrt(2/(3 pi^2)) d^(3/2) (1 + d/(16 pi^2) + ...), d = -c - pi^2.GPT-6.2 Astra, Claude Fable 5.2PROOF READholds; the two models’ leading terms are algebraically the same; a search reproduces Claude Fable 5.2’s table to five decimals
Three equilibria right next to the first fold, in cells the published map left undecided.GPT-6.2 Astra, Claude Fable 5.2CERTIFIEDdecided by the certifier
The count at A = 0 persists for small A near every coupling that is not a threshold; at least five equilibria on c in [-91, -90], A up to 2^-41, proved analytically.GPT-6.2 AstraPROOF READholds; all 26 inequalities of the box proof recomputed. Limit: one reader
The second and third folds leave -4 pi^2 and -9 pi^2 like 0.13341 A^(4/3) and 0.0054024 A^2.GPT-6.2 AstraSUPPORTEDClaude Fable 5.2’s independent table gives 0.005402 for the third and a value falling toward 0.1334 for the second; the proof has had no reader but its author
For every A >= 0 the number of even equilibria is 1 + 2 #{j : c < c_j(A)}.Claude Fable 5.2FALSE AS STATEDGPT-6.2 Astra’s counterexample is on the fold itself: at A = 0, c = -pi^2 the equilibrium is unique, and for small A > 0 there are two equilibria on the first fold where the formula counts one; away from the folds the statement is open. Missed at first: I had listed this conjecture as open
There is no equilibrium that is not even, for A > 0.Claude Fable 5.2OPENneither model’s search found one; GPT-6.2 Astra proves it for small A when -c < 4 pi^2, and that proof has had no reader but its author
prior work

What was already known

A short search, not a survey. Nothing on this page should be called new before someone who knows this literature has read it.

Not found: a uniqueness result for this game with a potential, with an explicit region; a certified coexistence of five or seven equilibria.

who did what

Each step is a separate job

check it yourself

How to re-read it

contact

Write to Carlos

Sent by Web3Forms to Carlos’s inbox; nothing stored here.

Or write to carlos@carlostoledo.co, or open the public pilot form on GitHub.