Repository navigation
Releases
UNTIL 0.3.0
Product-work repositioning and a repository-level proof.
- The site and README now lead with the owner’s outcome and six product-work stages: research, priority/spec, implementation, independent acceptance, release candidate and outcome observation.
goal.policy.releasePathspackages explicitly allowed workspace files into a local release candidate with a SHA-256 manifest, path-containment checks, size limits and idempotent writes.- A recorded repository cycle changes
src/activation.js, moves the real Node regression test fromexit 1toexit 0/ACTIVATION_TEST_OK, completes 6 role-separated tasks and 6 independent acceptance checks, and preserves two immutable candidate files after later workspace edits. - The deterministic proof now contains 14/14 checks; the complete suite contains 52/52 passing tests.
- The expanded raw ZIP includes the repository cycle, deterministic proof and test report alongside all 48 real local-model trials.
- Production deployment remains a separate boundary. The new repository proof is deterministic control-plane evidence, not an LLM-quality or business-ROI benchmark.
UNTIL 0.2.0 — verified autonomy, honest economics
UNTIL 0.2.0 hardens acceptance, exports real model evidence and publishes the complete same-model comparison — including the result that does not favour UNTIL.
Measured, not promised
A frozen protocol published all 48 selection and held-out trials, without discarding errors.
| Six held-out JSON-deliverable trials | Accepted | Median time | Median model calls |
|---|---|---|---|
| UNTIL 0.1 | 0/6 | 9.589 s | 1.5 |
| UNTIL 0.2 | 6/6 | 14.105 s | 3 |
| Direct 0.1 | 1/6 | 3.910 s | 1 |
| Direct 0.2 | 6/6 | 4.299 s | 1 |
The current direct arm is 3.281× faster than UNTIL. Both current arms received strict JSON Schema, so the acceptance improvement cannot be attributed to the outer loop alone. Local provider charges were $0; hardware and electricity were not measured.
This is a bounded JSON-deliverable experiment — not production coding, customer interviews, a competitor benchmark, a business outcome or evidence that a product team can be replaced.
Engineering changes
- Owner-supplied
goal.policy.acceptanceis merged into every contract and cannot be discarded by the model. Live contracts without external acceptance fail closed. - Strict role-specific JSON Schemas for compatible providers;
structuredOutputs: falseremains available per model or globally. - Opt-in raw response tracing, role/model economics and
until export --output FILE. - A downloadable raw bundle with 102 files, SHA-256 manifest and integrity checks covering every trial, response and accepted release receipt.
- Six-language RTL site with live GitHub stars/forks, sourced capability comparisons, explicit product gaps and an assumption-based pilot calculator.
Verification
- 50/50 unit, integration and integrity tests on Ubuntu, Windows and macOS.
- 10/10 deterministic proof checks.
- Production browser validation on Windows, including RU and RTL mobile layouts.
- CI: https://github.com/Balanov/until/actions/runs/42096500007
- Raw evidence: https://github.com/Balanov/until/blob/v0.2.0/app/public/until-proof.zip
- Comparison and protocol: https://github.com/Balanov/until/blob/v0.2.0/app/public/comparison.json
Known boundaries remain: no native Git/PR delivery or managed sandbox, no production deployment of the agent core, and no external effect without a configured risk boundary. Raw tracing is disabled by default because responses can contain sensitive data.
MIT. Fork it, inspect it, challenge it.
UNTIL v0.1.0 — Done means proven
Open, local-first control plane for evidence-gated product operations.
- Durable SQLite goals, tasks, artifacts, verdicts, releases and outbox.
- Role-separated research, ownership, building, independent verification and release.
- Real evidence is required before delivery; synthetic personas cannot authorize it.
- Crash recovery, epoch stop, capability boundaries and idempotent local release receipts.
- OpenAI-compatible and replay model adapters; no live provider claim.
- Bounded strategy selection on separate held-out fixtures; agents cannot change their own goals, permissions, budgets or evals.
Verification: 43/43 tests, 10/10 adversarial proof checks, 21/21 browser checks. CI passes on Ubuntu, Windows and macOS. Benchmarks and raw traces are included on the project site.
This is an engineering MVP, not a proven replacement for a product team. Live LLM integration and paired business pilots remain the next stage.
Manifest: https://balanov.pro/until/
License: MIT