Bid Engine

Screen 05 · Drafting Agent · fn_draft_response

Draft canvas

R-001 · Secure model deployment within a classified enclave. The agent may only cite evidence the retrieval policy returned.

Response draft

1,240 / 1,500 words · draft.v5

Neptune North deploys models into classified enclaves using an air-gapped build pipeline: artefacts are signed at build time, transferred through a one-way guard and verified against a hash manifest inside the enclave before any inference runs.

[E1]The same pattern is in production at a UK systemically important bank, covering 340 models across three security domains.

Model cards travel with the artefact. Each card records intended use, evaluation data, known limitations and the named accountable owner[E3] and the enclave refuses to load an artefact whose card is absent or unsigned.

Counter-view: an evaluator may ask about model update latency in a disconnected environment. Answer drafted in the appendix; scheduled for Gold review.

Retrieved evidence

Layer 06 policy
  • E1Aurelia enclave delivery case studysigned
  • E2NCSC secure AI development guidancesigned
  • E3Model card template v4signed
  • E4Internal enclave wiki pageunsigned

Model card

Layer 04
Model
claude-sonnet · drafting
Prompt
draft.response.v9
Temperature
0.3
Redaction
applied pre-call