published, archived,
and citable
Every claim on this site traces to an archived record with a DOI. Where we have no archived result, we publish no number.
Holosynthics: Composition Laws for Signed Digital Entities
The composition theory the whole product line rests on — the laws that govern how signed digital entities combine.
Verifiable Process Infrastructure: A Formal Framework for Cryptographically Provable Business Processes
Formal definition of Verifiable Process Infrastructure: four axioms, three theorems establishing soundness, necessity and sufficiency, and eleven regulatory regimes — including SOX, DORA, SEC 17a-4, MiFID II, GDPR and BCBS 239 — shown to require properties no existing tool provides. The GI Engine is its reference implementation.
Empirical Validation of the Holosynthics Periodic Table: Cross-Model Universality, Hallucination Detection, and Surgical Repair Across Transformer Architectures
From the archived preprint (DOI 10.5281/zenodo.20736159): The paper reports 3,000 sequential edits, ×9.08 surgical repair on Llama 3.1 8B, and validation across 11 transformer models from 9 organizations.
BIO56: Deterministic Decomposition of Biological Data into 56 Elements — Experimental Results
941 biological entities across 4 independent data domains decomposed into 54,986,846 atoms by a deterministic 7-law classification cascade.
ROBO56: Universal Robotic Sensor Decomposition — 10 Sensors, 2 Platforms, 74.4M Atoms
Robotic sensor data from an autonomous vehicle and a drone decomposed into 74,373,570 atoms by the same engine proven on biological data.
CounterFact × 4 Methods — Two Models, One Harness
ROME, MEMIT and AlphaEdit measured against Probe56 on the same 1,000 CounterFact edits, one harness, two unrelated architectures — Llama-3.1-8B and GPT-J-6B. Every per-row result, five reproducibility packages per model, and a signed certificate for each.
Does AlphaEdit Really Preserve Knowledge? Testing the ICLR 2025 Outstanding Paper's Guarantee
AlphaEdit's null-space projection is a real advance, and it does what the paper proves it does. It does not prevent collateral damage — we measured 202 of 492 nearby facts broken on Llama-3.1-8B and 143 of 499 on GPT-J-6B.
Read it → 6 min
ROME vs MEMIT vs AlphaEdit: A Reproducible Collateral-Damage Benchmark (Llama-3.1-8B & GPT-J-6B)
Three knowledge-editing methods, 1,000 edits, two unrelated architectures, one scoring harness. All three damaged facts the model already answered correctly — and the damage rate does not transfer between models.
Read it → 9 min
three production models, certified
Every scan runs against a world-standard benchmark suite spanning 49 categories — covering knowledge, reasoning, math, logic, reading, hallucination, safety, and commonsense, plus domain-specific categories. Full case studies available on request.
two unrelated architectures, zero collateral
Head-to-head against ROME, MEMIT and AlphaEdit under one scoring harness. Every method broke nearby facts on both models except one. Signed and checkable by anyone: github.com/growing-intelligence/counterfact-comparison
signed, and checkable by anyone
Every result referenced on this site is signed by the GI Engine, so an auditor can verify that a finding came from us and was not altered after the fact. Classification, detection, and repair methods. Protected by patents filed with USPTO. Application numbers are available on request under NDA.