Working demo · intelligence density harness

Build a verifier-gated Bonsai pack

Describe the workflow, watch the pack draft beside you, approve risky writes, and keep claims inside the proof envelope before anything leaves the boundary.

1 Source Sensitive note stays local. Clean triage record is the only egress target.
2 Recognize Bonsai proposes semantic identifiers, humans accept the recognizer set.
3 Verify The gate checks the actual outbound payload, not just the model output.
4 Structure Slack sees themes, next step, and risk flags. It does not see raw note text.
5 Publish Claims cite synthetic evals, gate proof, and runtime limits.

Wireframes 1A · 1B · 1C

Workflow interview to draft pack

Builder

I run care-navigation intake. We need to capture synthetic referral notes, remove identifiers, and send only clean triage themes to the team.

DeepAgent

I hear three signals: member identifiers, contact details, and family details. I will draft recognizers, a clean record schema, default-deny egress, synthetic eval rows, and safe claims.

Builder

The value story is internal adoption. Show leadership that the sensitive note stays local while the clean record reaches Slack.

DeepAgent

Pack draft ready in scratch space. It needs human approval before writing files or preparing an upstream benchmark report.

Draft assembles alongside the interview

The agent converts workflow pain into recognizers, policy, schema, eval rows, safe claims, and a benchmark report.

Evidence before claims

Verifier and benchmark state

pass

Synthetic data only

3 eval rows, 0 real notes, 0 raw examples in publish copy.

pass

Residual identifier scan

0 residual identifiers in clean Slack payload.

pass

Default-deny egress

Slack sink requires verifier.residual_identifiers == 0.

needs-review

Claim critic

Unsafe terms are blocked until copy says scrubbed or redacted.

pass

Prism benchmark export

Report includes hardware, model family, eval size, limits, and gate result.