LAB Technical work

Evidence lives
one level deeper.

Selected demonstrations, measurements, and methodology from AZIRE AI’s active development. Detailed enough to inspect; bounded enough not to overclaim.

01 Governed cycle

Follow one harmless request through the authority boundary.

This browser-only demonstration is illustrative. It does not contact Castor, issue real authority, run a model, or modify a repository.

SIMULATED TASK ENVELOPE

“Update a README heading.”

OperationPropose edit
Write scopeREADME.md only
Apply authorityNot granted
Grant behaviorSingle use
01Conductor plansInspect

The request is classified and routed into a proposal-only workflow. Planning does not itself create permission.

workflow: smith_proposal_only
02Authority narrowsInspect

Registry limits, task authorization, workflow constraints, and environment policy intersect. Authority can only become narrower.

effective = ceiling ∩ task ∩ workflow ∩ environment
03HEKATE grantsInspect

A short-lived opaque grant binds the exact context and step. Redemption is checked before repository or model work begins.

status: active → consumed
04Smith proposesInspect

Smith creates a patch and evidence package. The proposal may be reviewed, but it cannot silently become an applied change.

result: proposal_created
05Human reviewsInspect

Forge presents the bounded proposal and its evidence. Apply, commit, and push remain separate decisions.

apply_authority: false
OutcomeReviewable proposal created

No repository change. No implicit apply. Replay rejected.

02 Model Lab evidence

Models earn roles
through tests.

A bounded continuity bakeoff ran every candidate twice on the same Castor case at the same 64-token limit.

CandidateRecallSummaryAvg. latencyResult
Qwen3 4B Non-Thinking2 / 22 / 27.563sPass
Qwen2.5-Coder 7B2 / 22 / 28.028sPass
Phi-3 Mini2 / 20 / 212.834sPartial
Gemma E4B0 / 20 / 29.720sFail
Qwen3 4B Thinking0 / 20 / 210.160sFail
Method note

These results answer a narrow question: whether each model could satisfy this specific continuity case under identical constraints. They are not claims of general model superiority.

2 runs each · identical prompt · committed 64-token limit