Growing Intelligence
Probe56

the AI model health
platform

Scan any open-weight model across 56 functional elements and 5 dimensions, repair exactly what fails with zero measured collateral, and certify the result with an Ed25519 signature anyone can verify.

Contact us
How it works for you

five stages, one signed record

Scan
A full diagnostic pass over an open-weight model across 56 functional elements and 5 dimensions, ranked against the complete vocabulary rather than a shortlist. The output is a map of exactly where the model fails and where it is already correct — the second half matters as much as the first, because it defines what a repair must not disturb.
Repair
Targeted parameter-level repair applied to the failures the scan found. Every edit is verified against the full control set and ships with a measured collateral report: sample size, control set, and a 95% confidence bound. If any previously-correct answer degrades, the edit is rejected automatically.
Monitor
Gyro watches every inference at the neuron level and detects drift before it reaches production, with a confidence score on every answer and calibration per deployment. Neuron-level, not output-level: the difference is whether you learn a model has drifted from its own internals or from a customer complaint.
Protect
Guardian blocks prompt injections, jailbreaks, and unsafe inputs in under 5ms with zero GPU. A standalone safety layer, always on, running a deterministic ruleset with zero false positives — the same input always produces the same verdict, which is what makes it auditable.
Certify
Every result is Ed25519-signed and independently verifiable, producing a certificate that states model health before and after the repair. A signature is not a claim about our method; it is a property anyone can check, which is what an auditor can actually act on.
The evidence

Three production models certified with zero measured collateral damage. Every result is Ed25519-signed and independently verifiable.

3
Certified models
32% → 70%
Accuracy
0
Collateral damage
Ed25519
Signed proof

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.

Tested head-to-head against three leading weight-editing methods on two architectures — Llama-3.1-8B and GPT-J-6B. Zero collateral damage on both; the only method that achieved it.

2
Architectures tested
4
Methods compared
0
Collateral, both models
Ed25519
Signed comparison

Full comparison: the research page · github.com/growing-intelligence/counterfact-comparison

Probe56 FAQ

questions

What does Probe56 actually change in the model?
Repair applies a targeted, minimal fix. It does not retrain the model and it does not inject new facts into it — which is what weight-injection editors do, and what the forgetting literature reports going wrong at scale. Scan changes nothing at all: it is 100% read-only.
How is collateral damage measured?
Against a held-out control set of items the model previously answered correctly, evaluated under a no-teacher-forcing protocol. Each run reports the sample size, the control size, and a 95% confidence bound on the true collateral rate rather than a bare pass/fail.
How is this different from model observability?
Monitoring platforms such as Arize, Fiddler and WhyLabs detect drift and surface that a model is misbehaving. They do not repair it. Probe56 is a repair step: it acts on the specific wrong answer and then proves, against a control set, that the rest of the model's behavior held.
Can we run it on-premise?
Yes. Probe56 runs on-premise or in your VPC against open-weight models. Results are cryptographically signed and reproducible, which is what auditors in HIPAA- and GDPR-grade environments ask for.