Ztare

The Project Workbench becomes a first-party capability: a local web app over your projects, backed by a Python API that relays to the CLI. Derived judgments stay in the CLI and kernel; the server reads and writes bounded models only. From a fresh clone:

make forensic-workbench-live

This starts the API on http://127.0.0.1:8765 and the app on http://127.0.0.1:5174 and installs the app's web dependencies on first run. Requires Python 3 with the repo dependencies installed and Node 18+ / npm.

Workbench

Four surfaces for pressure-testing a project, each backed by a CLI command:

• Rubric review — critique a scoring rubric before spending a run. Covers gaming-surface coverage, evidence-anchor requirements, score-ceiling reachability without evidence, criterion independence, persona blind spots, and charter-spirit coverage, plus scenario validity and evidence gaps. Opened from the rubric health card.
• Research-map queries — the map search answers plain-language questions such as "what could falsify the thesis?" by resolving the question to a graph relation and anchor node and traversing the subject–predicate–object research graph. Runs without a model call.
• Single-claim falsification — from a project's Verdict, "How could this break?" runs an adversarial pass against the weakest node and returns inversion and refutation tests with explicit fail criteria.
• Document drafting — turn a memo, paper, or brief into a bounded starting mandate (question, thesis, next falsifier, non-claims) and pre-fill the create flow.

Workbench interaction:

• A general-purpose page-loading indicator driven by in-flight fetches, with a hard cap and silent background polling.
• Tooltips on non-obvious settings (for example, evidence-gap severity).
• File-viewer labels that name the file being opened.
• LeanMill start / open / verified cards aligned to the shared design system.

New server relays: /api/rubric-review, /api/falsify-claim, /api/project-draft — each confirms, calls the CLI, and parses the result.

Kernel

• rubrics/review_rubric.py — pre-run rubric critique with a --json contract for the workbench.
• reports/research_graph_query.py — model-free query over the subject–predicate–object research graph.
• validator/falsify_claim.py — side-effect-free single-claim adversarial review targeting the weakest node.
• scaffold/draft_project.py — mandate drafting from a source document.
• common/storage.py and common/file_io.py — a single file-backed storage-provider path.
• Workspace payload modules behind the Project Workbench contract, and refinements to trace health, compression-progress reporting, Obsidian export, eigenquestion generation, and isomorphism decomposition.

CLI

New verbs:

• ztare rubric review --project <project> [--rubric <rubric>] [--json]
• ztare research map-query
• ztare research falsify
• ztare project draft --doc <path> [--json]

LeanMill

• Solver and governance updates: agentic leaf, governed DAG search, isomorphism decomposition, move atlas, certified faithfulness, and void self-play.
• Phase timing, run standards, Lean-source handling, and agent-tool refinements.
• New and updated Lean formalizations under ztare_proofs/, including a cryptography track (Shamir secret sharing).
• A void SFT corpus with training and evaluation scripts under scripts/public/models/void_sft/.

Documentation

• README and docs/concepts/capabilities.md document the workbench pressure-test surfaces and their commands.
• Clarity pass across the concepts, guides, evidence-atlas, and reference docs.

Other run modes

docker compose --profile workbench up --build workbench        # one self-contained container serving the app + API at http://127.0.0.1:8765
make forensic-workbench-build && make forensic-workbench-api    # build once, then serve the app from the API at :8765 (no Node at runtime)
make forensic-workbench-data WORKBENCH_PROJECT=<slug>           # offline snapshot for audit or review

The container builds the React app once and serves it from the Python API on a
single port; projects/ and rubrics/ are mounted at runtime so it reads and
edits your real projects. All three images now live under deploy/
(deploy/Dockerfile, deploy/Dockerfile.workbench, deploy/Dockerfile.operator).

Also in this release

• Sample projects under projects/ to open in the workbench and copy for your own work.
• benchmarks/ ships the frozen-evaluator source run, so make first-run passes end to end.