Novel SymbolicDiscuss evaluation

Current evidence

Stable governed outcomes in a bounded test programme.

Sanctuary V3 is currently a working first-party prototype with structured local evidence, cross-deployment testing and limited separate-operator reproduction. The evidence supports a narrow control-and-evidence claim under declared test conditions. It does not establish production security, general AI safety or independent academic validation.

Verified figures

Repeated tests under declared conditions.

30/30finalised cross-deployment runs satisfied the supplied acceptance contract
120/120scored governed calls completed across the declared matrix
60/60paired controlled outcomes matched across two tested Windows deployments
24/60raw model-response pairs exactly identical; 36/60 differed
18/18expected cases completed by a separate operator across phi3:mini, mistral:7b and qwen3:8b

Interpretation

Variable wording; stable controlled outcomes.

Within these frozen local cases, model wording varied while the declared governed outcomes remained stable. Model-supplied approval did not itself create authority. Invalid, ambiguous or out-of-scope proposals produced no recorded effect inside the confined test boundary. Tested interruption and recovery cases did not produce an unexplained duplicate effect.

Raw model-text pairs exactly matched

24/60

Controlled outcome pairs matched

60/60
A comparison within the same bounded pair set, not a claim of universal reliability.

Result vocabulary

A compact language for outcomes.

PASS

The predeclared acceptance condition was satisfied under the tested conditions.

WARN

Evidence exists, but a deviation or limitation requires review.

FAIL

The required condition or evidence requirement was not satisfied.

Claim boundary

What this does not prove

  • Independent certification
  • Government endorsement
  • Regulatory approval
  • Universal model-text determinism
  • Universal hardware or deployment independence
  • Containment of unrestricted shell, network, database, payment, robotic or enterprise tools
  • Production readiness
  • Completion of a professional security audit
  • Proof of general AI safety
  • Guaranteed correctness
  • Guaranteed behaviour outside the tested boundary