The Graveyard

Every belief in this
system earned
its place.

Here is the permanent log of everything it refused to believe — and the measured reason for each refusal. No language model participates in any verdict. 141 claims tested. 141 retained. Zero deleted.

AI systems that build knowledge bases have a structural problem. The model that proposes a claim also writes the justification for it. The justification is never independent of the claim.

At scale this produces knowledge bases that are fluent, richly cited, and largely unverifiable. Every claim carries high stated confidence. None has survived a test that could have failed it.

Aby is different because it can say no. And it keeps the receipts.

Measured — 15 August 2026
693
Relations in KB
Every edge between concepts
141
Claims rejected
Retained permanently. Zero deleted.
4
Theorems found
From measurement. Without being told.
0
LLM verdicts
Every verdict is a measurement.

The live record

Live — Supabase
Pair Verdict Evidence Gate
FourierTransform ~ FourierSeries
FN1D · strongest in the tree
Identical 12.31 bits Surprisal gate
FourierTransform ~ HartleyTransform
FN1D · first gate-measured admission
Confirmed 5.54 bits Surprisal gate
GreenFunction ~ FundamentalSolution
PDE · same object, two names
Identical 5.43 bits Surprisal gate
CanonicalTransform ~ HamiltonianTransform
MECH · Arnold 1978
Identical 4.46 bits Surprisal gate
HubbardStratonovich ~ RenormalizationGroup
PDE · non-obvious kinship
Confirmed 4.26 bits Surprisal gate
EigenDecomposition ~ JordanNormalForm
MATRIX · structural kinship
Confirmed 6.70 bits Surprisal gate
LorentzTransform ~ GalileanTransform
GROUP · superficially similar, structurally distinct
Refuted 1.89 bits — below threshold Surprisal gate
AffineTransform ~ MobiusTransform
GROUP · different structure
Refuted 1.22 bits Surprisal gate
AbelTransform ~ HuffmanCoding
PROJ × CODE · cross-type claim rejected
Type error 0 bits · malformed Type gate
GapCausalTransform
Open gap · the gap-test mechanism
Gap closed Pearl 1995 already answers this Gap specification
What the system found without being told
12.31bits

FourierTransform is structurally identical to FourierSeries

Never proposed by the autonomous loop. Found by the engine running both transforms on real inputs and measuring seven properties. 12.31 bits — the strongest structural kinship in the knowledge base. The loop had been proposing weaker pairs for months.

4.46bits

Hamiltonian transforms are canonical — Arnold, 1978

The MECH probe battery measured seven properties for each transform and found identical signatures. The engine rediscovered this result from Arnold's Mathematical Methods of Classical Mechanics (1978) from numerical measurement alone. The result is now in the knowledge base, confirmed.

5.43bits

GreenFunction and FundamentalSolution are the same object

Physicists call it the Green's function. Analysts call it the fundamental solution. Identical measured property vectors. The ontology had them as two separate nodes — a data quality failure the engine caught and flagged as a duplicate.

4.26bits

Hubbard-Stratonovich and RenormalizationGroup share deep structure

A 1950s algebraic trick from statistical field theory and the scale-flow operator at the heart of modern quantum field theory. Not an obvious pair. Both are nonlinear field transforms, neither causal nor linear. Confirmed from measurement — 4.26 bits of shared surprisal.

The gap test — detecting already-solved problems
✓ Closed · 15 August 2026 · Gap-test mechanism

GapCausalTransform was not an open problem

The system had been spending part of every nightly run searching for a mathematical transform that handles causal interventions — non-linear on joint probability distributions, breaking independence structure when variables are intervened upon, reducing to ordinary conditioning when nothing is intervened.

Pearl's do-calculus (1995) satisfies every requirement. The mathematics has existed for thirty years. The gap was a missing ontology node, not an open problem. The system detected this, closed the gap, and redirected the freed budget to GapTimeTransform — the one gap with a genuine open specification.

This is what the gap-test mechanism was built to do. It worked on its first real application.

Tamper-evident audit record

Every verdict carries the SHA-256 hash of its own content plus the hash of its predecessor. Modify any row — the chain breaks at that point. An external auditor can verify the entire record independently. "Auditable" is now a mechanism, not a marketing claim.

d432144757bd04318bbc2f4a7121fd53eed89d84c43d9d525baf2b2425d7dfcdConfirmed
7f3a91cc28b4e56d0f9a2c18e77d4b3309aef1245cc8f761b3e9d20a4f8c1237Refuted
1b8e24d76c5f91a3b0289e4c47f63d58a12bc7fa901d3e5c82b41679d4f0e238Type error
a39f62e1d4780b5c2917f0e38d6a14b29c8e541f673b2d4910c8a37e56d1f490Identical

141 rows · 140 chained · trigger fires on every INSERT · first row has no predecessor (correct behaviour)

We are looking for a partner
with a genuine auditability
requirement.

Defence, safety-critical engineering, AI-for-science. The EU AI Act enforcement provisions took effect August 2026. The FY2026 NDAA requires DoD contractors to audit the AI they deploy. The requirement is no longer optional — and most AI systems cannot satisfy it.

We want to be the first organisation to deploy a verified autonomous knowledge system in a real operational context. The zero grounding ratio is not a weakness — it is why we need you.

What a partner gets

A system that builds knowledge honestly. An auditable record of every decision. The ability to say to their own auditors: "Here is the log of everything this system was told — and decided not to believe."

VyroNyx Private Limited

Bengaluru, India
nj@vyronyxgroup.com
hal.vyronyxgroup.com
Patent pending · Stamped 15 Aug 2026

Graveyard141 verdicts
Transforms typed64
Families measured7
Theorems found4
Gap closedGapCausal ✓
Hash chain140 linked
LLM verdicts0
IP stamped15 Aug 2026
Next run02:05 UTC
Graveyard141 verdicts
Transforms typed64
Families measured7
Theorems found4
Gap closedGapCausal ✓
Hash chain140 linked
LLM verdicts0
IP stamped15 Aug 2026
Next run02:05 UTC