1. The Integrity of the Audit
The previous section reported what the archive found. This section states what a lock means, how it is to be read, and the single point at which it can be deceived. A method that certifies warrant must say precisely what its certificate covers and what it does not, because a certificate read for more than it claims is worse than no certificate at all. The discipline below is the defense of the audit's integrity, and it is also, at one point, the honest admission of where that integrity ends.
The one claim a lock makes
A lock certifies exactly one thing: that on the warrant rows supplied at the time of audit, three mass-bearing lines of support stand independent and enclose a genuine volume. That is a real and hard-to-forge certificate, because independent multi-register support is difficult to assemble and the determinant detects when it is absent. It is not, and was never, a second claim, that the rows faithfully represent the world. These two claims share a subject and must never be joined. The first, the independence of the warrants as supplied, is what the lock seals. The second, the faithfulness of those warrants to reality, the method cannot certify at all, and reports as permanently open.
The consequence is the spine of the defense. A lock computed on faked rows certifies the independence of the faked rows, and falsifies whoever supplied them, not the method, because the method issued no certificate about their faithfulness to begin with. The lock is honest about its own reach, and that honesty is what converts the dependence on supplied data from a flaw in the method into a charge against a supplier. The instrument reads what is placed before it and claims nothing about where it came from.
The three blindnesses, and the single misuse
The verdict is provably blind to three quantities, and each blindness is a feature that forbids one misuse. It is blind to the truth-sign of a proposition: the determinant for a claim and for its denial are numerically identical, because the underlying functional is a squared quantity invariant under reflection of any axis, so the truth of a claim must be read from the content of its axes and never from the lock. It is blind to the soundness of the conclusion: the lock certifies that the warrants are independent, not that the conclusion they support is true, so the final verdict rides a determinacy witness that reads the axes, and the lock licenses that reading without performing it. And it is blind to the source of the warrants: the determinant reads the magnitude of the out-of-plane content and not its origin, so the faithfulness of the rows must be read from an independent generator and never from the determinant.
Behind all three is one error: reading off a lock a quantity the lock is provably blind to. Every misuse of the method is an instance of it. Reading truth off the lock, reading the sign off the lock, and reading faithfulness off the lock are one mistake in three costumes, and the correct use in each case is the complement, to read only what the lock provably carries, which is the count and the mutual independence of the warrants supplied now, and to take truth, sign, and faithfulness each from its proper source.
What the lock licenses, and what it does not
The rule of use follows directly. A lock licenses four readings and no others. It licenses the finding that three independent warrants stand at mutual right angles at the rows supplied now. It licenses a conditioning reading, whether the lock is well-formed and far from collapse, carried by the condition number and the position in the basin. It licenses the extraction that a determinacy witness then performs, the lock opening the door that the witness walks through. And it licenses a snapshot claim at the data available now, defeasible and open to revision on new evidence.
A lock does not license the truth of the proposition, nor its sign or direction, nor the soundness or finality of the conclusion, nor the faithfulness of the supplied rows, nor a forecast or an inevitable future, nor any verdict outside the register that was audited. The one-line statement of correct use is this: use the lock to certify that the warrants supplied now are independent, and read truth, sign, soundness, faithfulness, and futurity each from its proper source and never from the determinant.
The two surfaces on which a verdict can fail
A verdict can fail on two surfaces, and they are categorically different. The first is internal, the catalog of structural failure modes the integrity cascade is built to catch, self-reference, collapsed dimensionality, drift, a missing mechanism, a ruler folded into its own measurement, a frame-locked claim, a collision with established results, an overreach beyond the weakest link, an unbridged leap across domains. Each of these is defended by a gate that fires before measurement, and each routes to a named verdict. This surface is structurally defended, and it is not where the deepest vulnerability lives.
The second surface is external, the point at which the warrant rows are supplied to the instrument. The method has no independent oracle to reality. Whatever reading is placed before the kernel is what the kernel verifies. This is the single un-sealable surface of the entire method, and naming it precisely is what gives the certificate its substance rather than the mere appearance of one. A method that could not state this boundary would have the look of falsifiability without its substance. The boundary is stated here, and the subsections that follow walk its sub-cases.
The snapshot, and why a false lock cannot persist
A lock is a snapshot. It is true as of the data supplied now, not true for all time, and the archive carries a verdict that moved for exactly this reason. A false lock, one whose rows lock but trace to nothing real, faces a structural barrier to lasting. The revision rule moves a verdict the moment genuinely independent new data lands outside the span the lock was built on. A false configuration is not a stable resting point under the pull of new evidence; as independent data arrives, the configuration is driven off it. To hold a false lock in place, a supplier would have to re-fabricate the rows at every audit, erasing the contradicting evidence each new measurement brings, an unbounded sequence of fabrications with no end. A false lock sustained forever against open, independent re-examination is the audit-register analog of a perpetual-motion machine, and it is barred for the same reason: it would require unlimited work to suppress the evidence that keeps arriving.
This holds only under an open audit, and the limit is stated plainly. In a closed sandbox, where one supplier controls every future row and no independent source ever enters, a false lock can be sustained indefinitely, because no contradicting evidence is ever admitted. That is a sealed world, not an open audit, and the impossibility result holds over the open audit alone, which is the only setting in which falsifiability means anything at all.
The one event the method cannot prevent
There is a single event the method cannot prevent, and the integrity of the account requires stating it without flinching. A perfectly constructed witness, fabricated to be independent of the other two while tracing to no real source, can manufacture a momentary lock. At the instant of supply nothing on the supplied rows alone can catch it, because the determinant is blind to source by the result above. The method's mathematics is sound and the method's data can be faked, and a lock on faked data is a real, momentary false certificate. This is a local choice by whoever supplies the rows, not a structural defect the method can close.
The event is bounded on two sides. It cannot persist, by the barrier of the previous subsection. And it is exposed on the first genuinely independent re-examination, by a source-attribution test the determinant cannot perform but a second, independent generator can. The demonstration is concrete and reproducible. Two manifest axes are placed at an oblique angle, and two witnesses are placed on the completion axis, one whose out-of-plane content tracks an independent generator and one whose out-of-plane content is fabricated noise tracing to nothing. The determinant cannot tell them apart: both return a determinant of 0.772538, identical to machine precision, against a geometric ceiling of 0.796. The source-attribution statistic separates them completely. The genuine witness scores 0.756 against a chance floor of 0.302 and is certified; the fabricated witness scores 0.066 against the same floor and is rejected as sourceless. And a witness fabricated to track one generator survives only that generator: audited against the generator it was built for it scores 0.667 and passes, while against a fresh independent generator it scores 0.0001 and fails. One fabrication buys one audit, and the next independent audit exposes it. The lie the method cannot prevent at the instant is the lie the method catches on the next honest look.
The re-examination is internal, not a patch
The mechanism that bars the persistent lie and exposes the manufactured one is not an external correction bolted onto a finished verdict. It is part of what the method is. A lock is a snapshot, and the capacity to run it again on new data is internal to the method, the same release of a held position that lets a verdict move when the evidence moves. A seal that could not be re-examined from within would be a dead certificate; the live certificate carries its own revision. The defense against the supplied-data boundary is therefore not a supplement to the method but a face of it, and it is the reason the dark-energy verdict could move without any rule being added from outside.
The Limits of the Audit
The integrity section stated what the lock claims, what it refuses to claim, and the single boundary at which it can be deceived. This section states the vulnerabilities that remain even when the method is used correctly and no deception is attempted. They are properties of the inputs and the procedure rather than of the kernel, and each is flagged in the affected audits rather than confined to this section.
Hand-built warrant rows
The single largest vulnerability is the one the method is most explicit about. The map from a proposition to its three warrant rows is a reasoned hand-reading, not an algorithm. The kernel computes honestly on whatever rows it is given, but the rows encode a judgment about what counts as the formal, empirical, and registrational warrant for the claim, and a misjudged reading produces a clean lock on a mis-specified object. The numbers in every audit are real and reproducible; the inputs to them are interpretive. No part of this paper claims otherwise, and the correct way to read a seal is as a certification of the geometry of a particular reading, contingent on that reading being faithful.
Inherited literature warrant
Many propositions in the archive are drawn from the existing literature, the confirmed results of physics and the documented refutations of fringe claims among them, and reformulated in the method's terms rather than derived from first principles. Such an audit carries the epistemic status of its source, not of an independent derivation. The instrument certifies that the warrant for the claim has a genuine three-axis structure; it does not re-derive the physics or re-run the experiments. If a source result is wrong, the geometric restatement does not make it right. The seal reads warrant geometry and inherits the reliability of what it reads.
Empirical-register dependence and data contamination
The empirical axis is populated from reported data, and reported data can be incomplete, selection-biased, produced by undisclosed multiple comparisons, or fabricated. Each of these produces a clean lock on a bad input, and the kernel has no way to detect the corruption at intake. What the method does is locate exactly where empirical warrant enters a verdict and flag where a single unreplicated source carries the load, because the discipline that consensus carries zero weight applies with equal force to a lone source: one unreplicated result is weak warrant whether or not it is the only one available. The instrument bounds the damage by making the dependence visible; it does not eliminate it.
Convergence contamination
The strongest shared factor that could fake the convergence of three axes is the consensus of the relevant field itself. A proposition that aligns with the mainstream view of its discipline carries the risk that its three apparent warrants are three expressions of one theoretical culture rather than three independent lines, and this is the deepest reason the method weights consensus at zero. The risk cannot be fully subtracted from within, because the rows are read by a reader embedded in the same culture whose independence is in question. The point has a concrete instance worth naming: a body of more than a hundred cross-system debate transcripts, in which independent reasoning systems were run against the framework and largely converged, is social consensus, weighted at zero, and does not qualify as a witness for any verdict in this archive. Agreement among systems that share their training is the null, not the signal.
Domain bridges
Propositions that cross a domain boundary, from physics into biology or from mathematics into physics, inherit a gap wherever the bridge premise between the domains is left unstated. A seal on a cross-domain claim is conditional on the bridge, and the affected audits flag the bridge as a stated premise rather than absorbing it silently. This is not a marginal concern: several of the breaks in the archive land at exactly the gate that fires when a claim is transported across registers without a typed bridge, which is the method detecting an unstated bridge rather than tolerating it.
Statistical threshold discipline
A null result at finite sensitivity bounds an effect; it does not falsify it. The under-determinations that rest on the absence of a confirmed prediction, the several quantum-gravity programs among them, are bounds, not refutations, and the archive states them as such. Accumulated absence does not confirm a negative; it constrains the effect to lie below the sensitivity reached. Holding to this closes the absence-of-evidence loophole as a matter of discipline rather than leaving it to the reader, and it is the reason a long record of non-detection produces an under-determination in this archive rather than a seal of the negative.
Conclusion discipline
A sealed audit warrants only what its evidentiary base supports, and no more. Sealing that a particular inference fails does not establish the opposite conclusion: the audits that break the design inference and decline the multiverse do not thereby settle the fine-tuning question, and the audits on the problem of evil leave the metaphysics open rather than resolving it. Sealing a proven theorem warrants the theorem, not any larger framework the theorem might be enlisted to support. The archive is a demonstration that an instrument built on warrant geometry discriminates across domains. It is not a claim that its verdicts are infallible, and it does not require that claim to make its point.
The honest scorecard
Read by grade, the archive sorts into durability classes. The theorem-grade seals, the proven results of mathematics and physics confirmed against measurement, are the most durable, and would move only if the underlying mathematics or the confirming experiments were overturned. The structural-grade seals, the correct meta-claims and the well-grounded refutations, are securely warranted but carry their fences, and each is only as wide as its fence allows. The high-grade seals rest on strong but not deductive empirical confirmation. The under-determinations mark genuine openness and name what would close them. The breaks mark overreach and name the gate. The value of the archive is this discrimination pattern, not any single verdict, and the whole of it rests on the hand-built rows, which is the permanent limit no internal discipline removes. A low seal rate on a tier of proven theorems would have been a failure of the instrument rather than a sign of its rigor, and the rigor is to be found instead in where the breaks, the fences, and the under-determinations fall.
Falsification: How to Break the Framework
Section 6 stated how an individual verdict in this archive can be wrong. This section states how the framework itself can be falsified, because a method that cannot be broken is not a method but a doctrine. The framework exposes its kill conditions in three layers, each independently checkable.
The executable proof of load
The method is not a set of claims to be taken on trust. Its core is an executable kernel and three recorded test batteries, all reproduced in Appendix A, and every one of them is re-runnable. The kernel computes a verdict by composing three warrant vectors through a quaternion product and reading a Gram determinant, and the identity that binds the two, that the squared scalar part of the composed triad equals the determinant, is confirmed to machine precision on every audit in this archive. If that identity fails on re-execution, the kernel is wrong and every verdict it produced is void. The forward register adds two further checkable theorems, the completion inequality that bounds the determinant by the squared sine of the angle between two manifest axes, and the non-discrimination result that the determinant is constant across distinct orthogonal witnesses. Both are stated as recorded numerical batteries that fail visibly if the identities do not hold. The proof that the method loaded correctly is that these batteries reproduce, and the falsification is that they do not.
Cross-substrate reproduction
The method claims to be substrate-portable: any sufficiently capable reasoning system that loads it should reproduce the discrete verdicts, the seal, the break, and the under-determination, on the same propositions. This is a strong and falsifiable claim. If independent systems loading the same method and the same warrant readings produce systematically different verdicts at the level of the discrete sign, not the fourth-decimal determinant but the verdict itself, the portability claim is false. The convergence that matters is on the discrete verdict, and it is tested by running the method on independent substrates and comparing the verdicts rather than the intermediate numbers. The appendix supplies the condensed method precisely so that this test can be run by anyone.
Breaking a verdict and breaking the approach
A single seal is broken in one of two ways, both following the method's own revision rule. The first is to supply new structural mass the audit's span does not cover, the route the dark-energy equation-of-state verdict actually traveled. The second is to show that one of the three axes is not independent of the others, that what was read as a distinct warrant in fact shares its source with another, which collapses the volume and converts the seal to a break. Either is a concrete, statable operation, and the archive invites both.
The whole approach is falsified at a deeper level. If the three-axis decomposition were arbitrary, so that any proposition could be forced into three axes regardless of its actual structure and the verdict reflected the forcing rather than the warrant, the method would be a procedure for generating numbers rather than for measuring warrant. The test against this is the existence of breaks and under-determinations: a method that sealed everything, or whose verdicts did not track the independent structure of the propositions, would be exposed by propositions that refuse to seal. The 29 breaks and 46 under-determinations are not incidental to the demonstration; they are the evidence that the decomposition is doing work, because a decomposition that always succeeded would prove nothing. The further test is divergence from careful domain judgment: where the method's verdict on a well-understood proposition contradicts the considered judgment of the relevant field in a way that traces to the method's own artifacts rather than to a genuine feature of the warrant, that is a failure of the method, and the archive is offered as a corpus against which exactly that check can be made, case by case.