Diagnostic research for reliable AI systems
Find the failure before you trust the result. Velorin tests the points where AI systems quietly become unreliable: memory that survives correction, evaluations that reward the wrong behavior, verification that misses incomplete work, and evidence that breaks before a result can be trusted.
Diagnose a failure See the decisions Follow the field
Published mathematical result
Independent verification completed · 17 November 2026
Velorin Intelligence Establishes Improved Coding-Theory Bound E 3 (12,8) ≥ 12
An AI-assisted computational search produced a 12-word ternary edit code improving the published lower bound for E 3 (12,8). Independently verified by Abraham Yeung, the result is now included in his maintained bounds table and preserved in a permanent research archive.
01 / Current diagnostic · Execution evidence
The run reports success while the process is still running. If accepted, an incomplete experiment can enter the record as a finished result.
Consequence The execution boundary was never established, so the result cannot yet be trusted.
Decision Block scored trials until termination evidence reproduces independently. Inspect the qualification record
Each record begins with an observable condition, then identifies the risk, the smallest useful test, and the resulting decision.
Persistent memory
Your agent keeps acting on information you already corrected. Stale, withdrawn, or unauthorized memory can steer later decisions long after the original source has changed.
Risk if trusted The system may repeat a decision that no longer has valid authority or evidence.
Diagnostic Trace every retained claim to its source, permission, correction chain, and revocation state.
Decision Build source-backed, correctable memory. Retention alone is not progress. Inspect the supporting evidence
Verification
Every check passed. The requested work is still incomplete. A compliant process can still produce an unusable or unfinished result.
Risk if trusted A green verification report can conceal failure against the actual acceptance criteria.
Diagnostic Compare the evidence produced by the verifier with the outcome the task required.
Decision Constraint Sentinel did not add enough value over simpler controls. Product development stopped. Read what failed
Evaluation
You may be maintaining an evaluator that never changes a decision. Extra evaluation machinery adds cost and false confidence when a simpler established framework reaches the same conclusion.
Risk if trusted Complexity is mistaken for stronger evidence.
Diagnostic Run the same frozen cases through the custom layer and the simplest credible baseline.
Decision Keep only a reproducibly distinct contribution—or retire the custom runner. Inspect the decision test
Evidence lineage
The conclusion looks supported, but part of the evidence chain is missing. A plausible claim can survive review even when one of its supporting steps cannot be traced.
Risk if trusted An unsupported claim is released because local checks never exposed the missing lineage.
Diagnostic Require each released claim to close against inspectable source and transformation records.
Decision Keep the bounded positive result local and preserve the external negative result. Inspect the evidence record
02 / Current work
What the evidence must decide next 01 Can persistent memory remain useful after correction or withdrawal? The next usable proof is not recall. It is an approved fact recalled with its source, revoked, and then absent from later recall.
Current direction 02 Does VCE reveal anything a simpler evaluator does not? If the answer is no, the custom runner goes away. The frozen cases and decision rules remain.
Decision test 03 Can discovery evidence be trusted after the search ends? SDPS cannot begin a scored search until the execution boundary and terminal evidence survive independent review.
Qualification record
01 Name the symptom. State the condition someone can observe without requiring the project taxonomy.
02 Show the consequence. Identify what could be falsely trusted, released, or deployed.
03 Run the smallest decisive test. Compare against the simplest credible alternative and preserve the evidence.
04 Continue, repair, block, or stop. The result must produce a bounded engineering or portfolio decision.
04 / Inspect the record
Use the diagnosis first. Audit the technical evidence underneath it.