The math is public. You can check it.
Security claims are worth exactly what the evidence behind them is worth. So here is the evidence — the physics, the arithmetic, and our measured test results, with sources you can verify independently.
Every constant on this page comes from published physics. Every test number comes from our own suites. What stays private is the architecture that makes the arithmetic apply — not the arithmetic itself.
The energy floor
Erasing information costs energy. That is not an engineering limit — it is thermodynamics, and it puts a hard floor under any exhaustive search.
The Landauer limit gives the minimum energy to irreversibly erase a single bit at temperature T:
Multiply that floor by the number of operations an exhaustive search requires. The results are not close calls:
| Search width | Operations | Minimum energy | Comparison |
|---|---|---|---|
| 2128 | 3.4 × 1038 | ~9.8 × 1017 J | A few hours of global energy production. Feasible in principle. |
| 2256 | 1.2 × 1077 | ~3.3 × 1056 J | ~1012 times the Sun’s entire lifetime output. Not reachable with stellar energy. |
| 2512 | 1.3 × 10154 | ~3.9 × 10133 J | ~1063 times the total mass-energy of the observable universe. Categorically excluded. |
Aegis Eternal operates above the second line. The exact width is reserved — but the threshold behavior is what matters, and it is public arithmetic you can reproduce from the constants above.
What quantum actually changes
Quantum computing is treated as a wildcard in most security marketing. It isn't. Its effects are specific, published, and bounded.
Shor's algorithm
Factors integers and solves discrete logarithms in polynomial time. This breaks RSA, Diffie–Hellman, and elliptic-curve cryptography — the asymmetric primitives that secure most of the internet.
It does not apply to symmetric primitives, and it does not apply to unstructured search. It is not a general-purpose codebreaker.
Grover's algorithm
Gives a quadratic speedup on unstructured search — √N instead of N. In practice this halves the effective exponent of a search space.
A search space built with that halving already accounted for lands back on the same energy wall. Grover changes the arithmetic. It does not change the conclusion.
Post-quantum posture is aligned to the NIST standards track — FIPS 203 (ML-KEM), FIPS 204 (ML-DSA), and FIPS 205 (SLH-DSA), standardized August 2024. Verifiable, unglamorous, and the correct answer to “what about quantum.”
The problem nobody budgets for: where does it go?
Conventional brute force has a compact failure. You test a key, receive one bit of answer, discard it. Storage cost per attempt is effectively zero. Every keyspace-security estimate ever published rests on that assumption.
Against this architecture, that assumption does not hold. There is no rejection signal — no one-bit “no.” Every attempt returns a complete, well-formed file that has to be transported across a network, held somewhere, and inspected before the next attempt means anything.
The failure is itself a full artifact. That converts the attack from a compute problem into a logistics problem — and logistics do not yield to Moore’s Law.
There is no medium to write them to, no network that could carry them, and no way to defer evaluation to reduce the load. You can add cores. You cannot add universe.
Why no shortcut exists
Every attack technique that beats brute force depends on a gradient — some signal that says you're getting warmer. Hill climbing, evolutionary search, differential cryptanalysis, partial-key recovery, side-channel correlation. All of them convert a search problem into an optimization problem.
Here, proximity to the correct key confers no advantage whatsoever. A key that is one element wrong produces a result indistinguishable from one that is entirely wrong. There is no gradient, no partial credit, and no warmer-or-colder signal.
No search strategy outperforms exhaustive enumeration — and no optimization technique, including machine-learning approaches, has anything to optimize against. Machine learning is descent over a loss landscape. On a flat landscape there is no descent. The technique is not slow here; it is inapplicable.
That is the honest form of the “AI-resistant” claim. An attacker using a model to triage results still runs inference on every candidate — which is more expensive than the operation it replaces, not less — and a fraction of results pass superficial validity checks by chance, which corrupts the filter the same way it corrupts a regex.
Four independent walls
Flatness forces enumeration. Enumeration then fails four separate ways, each of which is sufficient on its own.
Assembly
Stored fragments carry no linkage information. An attacker holding every fragment cannot determine which constitute a file, or in what order, without already possessing the key. The search space is not the keyspace — it is the keyspace multiplied by an assembly space that isn't enumerable from what's on disk.
Oracle
No rejection path exists. No error, no lockout counter, no timing signature separating accept from reject. Every submission produces a complete result, and evaluation cannot be safely delegated to an automated filter.
Economics
Every attempt requires full retrieval and reconstruction. There is no fast-reject path — no cheap test that lets an attacker sweep at hardware speed. The per-attempt cost is a full cycle, not a comparison.
Logistics
Every failure is a complete artifact requiring transport, storage, and inspection. The attacker cannot batch and defer. Physical handling becomes binding long before the cryptography is meaningfully tested.
This survives a cipher break
The sharpest question a reviewer can ask is: what happens when someone publishes an attack on your underlying cipher?
For nearly every security product on the market, the answer is “everything fails.” A single cryptographic assumption carries the entire system.
Legs I and II above are not cryptographic assumptions. Break the cipher tomorrow and an attacker still faces: which fragments belong together, in what order, and how would I know if I had succeeded? None of those questions is answered by breaking a cipher.
We also satisfy Kerckhoffs’s principle — an attacker who knows exactly how the system works still cannot proceed without the key. That is not an assertion: it is a tested property, with a dedicated attacker-knows-scheme suite in the core library that passes on every build.
What we actually measured
Architecture arguments are only as good as the implementation behind them. These are our own suites, run on every build.
The attacker-knows-scheme, cascading-errors, and independence suites are the three that matter most for the claims on this page. They test the failure modes rather than the happy path.
Verify it independently
You should not take our word for any of this. Here is what you can check without us.
- 1Reproduce the energy arithmetic
Take the Boltzmann constant from the SI definition, compute kB·T·ln(2) at 300 K, and multiply by the operation counts in section 01. Every figure in that table follows from published constants.
- 2Check the quantum claims against the literature
Grover’s quadratic bound and Shor’s polynomial-time factoring are both published and unambiguous. Confirm that neither changes the conclusion at the widths in section 01.
- 3Confirm the post-quantum standards
FIPS 203, 204, and 205 are published by NIST. Our alignment to that track is verifiable against the public standards.
- 4Bring a red team
Under NDA, we will host an adversarial evaluation against a live deployment — with the architecture disclosed to your reviewers, so the test runs under Kerckhoffs conditions rather than against obscurity.
What’s on this page, and what isn’t.
- ·Published physical constants and the arithmetic from them
- ·Published quantum-algorithm bounds
- ·NIST post-quantum standards alignment
- ·Our measured test results and pass rates
- ·Security properties stated as outcomes
- ·Search-space width and construction
- ·Fragment topology and distribution
- ·How assembly information is withheld
- ·Result-generation methodology
- ·Patent claim subject matter (STIMS priority chain)
The distinction is deliberate. Everything needed to evaluate the claims is public. Everything needed to rebuild them is not.
See it running.
The demos run the same pipeline as production, against synthetic data. Nothing you click touches a real record.