Bonsai Pack Builder

A Deep Agents and SvelteKit seed repo for turning regulated workflow pain into verifier-gated Bonsai packs.

Live demo: chaiwithjai.github.io/bonsai-pack-builder

This is the next ecosystem artifact after the iPhone PHI scrubber demo. The prior demo proved a narrow workflow. A user can capture a synthetic note, scrub identifiers inside a first-party edge boundary, run a verifier gate, and send only a clean structured record to Slack. This repo turns that lesson into a pack authoring harness. The harness helps a user move through this sequence:

• interview the workflow;
• draft the pack;
• create synthetic evals;
• run gates;
• collect evidence;
• prepare a contribution.

The core claim is intentionally modest: density models such as Bonsai become useful when they sit inside harnesses with explicit policy, evals, human approval, and default-deny egress. They are components, not oracles.

What is here

• Working value demo screenshot - browser proof for the sensitive-note to clean-artifact path.
• Working pack builder screenshot - proof for the interactive pack inspection path.
• Advisory brief - the short review note for PrismML founders, advisors, and ecosystem reviewers.
• PRD - product scope, users, MVP, acceptance criteria.
• RFC-0001 - proposed architecture and delivery plan.
• Demo runbook - how to review the seed and what the full demo should show.
• Verified seed UI screenshot.
• Implementation plan - the phased path from seed repo to working pack builder.
• Completion audit - requirement-by-requirement proof against the original goal.
• Failure analysis and demo acceptance.
• FSM screen map - route, screen, component, and gate map from the wireframes.
• UI state model - address, local, draft, server-cache, and derived state.
• System architecture - Deep Agents, SvelteKit, Bonsai adapter, verifier, and exporter.
• Network topology and flows.
• Service map.
• Means, Motivation, Opportunity.
• PrismML proof timeline - citation-backed founder, partner, and course-delivery proof points.
• PrismML benchmark template.
• Care navigation sample benchmark report.
• Care navigation example.
• Care navigation synthetic pack.
• Wireframes.
• Source index.

Run the demo

npm install
npm run dev

Then open the printed local URL. The UI is a working static demo. It lets you:

• run the workflow interview and assemble a draft pack;
• inspect the IDE-style file diff;
• walk the guided wizard;
• approve the write;
• inspect the registry, evidence table, and fork path;
• run claim lint and switch to safe copy;
• review the CLI equivalent of the same harness.

The visible state panel shows address, local, draft, server-cache, and derived state while you click.

Verify

npm run check
npm run build
npm run docs:check
npm run pack:check

The public static demo is built with:

npm run build:pages

Guardrails

• Synthetic examples only.
• Say scrubbed or redacted; do not claim legal de-identification.
• Do not claim HIPAA compliance.
• Bonsai is a local inference component behind policy and verifier gates.
• Human approval is required before writing generated packs, running sinks, or publishing external claims.

Source context

This repo preserves source context from the prior build under docs/research/ and docs/strategy/. The public contribution target is the benchmark-template pattern in PrismML-Eng/Bonsai-demo, especially the community-benchmarks/bonsai report templates.