Proof, IP and disclosures

For diligence

How each result’s own check works

Each result on this site has a check written by the lab that re-reads or re-runs its evidence. Below, for each one: what the check does, how the lab tried to make it fail, and what that means for a buyer. No outside firm has audited any of them.

• A fast coupling model, graded by outside solvers

  • The graders, FastCap and Palace, are programs the lab did not write; the lab ran them itself, against a baseline it defined.
  • The lab’s own check was run against a deliberately broken copy and failed as it should.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • That test could not be run for this result, because the program found no number in the claim that it could change.
  • In a clean copy of the code the Palace step fails because the solver is not where the check looks.

  What it means for a buyer. The graders, FastCap and Palace, are programs the lab did not write; the lab ran them itself, against a baseline it defined, so a buyer would want to repeat the grading on their own installation of both programs.

• Fixed-size memory for reordering AI network data

  • The lab’s own check recomputes the larger headline ratio from the saved results of the simulated grid and failed as it should when one measured cell was doubled.
  • It does not re-run the simulation, so a wrong simulator would still pass it, and the figure against STrack is checked only to two decimal places.
  • Every number comes from the lab’s own simulator, and nobody outside the lab has checked any of it.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • That test could not be run for this result, because the program found no number in the claim that it could change.
  • The lab also ran adversarial tests against the design, which it calls an “escape sweep”; STrack, the published design the main figure is measured against, was left out of that sweep by name until 2026-08-19.
  • The lab says it did not move its thresholds (its pass marks) to fit STrack.

  What it means for a buyer. Every number comes from the lab’s own network simulator, and nobody outside the lab has checked it. The lab’s check recomputes the larger headline ratio from the saved simulation results (the figure against STrack is checked to two decimal places), so it would catch a mis-copied figure but not a wrong simulator; a buyer would want to re-run the simulation, or measure on hardware.

• Checked brightness ranges for chip prints

  • The lab’s own check confirms that its saved results file is unchanged and still carries the claim’s figures.
  • It failed as it should when a claimed figure was changed, but it does not re-run the simulation.
  • No outside tool or party has checked the result.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted, because the lab’s saved record of it no longer matches its signature (the stored mark that shows a file has not been altered).

  What it means for a buyer. Every number is relative to the lab’s own imaging simulation, and no outside tool or party has checked it. The lab’s check confirms that the saved results still carry the claimed figures but does not re-run the simulation, so a buyer would want to re-run it on their own imaging model.

• Crash recovery for AI assistants’ multi-step changes

  • The check is the lab’s own script for this result.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The check caught the change and failed, as it should.
  • That separate program is still inside the lab’s own programme, and the test world, the test scenarios and the crash points are the lab’s own.
  • So this shows that the check can fail; it is not an outside grading of the recovery layer.

  What it means for a buyer. The lab showed that its crash check can fail: when a separate program changed a number the check reads, the check caught it. The test world, the scenarios and the crash points are the lab’s own, and no outside party has graded the layer, so a buyer would want to run it against their own tools.

• The capacitance flat models leave out

  • The lab’s own check compares its solver with shapes whose exact answers are known and, when FastCap is installed, with FastCap; if FastCap is missing, the check skips that step instead of failing.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change the figure the claim quotes, in a file that sits outside the claim’s own evidence.
  • The check caught the change and failed, as it should.
  • But the lab’s own list of which figures its checks read shows that no check in its codebase reads this claim’s headline figure.
  • So the test shows that a check can fail when a figure is changed; it does not show that the headline figure itself is guarded.

  What it means for a buyer. The lab’s solver was checked against shapes with exact answers and against FastCap, an outside solver, run on the lab’s own mesh. No automated check in the lab’s code recomputes the headline share on each run, so treat it as a measurement reported once.

• An Ultra Ethernet recovery design, proved in a model

  • Four checking tools (a solver, a breadth-first search and two model checkers) each confirm the property on the lab’s one model, and the standard’s own sentence is what the model encodes.
  • But all four check the lab’s single encoding, so a modelling mistake would pass all four together.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted, because the lab’s saved record of it no longer matches its signature (the stored mark that shows a file has not been altered).

  What it means for a buyer. Four checking tools confirm the property, but all four read the lab’s single model of the standard, so one modelling mistake would pass all four. A buyer would want the same rule checked in a model they wrote themselves.

• Pair-by-pair coupling estimates overstate the worst case, in a model

  • The lab’s own check re-runs the search and failed as it should when the stored minimum was made tighter.
  • An earlier test that changed only the printed headline figure did not make it fail; the lab’s record calls that result inconclusive.
  • No tool outside the lab grades this: the model, the search and the check are all the lab’s own.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted, because the lab’s saved record of it no longer matches its signature (the stored mark that shows a file has not been altered).

  What it means for a buyer. The lab’s check re-runs the whole search, and it failed as it should when the stored minimum was made tighter. The model, the search and the check are all the lab’s own; no outside tool grades this.

• Tool combinations that leak a secret

  • The lab’s own check re-runs the analysis and failed as it should on a deliberately broken input.
  • The tools, the test world, the secret and the judge are all the lab’s own.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted, because the check rebuilds the file that holds the number each time it runs, which wipes out the change before the check reads it.

  What it means for a buyer. The lab’s check re-runs the analysis, and it failed as it should on a deliberately broken input. The tools, the test world, the secret and the judge are all the lab’s own, so a buyer would want the same run on their own assistant’s tools.

• Where deletion receipts can be fooled

  • The lab’s own check re-runs the whole test campaign, in which the lab tries to fool both of its receipt checkers.
  • To show that the check can fail, the lab changed one checker so that a step which should refuse a bad receipt accepted it, and the check failed, as it should.
  • The check confirms only the campaign’s overall result, not each of the six ways the campaign shows a receipt can be fooled.
  • Everything runs over an ordinary file in a simulation, and nobody outside the lab has graded it.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted, because the check rebuilds the file that holds the number each time it runs, which wipes out the change before the check reads it.

  What it means for a buyer. The lab’s check re-runs the whole attempt to fool both receipt checkers, and it failed as it should when one checker was deliberately weakened. It confirms the overall result, not each failure mode one by one, and nobody outside the lab has graded it.

• Signing off some tiles near a chip-mask edit, in simulation

  • The lab’s own check confirms that the evidence file is unchanged and still carries the claim’s figure.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The check caught the change and failed, as it should.
  • The check does not redo the point-by-point comparison.
  • Another, longer command in the lab’s codebase does redo it, and nothing on this site runs that command.
  • Everything is relative to the lab’s simulator, not to silicon.

  What it means for a buyer. The lab’s check confirms that the evidence file is unchanged and still carries the claimed figure; the point-by-point comparison itself is redone by a longer command in the lab’s code, which this site does not run. Everything is relative to the lab’s simulator, so a buyer would want to re-run it on their own masks.

• An automated attack search against quantum-safe Wi-Fi sign-in

  • The lab’s own check re-runs the search.
  • As a control, when the lab re-opened one known attack (KRACK, a key-reinstallation attack) in its models, the same search found an escape and the check failed as it should.
  • The attacker, its list of moves and the verifier are all the lab’s own, and the target is the lab’s Python models, not real Wi-Fi software.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted, because the check rebuilds the file that holds the number each time it runs, which wipes out the change before the check reads it.

  What it means for a buyer. As a control, when the lab re-opened one known attack (KRACK, a key-reinstallation attack) in its models, the same search found it, so the search found the one known attack the lab put back. The attacker, its moves and the models are all the lab’s own; it is not a test of real Wi-Fi software.

• A memory ceiling for post-quantum Wi-Fi, proved in a model

  • The Lean proof checker, a tool the lab did not write, confirms that the proof follows from the lab’s statement of the model.
  • The lab’s own check was deliberately broken twice, with a doubled per-connection buffer and with a step of the proof left unfinished, and failed both times as it should.
  • The Lean model is written by hand to mirror the Python code and nothing extracts it from the code.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted: the program could reach only one number, and the check kept passing even with the file that holds it emptied, so the check does not read that file.

  What it means for a buyer. The Lean proof checker, a tool the lab did not write, confirms that the proof follows from the lab’s model. The model is written by hand to mirror the lab’s Python code, so a buyer would want the same cap checked against their own firmware.

• The least bookkeeping a shared AI cache needs, in a model

  • The Lean proof checker, a tool the lab did not write, confirms that the proof follows from the lab’s statement of the cache model.
  • It does not check that the statement matches any real cache’s permission rule.
  • The lab’s own deliberately broken copy changed only a worked example, not the general theorem, so it does not show whether a change to the theorem itself would be caught.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • That test could not be run for this result, because the program found no number in the claim that it could change.

  What it means for a buyer. The Lean proof checker, a tool the lab did not write, confirms that the proof follows from the lab’s model of the cache. Nothing checks that the model matches a real cache’s permission rules.

• A near-constant-size sign-off record for a photomask

  • The lab’s own check re-hashes the two files it cites and confirms that the claim’s figures appear in them; it does not rebuild or re-verify a record.
  • It failed as it should when a figure was changed.
  • Another, longer command rebuilds the record.
  • No tool outside the lab is involved.
  • The lab also tried to break its own check on purpose.
  • It had a separate program, one that did not build this result and could not see what the check expects, change a number the check reads, to see whether the check would notice.
  • The result of that test cannot be counted, because the lab’s saved record of it no longer matches its signature (the stored mark that shows a file has not been altered).

  What it means for a buyer. The lab’s check confirms that the two files it cites carry the claimed figures; it does not rebuild the record, and no tool outside the lab is involved.