Carlos Toledo · cert-machine

The machine proves it — or breaks it

AI produces mathematical claims faster than anyone can read them. This machine does both ends of the job with the same exact arithmetic. It has settled a value conjectured open since 2019, certified what is — to our knowledge — the first hot-spots domain beyond every class analysts have reached, and refuted a published constant at its twelfth digit. Every verdict is a re-runnable certificate: proved, disproved, or honestly refused — never a probability argument.

No probability arguments and no digit-matching. A claim is admitted only by exact arithmetic on whole numbers, and an instrument that cannot decide refuses instead of guessing. When a page here says REFUTED, that is a proof, and the falsifying witness is printed beside it.

start here

Two theorems and two audits, in plain words

a theorem · hot spots, asked 1974The hot spot sits at one corner — provedWhere does heat concentrate in a still room? Rauch asked in 1974; fifty years of proofs cover triangles (an Annals paper), symmetric shapes, and little else — and in high dimension the conjecture turned out to be FALSE, which makes the flat convex case the live one. For a lopsided trapezoid outside every proven class, this machine certified the answer: the extremes live on the boundary, and the hottest point is exactly one vertex. To our knowledge the first certified hot-spots domain beyond every analytically proven class.the whole proof re-runs in ~2 minutes · archived with a DOI a theorem · erdős #510A number that was a conjecture is now exactMercer proved the first two values of Chowla’s cosine dip in 2019, conjectured the third, and wrote that he did not know how to evaluate it. The machine executed his own strategy to the end: λ(4) is the root of 512y³ − 1227y² + 600y + 125 near 1.51956 — a cubic no paper had printed — with every threshold derived rather than transcribed and every finite case decided in exact arithmetic. An independent audit sharing no code with the prover walked all 25,819 cases: zero holes, zero refuters.re-proved at every build · archived with a DOI a refutation · erdős #852The constant that was a rounding errorA constant for an open Erdős problem, published with frontier-model help, quoted to 13 decimal places, with no error bound. It is wrong from digit 12 — and the wrong digits are exactly what an ordinary floating-point loop prints. The constant was not near-right with unlucky endings. It was the bug, published. The corrected value is certified here and the correction is now public in the problem’s own thread.refuted at digit 12 · correction certified and public six theorems · re-verifiedWe checked the AI’s homeworkSix mathematical results produced with frontier-model help — among them a counterexample to a bound of Maxwell’s and an Erdős problem open since 1958 — re-verified here from the manuscripts, by code that never ran a line of the authors’. 5 held, 1 came back partial, none was refuted. The finding is not the tally: in all 6, the part a machine can check held, and the part that carries the theorem stayed out of reach.92 checks · 21 deliberate forgeries, every one rejected
the theorem, in full

How a trapezoid got a theorem

The domain is deliberately inconvenient: a convex trapezoid with side slopes 6 and 18/5, no symmetry axis, nothing any existing proof technique can grab. The machine proves its second Neumann eigenvalue is simple by certifying a spectral gap — upper bounds from an interval Galerkin method, lower bounds from exact-rational finite elements with eigenvalue counts by interval inertia. It then builds a trial function that solves the eigenvalue equation EXACTLY — a sum of Bessel fans anchored at the four corners — and certifies that its boundary defect is a hundred-thousandth, which pins the true eigenfunction within an explicit distance of the trial.

Then the geography: every interior point is assigned, in exact rational arithmetic, to a deep core, a boundary collar, or a corner sector — and each region is killed by its own argument. Core and collar cells die by comparison against certified interior witnesses; the corner sectors die by certified series expansions, where the delicate corner needs a Bessel ladder identity to tame a divergent second derivative whose singular part arrives, provably, with the helpful sign. Zero cells survive. The extremes are on the boundary, the hottest point is vertex A and nowhere else, and the whole chain — eight records, every red control firing — re-runs from one command in about two minutes. The full account:

the report · the archived release.

the refutation, in full

How a floating-point bug became a published constant

Erdős problem #852 has a constant attached to it. In 2026 a value for that constant appeared, produced with frontier-model help and quoted to 13 decimal places. This machine enclosed the same constant in exact arithmetic. The two agree for a while and then they do not.

published  0.0752403861777
certified  0.0752403861783092455893…
                        ↑ wrong from here

The interesting part is which wrong digits. The obvious way to compute this constant is a loop: take a few million prime numbers, turn each into a factor slightly larger than one, and multiply them together in ordinary floating point. Past a certain size those factors are so close to one that rounding makes them exactly one, and they stop contributing anything at all. The running product goes still. Going still is what convergence looks like, so the loop appears to have settled, and it prints.

The published constant is what that loop prints, digit for digit. It was not approximately right with unlucky endings. It was the bug.

That is why one wrong constant is worth a whole page. Every ordinary defence fails against this. Rerunning it reproduces the same wrong digits, because two independent floating-point implementations agree with each other rather than with the truth. Spending more compute changes nothing, because the loop is already ignoring almost every factor you would be adding. Checking the digits against a reference value fails whenever the reference came out of the same kind of pipeline — which is how a bad number gets into an answer key and stays there.

What settles it is arithmetic that never rounds. The refutation here is a strict inequality between two whole numbers, with a denominator millions of digits long and no approximation anywhere in it. The corrected value is trapped between two exact fractions that agree for 15 decimal places, so the true constant cannot be anywhere near the published one.

The correction has been public in the problem’s own thread on erdosproblems.com since 27 August 2026. The full audit → — the refutation as integers, the certified correction, and a catalogue of the other ways a mathematical answer key goes wrong without anyone noticing.

the shelf

The record, ordered by weight

the eval · live boardThe matmul eval: ground truth is a proofFrontier models are asked for exact rank-R matmul tensor decompositions; every proposal is certified or refuted in exact rational arithmetic. No judge, no rubric, no answer key to contaminate — a proof either exists or it does not.364 proposals graded · zero subtly-wrong survivors audit · AI-discovered algorithmsThe AI-discovered algorithms, certifiedAlphaEvolve’s rank-48 ⟨4,4,4⟩ certified over Z[i]; AlphaTensor’s rank-47 verified over F2 and REFUTED over Q — the speedup provably requires characteristic 2. Both decided from commit-pinned bytes at every build.10 algorithms re-decided each build certified theorems · MFG splittingThe crowd splits — MFG beyond the uniqueness wallA congestion-averse crowd in a single-well cost landscape provably settles into TWO density peaks — or THREE. To our knowledge the first validated-numerics equilibrium enclosures for a mean-field game, and the first whose certified peak count strictly exceeds the potential's wells. Two computer-assisted theorems and a bracket table of certified instances whose enclosure balls fix the exact peak count of the exact solution; the discount-free crossover σ* = 1/(8π²) decided in exact rationals; certified multiplicity where uniqueness theory is silent; an EXACTLY-3 census; the 21,567-cell regime map.two theorems + a bracket table · honest counting, every bound displayed erdős #1038 · the supremum sideHow shallow can a lemniscate stay?Tao reformulated Erdős #1038 over discrete measures and conjectured the supremum of |{U<0}| is 2√2 — "this may be hard to prove completely." For rational weights it is a per-degree polynomial question, and the machine decided the first seven degrees: odd degrees fall strictly below 2.82 < 2√2 (branch-and-bound certificates, 127 boxes for the cubic); even degrees localize the supremum to [2√2, 2.82845] with the two-atom witness within 2.3×10⁻⁵ of optimal. Plus the certified landscape: an interior cubic champion and a quintic cliff where the sublevel set changes topology.per-degree theorems · the conjecture itself stays open · not peer-reviewed verified reward · demoThe verifier in the loopThe reward channel in closed loop: a model proposes, the grader answers every failure with its own refutation mechanism — the exact violated equation, nothing more — and the model retries. Every trajectory rendered from the append-only ledger.feedback template-locked · zero coaching the eval · forecastingThe Forecast Gym: the test set that cannot leakForecasts are sha-committed to an append-only ledger before their outcomes exist and scored afterward with a proper score in exact rationals — contamination impossible by construction, hedging and overconfidence both priced, admission prune-only with an exact binomial certificate.130 forecasts sha-pinned before their targets

All 42 reports → — the AI-verification shelf, the Erdős problems, the applied fronts in aerospace and energy, and the classical ground the instruments were proven on first. Every number on every page is recomputed from the certificates and records at build time, and a build that drifts refuses to ship.

why you can trust this

One rule, and what it costs

Everything here rests on a single rule, and most of the engineering is the cost of keeping it.

the machine

How a claim becomes a certificate

the machine
Generate at scale, screen in float, certify the survivors exactly. Select any stage for what it does — every count is read off ledger.json at build time.
ENUMERATE · 11 FAMILIES 819,152 objects SCREEN · FLOAT 17,741 pass CERTIFY · EXACT 16,943 decided HIT · CERTIFIED 2,274 REJECT · PROVED 14,668 REFUSED · HONEST 1 LEDGER · THE GATES ledger.json · 47/47 batteries dedup by key only certificates
The loop, in five stops. The screen may only prune; the instruments alone decide; REJECT and REFUSED are terminal by design. The full drawing — every family, every instrument, the closed-form hunt — is on the control page.

The control page carries the full drawing, live: every family, every instrument, every battery executed at its build (never remembered), the full ledger decomposition, drift status.

the tally

What it has decided so far

theorem programs
3
λ(4) exact and λ(5) closed (λ(6) one family from the non-monotonicity theorem); the hot-spots trapezoid with its vertex-A corollary; the MFG splitting pair with its bracket table — every one re-proved or re-walked at build
published claims decided
1 refuted · 1 corrected
a constant on Erdős #852, wrong from digit 12 · one printed Ramanujan Machine row, a transcription slip — both replacements certified
AI-claimed theorems re-verified
6
5 held, 1 partial, none refuted — checked from the manuscripts, never from the authors’ code
AI-discovered algorithms re-decided
10
at every build, from commit-pinned bytes — AlphaEvolve’s rank-48 ⟨4,4,4⟩ certified over Z[i]; AlphaTensor’s rank-47 verified over F2 and refuted over Q
model proposals graded
364
115 certified, each an exact theorem · none subtly wrong — nothing false has ever been graded correct
closed forms disproved
54,628,296
each one an exact proof, out of 54,629,173 tested — and zero discoveries claimed
Hénon points, counted exactly
EXACTLY 1,696
period 16, with a proof that there are no others anywhere in the plane — 1,419,655,025 boxes exhausted. The classical case the instruments were calibrated on before they were pointed at anything new.
the same instrument, pointed at the world

SkyAudit — one real day over New York, every flight decided

A real, hash-pinned day of New York helicopter traffic — 82 aircraft, 382 flights — replayed on a map. Every trail is coloured by a verdict rather than by telemetry: each flight is re-flown on paper by an electric aircraft, using that manufacturer’s own published numbers, under the FAA’s energy-reserve rule, and decided.

Beta’s ALIA can certifiably cover 100 of the 382 flights and needs exactly 10 aircraft to do it — the lower half proved by pigeonhole, the upper half by an actual schedule. Joby, Archer and Eve publish too little to certify a single fleet, and the app says NEEDS DATA rather than guessing, with the exact number that would flip it printed beside the verdict.

Open SkyAudit → — the replay, the certificate panel, the fleet frontier and the what-if sliders, live. Data © adsb.lol (ODbL). Also live: the São Paulo pack — the world’s busiest urban helicopter market, decided under Brazil’s own reserve rule.

for people building evals

A grader that cannot be fooled

The instrument that audits a published claim also grades a model’s output, and that is turning out to be the more useful job. A model proposes an exact object; the grader re-derives it from whole numbers and answers CERTIFIED or REFUTED. There is no judge model, no rubric, and no stored answer key — so there is nothing to leak into a training set and nothing to game. The gap between “graded correct” and “is correct” that a policy would learn to exploit does not exist, because the grade is the proof.

That is measured, not asserted. Every campaign opens with deliberate forgeries — including one wrong by a billionth, invisible to any floating-point check — and if a single forgery grades as correct the run aborts before it touches real work. Across every real-model campaign so far, 364 proposals, nothing false has ever been graded correct and no certified row has ever turned out to be wrong.

The honest limit: this works for claims that come down to finitely many exact arithmetic facts — exhibit a witness, verify an identity, bound a quantity. It does not work for mathematics at large, and nothing here pretends otherwise. Inside that boundary the same grader can sit unchanged inside a training loop, with the verifier strictly stronger than the thing it is grading.

The oracle, packaged → — one curl, no dependencies, running on your laptop in under a minute: the claim schema, the tool definition a model calls mid-generation, the paste box, the paper draft, the ledgers.

rerun a proof

Check a result yourself, in ten seconds

Every headline claim detaches into a certificate — a JSON file of exact numbers — plus a verifier in plain Python: standard library only, nothing to install, zero code shared with the engine. Each verifier re-derives the mathematics from the certificate alone, re-hashes the pinned sources, must refute a deliberately forged value before it will exit green, and prints the sha256 of the certificate it checked.

python3 verify/verify_erdos852.py certs/erdos852-certificate.json
python3 verify/verify_keller.py   certs/keller-certificate.json
python3 verify/verify_strassen.py certs/strassen-certificate.json
sources

Run from a clone of the repository (add --sources corpus/sources to re-hash the pinned source bytes too), or download the certificate and verifier right here — the proof travels without the machine.

the certificates

Proofs that travel without the machine

certificatewhat it holdsre-verify
erdos852-certificate.jsonBoth Erdős #852 constants as exact data: the c0 window re-decidable at 130 digits, the C∗ refutation as strict integer inequalities with no tail bound.verify_erdos852.py
erdos852-h-records.jsonEvery record run of pairwise-distinct consecutive prime gaps the scan has closed — index, opening prime and length. The head reproduces OEIS A078515/A079889 term for term; the tail passes them. Each record beyond the published terms is re-proved at build by an independent Miller–Rabin verifier.battery-gated
ai-claims-summary.jsonThe six-lane AI-claim audit as the build recorded it: every lane’s verdict, scope, check count and mutation-control count, written by the report builder from a live run of all six verifiers. Not a certificate — a record of what the verifiers said, so the shelf card and the page cannot quote different numbers.battery-gated
keller-certificate.jsonThe Jacobian/Hessian counterexample corpus — every polynomial as explicit exact rational monomials; determinants and collisions re-derivable from the file alone.verify_keller.py
strassen-certificate.json10 fast matrix-multiplication algorithms as exact tensor identities over Q and F2 — including AlphaTensor’s rank-47, decided both ways.verify_strassen.py
bilinear-certificate.json9 bilinear algorithms for POLYNOMIAL multiplication over F2 — full, truncated and cyclic products — each found by the generation front’s free flip-graph walk and decided by instruments/bilinear, which rebuilds the target tensor from its name rather than trusting the scheme handed to it. Every entry stores its scheme in full, so any reader can re-decide it; the published bounds each one is measured against live in corpus/bilinear-bounds.json and are not results of this repository.battery-gated the report
lambda4-audit.jsonThe adversarial audit of the lambda(4) proof: an independent clause walk sharing no code with the engine — inner products by direct trigonometric summation, family and subfamily membership by plain integer arithmetic, thresholds read from the campaign record, finite-clause sets re-certified fresh by the calibrated instrument. Every gcd-reduced 4-set in the box must be reached by an explicit clause of the proof (generic, family dot, closure, finite, delegated, or the definitional witness); a set with no clause is a hole and aborts. Also carries the full theorem sweep of the box: zero refuters.battery-gated
lambda4-campaign.jsonThe lambda(4) campaign record. Phase 0: Mercer's Section-5 strategy (INTEGERS 19 (2019) #A4) executed mechanically — lambda(2) and the whole lambda(3) proof re-derived with exception families discovered and thresholds derived, not transcribed; the lambda(4) generic case with its 14 exception conditions discovered and matched against the paper's hand-written list; the measured reduction: five of the fourteen carry strictly negative delta, so NINE families remain. Phase 1, in progress: each family closed so far carries its full derivation here — second-level dot theorem, subfamily cones, derived thresholds, decided finite parts. d = 2c is CLOSED. Re-derived symbolically at every build by the lambda4 battery; finite parts pinned here and sampled at every run.battery-gated
sublevel-tao179.jsonThe Tao #179 sublevel campaign (Erdős #1038, supremum side): rational-weight discrete measures on [-1,1] are monic root-constrained polynomials via |q| < 1, and this record holds certified sublevel measures — the 2√2 witness, grid champions including the interior cubic champion near 2.7542 and the quintic transition peak near 2.8011 — plus per-degree branch-and-bound THEOREMS: odd degrees strictly below 2.82 < 2√2, even degrees localized to [2√2, 2.82845] with (x²−1)^{N/2} attaining the left end. Every measure an outward enclosure from BigInt Sturm isolation; the box bound calibrated to equal the measure on thin boxes.battery-gated
mercer-mu5.jsonThe mu(5) ladder, mu(5) ≤ 1 + π/m rung by rung to m = 20 — every exceptional tuple closed by one exact rational evaluation.battery-gated
mu-table.jsonThe Newman min-modulus table: every set in the named boxes exhausted, champions certified, orbits classified, conservation per row.battery-gated
mu-table-40.jsonThe wider-box extension of the mu table — billions of sets exhausted, the narrow-box crowding artifacts corrected.battery-gated
lambda-table.jsonThe lambda table: the source lab’s rows reproduced exactly, plus rows that lab’s record does not hold and no table this lab has read holds — a claim about the certificates, not about priority.battery-gated
census-high-periods.jsonThe Hénon high-period census, p = 13..16: every period point of the classical map found and counted exactly, the plane exhausted box by box, each row re-checked with zero unmatched — the proving ground the interval instruments were calibrated on.battery-gated
entropy-henon.jsonThe certified entropy lower bound for the classical Hénon map: h-sets, covering relations, and the exact spectral argument.battery-gated
erdos290-tail-ext.jsonThe Erdős #290 sweep extension: degrees closed beyond the cited page, the constant’s enclosure tightened degree by degree.battery-gated
mfg2p-regime-map.jsonThe TWO-population regime map: every cell of the coupling plane with its verdict and its exact witness — the symmetric cross-coupling s against the attack-defense asymmetry d, two disjoint enclosures where uniqueness provably fails, the Lasry-Lions monotone strip where it does not, and the refusal reason everywhere else. There is no single-file certifier for this map: it is decided by labs/mfg2p/box2p.js and re-gated by that lab’s battery at every build of its report.battery-gated the report
mfg-regime-map.jsonThe mean-field-game regime map: every cell of the coupling–potential plane with its verdict and its exact witness — two disjoint enclosures where uniqueness provably fails, the monotone enclosure where it does not, and the refusal reason everywhere else.mfg-certify.js
matmul-eval-ledger.jsonlThe matmul eval’s append-only ledger — every campaign row, every verdict, every tag; the leaderboard is built from this file.battery-gated
chowla-records.jsonCertified Chowla merits c(A) = -min_x sum cos(ax) / sqrt|A|, one row per set size n, each an exact UPPER bound proved by instruments/trigmin for the set stored beside it. THIS FILE MAKES NO CLAIM ABOUT CHOWLA’S COSINE PROBLEM. Chowla’s question is asymptotic — whether c can be driven to 0 as n grows — and a low c at one n is a fact about that n. The measured trend here RISES with n (0.6558 at n=10 to 0.8205 at n=30), which is consistent with Chowla’s conjecture that the order is sharp, i.e. evidence against the direction, not for it. Explicit sets with small c are occupied literature (Mercer, INTEGERS 2019; Bedert, arXiv:2509.05260); these rows beat only the classical families recomputed here beside them.battery-gated
matmul-eval-corrections.jsonCorrections to rows already written in the matmul eval ledger. The ledger is append-only and is never rewritten, so a row that turns out to be MISLABELLED is corrected here and the correction is applied when the report displays it — currently one: 90 rows tagged v4-effort-low ran at the API’s DEFAULT effort, because the harness dropped its --effort argument in campaign mode. A correction naming a tag no row carries, or claiming a row count the ledger does not hold, refuses the report build.battery-gated
matmul-loop-ledger.jsonlThe verifier-in-the-loop ledger — every trajectory round with its verdict and the exact feedback sent; the loop report is built from this file.battery-gated
skyaudit-forecast-ledger.jsonlThe prediction ledger — interval FORECASTS committed before their target day (sha-pinned, append-only) and scored after in exact rationals; coverage claims are conformal counting theorems, never model faith. Wrong forecasts stay forever.battery-gated
forecast-gym-ledger.jsonlThe Forecast Gym’s append-only ledger — every proposer’s forecast sha-committed before its outcome exists, every score an exact Winkler rational; the gym report and its admission board are built from this file.battery-gated
lambda56-campaign.jsonThe lambda(5)/lambda(6) campaign record (Chowla’s cosine dip, the non-monotonicity program): lambda(5) = −L(1,2,4,5,6) closed in full — the generic case certified, all eight exception families closed with derived thresholds, 1725 finite sets decided, the extremizer walled — and lambda(6) in progress with nine of its ten families closed in this record. The double-sum-core obstruction theorem (no classical weight works on b+c = a+d = e) and the Fejér–Riesz comb weight that beats it are re-proved by the battery at every run.battery-gated
terra-sigmastar.jsonThe exact crossover of the MFG splitting program: σ* = 1/(8π²), discount-free, DECIDED IN EXACT RATIONALS — the crossover polynomial factors, the γ-coefficient is identically zero (checked k = 2..12), the band-pass identity and both harmonic windows are exact, π enters only as a Machin bracket of width 1.3e-44.battery-gated the report
terra-bracket-table.jsonThe bracket table under the splitting theorems: seven certified rows straddling both predicted thresholds — negatives below the amplitude threshold and past the crossover, replications, and the threshold pin r_c ∈ [0.13, 0.14] with the exact-rational prediction landing inside. Honest counting lives here: two theorems plus a table, never eight.battery-gated the report
mfg-cap-multiplicity.jsonCertified multiplicity for the ergodic quadratic MFG past its pitchfork: at each of six couplings, at least THREE distinct exact solutions enclosed in pairwise disjoint uniqueness balls with certified positive density — exactly where Lasry–Lions monotonicity is silent. The c = −9.5 monotone-regime boundary, where the branch collapses and no claim is made, is recorded too.battery-gated the report
attnflow-theorems.jsonThe attention-wing theorems: a rational-kernel token flow chosen so equilibrium and stability are DECIDABLE in exact ℚ — the consensus spectrum proved β- and p-free by exact dual-number expansion, the two-cluster cross-weights identically zero with the honest p = 1 boundary, the reduced flow’s double zero decided by exact division (every pitchfork claim refuted), and the phantom-bifurcation taxonomy with its live artifact.battery-gated the report
facelaw-theorem.jsonThe face-dimension law k = |shared| − cons + z, decided against the exact ℚ null space on two seeded 4,000-network ensembles; every instance where the natural shortcut fails (precisely the z > 0 cases) is ENUMERATED here so any reader can re-run any one.battery-gated the report
terra-recert-t1.jsonThe T1 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t1.jsonThe certified peak count for T1: the exact number of strict maxima of EVERY density in the T1 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
terra-recert-t2.jsonThe T2 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t2.jsonThe certified peak count for T2: the exact number of strict maxima of EVERY density in the T2 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
terra-recert-t3.jsonThe T3 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t3.jsonThe certified peak count for T3: the exact number of strict maxima of EVERY density in the T3 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
terra-recert-t4.jsonThe T4 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t4.jsonThe certified peak count for T4: the exact number of strict maxima of EVERY density in the T4 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
terra-recert-t5.jsonThe T5 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t5.jsonThe certified peak count for T5: the exact number of strict maxima of EVERY density in the T5 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
terra-recert-t6.jsonThe T6 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t6.jsonThe certified peak count for T6: the exact number of strict maxima of EVERY density in the T6 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
terra-recert-t7.jsonThe T7 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t7.jsonThe certified peak count for T7: the exact number of strict maxima of EVERY density in the T7 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
terra-recert-t8.jsonThe T8 enclosure of the congestion-MFG splitting program: a radii-polynomial certificate around the stored candidate — exact instance parameters, certified ℓ¹_ν ball radius, contraction bound Z₁, closure margin, and density floor — re-certified here against the frozen published verifier extended with the data terms; nine falsifiers fire per instance at every battery run.battery-gated the report
terra-peakcount-t8.jsonThe certified peak count for T8: the exact number of strict maxima of EVERY density in the T8 enclosure ball, derived from CERTIFIED region signs only (instruments/critcount) — outward coefficient products, ball Lipschitz folded into the cell pad, never a float sign.battery-gated the report
mfg-cap-census-N2-c-12.jsonThe Krawczyk exhaustion census at Galerkin level N = 2: every subbox of the printed box eliminated by interval residual or Krawczyk exclusion, each survivor isolated by K(X) ⊂ int(X) — EXACTLY three solutions of the truncated even system at c = −12, matched one-to-one to the candidates. Box-bounded, truncation-level; the PDE-level count is the stated open problem.battery-gated the report
mfg-cap-census-N3-c-12.jsonThe Krawczyk exhaustion census at Galerkin level N = 3: every subbox of the printed box eliminated by interval residual or Krawczyk exclusion, each survivor isolated by K(X) ⊂ int(X) — EXACTLY three solutions of the truncated even system at c = −12, matched one-to-one to the candidates. Box-bounded, truncation-level; the PDE-level count is the stated open problem.battery-gated the report
mfg-cap-census-N4-c-12.jsonThe Krawczyk exhaustion census at Galerkin level N = 4: every subbox of the printed box eliminated by interval residual or Krawczyk exclusion, each survivor isolated by K(X) ⊂ int(X) — EXACTLY three solutions of the truncated even system at c = −12, matched one-to-one to the candidates. Box-bounded, truncation-level; the PDE-level count is the stated open problem.battery-gated the report
mfg-cap-census-N5-c-12.jsonThe Krawczyk exhaustion census at Galerkin level N = 5: every subbox of the printed box eliminated by interval residual or Krawczyk exclusion, each survivor isolated by K(X) ⊂ int(X) — EXACTLY three solutions of the truncated even system at c = −12, matched one-to-one to the candidates. Box-bounded, truncation-level; the PDE-level count is the stated open problem.battery-gated the report
ember-spectrum.jsonStage 1 of the hot-spots chain: two-sided spectrum localization for the trapezoid — interval Galerkin Rayleigh–Ritz uppers (floats only pick the subspace), exact-rational Crouzeix–Raviart assembly with interval LDLᵀ inertia counts and Liu’s framework for the lowers; the certified gap makes μ₁ SIMPLE. The rectangle regression encloses π² at every run.battery-gated the report
ember-defect.jsonStage 2: the frozen Helmholtz trial’s boundary defect — order-2 interval Taylor jets along every edge with Bessel-ODE closure, a certified midpoint-Taylor cell rule, value AND derivative bridges against an independent float evaluation, and the exact trial coefficients frozen for every downstream stage.battery-gated the report
ember-eigenpair.jsonStage 3: the eigenpair certificate — the boundary-residual identity with the rational star-shaped trace constant and the CR localization tightens μ₁ by a factor of ~105 and encloses the eigenfunction in L²; also the H¹ error and the δλ bound the pointwise machinery consumes.battery-gated the report
ember-pointwise.jsonStage 4: the solid-mean pointwise machinery — kernel norm I₀ = 5/48 DERIVED in exact rationals, witness balls decided inside Ω in exact rationals, and every CORE cell of the 1/100 grid (corner-min depth ≥ 3/40, exact by concavity) killed on both sides with zero survivors.battery-gated the report
ember-collar.jsonStage 5: the collar sweep — every sub-core cell killed by the value argument with REFLECTED pointwise bounds across its nearest open edge (the single layer bounded by the certified per-edge flux sups), kill-or-refine to 1/800, ZERO residual cells.battery-gated the report
ember-corner.jsonStage 6: the four corner-tip certificates — Bessel–Fourier coefficients certified by annulus L² extraction AND re-extracted at a second annulus (the enclosures must intersect — a condition of entry), value kills at B/C/D, radial monotonicity at A, and the ladder-identity wedge bound at C where the boundary minimum lives.battery-gated the report
ember-cross.jsonThe independent cross-derivations: I₀ = 5/48 in exact rationals, the trace constant re-derived on directed dyadic big-floats from the exact-rational star geometry, μ₁ bounded above on an independent conforming P1 basis, and the two-annulus corner condition re-asserted.battery-gated the report
ember-theorem.jsonThe assembled hot-spots theorem: cross-record chain consistency (every stage’s inputs equal the upstream outputs), the interior partition RE-DECIDED IN EXACT RATIONALS, every sweep and tip verdict re-checked, the honest framing with its fence list, and the sha256 of every input record.battery-gated the report

Code, corpus, and full provenance: github.com/carlostoledo1891/cert-machine — MIT, no dependencies. Instruments lifted from a private source lab are hash-pinned in PROVENANCE.json; patches are declared so they can never be mistaken for drift.

the limits

What this is not

This is one machine and one operator. The trust base is stated rather than hidden: V8’s big-integer arithmetic and IEEE-754 correct rounding are assumed correct, and a handful of named external theorems are consumed and cross-checked rather than machine-proved. Nothing here is a formal proof in the sense of Lean or Coq, and no page claims to be.

What it does meet is the working standard of the computer-assisted-proof tradition — Tucker on the Lorenz attractor, Galias on the Hénon censuses, whose published counts this machine reproduces independently. That is one rung below a formal proof and several rungs above a decimal that looked convincing.