# Build 1 Milestone 1 Kickoff

Date: 2026-06-02
Task: `task:fg-goals-experiment-build1-kickoff-20260602`
Run level: `preview_bound` for the original kickoff; content-only publication is now `execute_bound` for the already-provisioned app.

## Final Check Answer

1. There is no remaining blocker to starting milestone-1 work.
2. The Rust `publish-site` adapter plus release-gate route is sound, but it belongs to milestone 2. The smallest non-blocking first step is a release-plan and adapter contract skeleton only; no public deploy or execute-bound publish action is attempted here.
3. The frozen eval set and external-source manifest are good enough for an existence-first run. The S1-S7 manifest hashes were rechecked with `/usr/bin/openssl` before this kickoff. Eval perfection is not a blocker; v1 stays fixed for Build 1 and updates through a v2 recalibration after human usage evidence.

## What Was Built In This Kickoff

- `local-shell/index.html` - locally reviewable Varro V1 shell adapted from the cromer platform shape.
- `kg/kg-0-draft-records.ndjson` - KG-0 node records conforming to the crystallised KG contract; no `source_backed` promotion.
- `kg/kg-0-relationships.ndjson` - first-class KG relationship records.
- `kg/kg-0-source-trace.md` - human-readable extraction and trace notes.
- `kg/kg-1-draft-records.ndjson` - KG-1 draft rebuild, folding public deploy evidence into the graph.
- `kg/kg-1-relationships.ndjson` - first-class KG-1 relationship records, including `resolves` and `changed_by` links.
- `kg/kg-1-source-trace.md` - KG-1 rebuild input set, active gaps, and resolved-gap evidence.
- `local-shell/tutor-citation-checks.html` - cycle-6 static tutor citation-check surface generated to the KG-1 limit.
- `eval/varro-tutor-eval-v1-bridge-prompt.md` - bridge prompt for scoring the frozen v1 eval set without exposing expected answers.
- `eval/varro-tutor-eval-v1-response-schema.json` - JSON schema for eval-mode output shape.
- `eval/varro-tutor-eval-v1-structured-output.json` - structured eval output plus cost/usage metadata.
- `evidence/build-1-cost-wall-clock.ndjson` - task wall-clock and model-cost ledger.
- `evidence/round4-clean-gate-claude-review.md` - Claude review of the round-4 clean-gate remediation.
- `release/publish-site-release-plan.md` - milestone-2 release plan only.
- `evidence/local-shell-viewport.png` - browser screenshot of the local shell preview.

## State Diagram

```mermaid
flowchart LR
  Start["Round-3 clear to start"]
  Task["Helios task executing"]
  Shell["A: local shell"]
  KG["B: KG-0 draft records"]
  Eval["Eval spike"]
  Scorecard["Build-1 scorecard and SPC point 2"]
  Review["Claude Code review gate"]
  Deploy["Milestone 2: gated public deploy"]

  Start --> Task
  Task --> Shell
  Task --> KG
  Task --> Eval
  Shell --> Scorecard
  KG --> Scorecard
  Eval --> Scorecard
  Scorecard --> Review
  Review --> Deploy

  Deploy -. "blocked until Rust publish-site adapter, release evidence, and Philip approval" .-> Deploy
```

## Human Reviewer Snapshot

What changed since the last review:

- Preparation moved into first execution artifacts.
- S1-S7 external source hashes were verified against the manifest.
- The first local Varro site shell and KG-0 draft record set now exist.
- A structured eval output contract now exists and has passed once against the frozen eval set.
- Cost/wall-clock instrumentation now exists for the kickoff scorecard.
- Cycle 6 exposes all 15 frozen tutor-eval items as a static citation-check surface in the public shell.

What has been achieved:

- Milestone 1 is no longer only a plan. It has a concrete local shell, draft KG seed, and eval prompt ready to run.
- KG-0 has been conformed to the crystallised `KnowledgeGraphRecord` / `KnowledgeGraphRelationship` contract.
- The original non-JSON eval finding has a structured-output remediation.
- Public content deploy is live through the already-provisioned app; source-backed KG promotion remains gated; build-gate advancement remains gated.
- The citation-check surface is static and reviewable. It does not claim a live tutor runtime.

Next decision requested from Philip:

- Review whether the KG-1 draft rebuild correctly resolves the public deploy gap without weakening the source-backed promotion and tutor-runtime gates.

Next engineering slice:

- Integrate a live tutor answer runtime only after the static citation-check surface has passed review and route verification.
- Keep source-backed promotion behind the promotion adapter and authority gate.
- Send each content increment through the Claude Code review gate before deploying through the runbook.