Repository navigation
Phase 0 Plan
Joey Hammoth edited this page Nov 20, 2026
·
4 revisions
Goal: nothing may depend on an unverified assumption when the first task is authored.
Entry state (2026-11-21): protocol v0.4 exists, the repository has governance, CI and a harness skeleton, and the metric definitions are tested. No Bend artifact exists yet, and no task package exists.
| Item | Evidence |
|---|---|
| Protocol amended for agent-authored laws | Appendix J, merged in PR #1 |
| Phase 0 gaps closed against primary sources | Appendix K, docs/research/2026-11-21-phase0-source-verification.md
|
| Licence split, conduct, contribution, security, governance, citation | Merged in PR #2 |
| CI, hygiene tools, locked-file guard |
.github/workflows/, tools/
|
| CLI surface with metric definitions and known-answer tests | PR #4, 27 tests |
| GitHub Pages developer site | PR #5, site/
|
| Issue backlog on the project board | Issues #6-#15 |
| Bend 2.0.40 pinned by archive and source digest |
docker/versions.lock, docker/bend-sources.sha256, bend-image CI job (PR #19) |
model.mjs vector path proven from Node, one-cycle offset detected |
bendeda conformance compile/trace, tools/model_vectors.mjs, vector-path CI job (PR #19) |
| BendBV mask layer proved, and the unmasked control refused |
bv/bv.bend, bv/negative/unmasked_add.bend, tests/test_bv.py
|
| # | Item | Why it blocks everything after it |
|---|---|---|
| #6 | Pin the OSS CAD Suite image (Bend is done) | A result that cannot name its image digest cannot be reproduced |
| #9 | Canary suite + verdict_available
|
A checker that accepts a false law invalidates every Bend arm |
| #10 |
law lint and law mutate
|
LA3-LA7 cannot be evaluated by hand at 12 tasks |
| #11 | Proof cookbook and PEH list | Distinguishes "law is hard" from "agent cannot prove" |
| #12 | Self-hosted tools runner | Otherwise the EDA gates are skipped, and skipped is not green |
| #13 | Sealed-oracle tooling | The oracle commitment is what makes pre-registration verifiable |
Phase 0 is done when all of these are true, not when the tickets are closed:
- Three images build offline and record their digests. Bend: built in CI from the pinned digest. OSS CAD Suite: built in CI from the verified digest, tools answer, and an
sbyBMC proof passes. Harness: not yet built, and no image has been shown to build offline. -
bend model.bend -o model.mjsdrives a model from Node and a one-cycle injected offset is detected. Done —bendeda conformance trace --inject-offset 1reportsoffset detected. - BendBV's lemmas all check with no escape hatch. Partly done — the mask layer checks (
mask.idem,bv_add.comm, the concrete 8-bit wrap); closure, identity, wrap and associativity are open and listed inbv/OPEN.md. - 30+ canary controls, zero false accepts,
verdict_availablerecorded with evidence. -
law lintcatches each seeded defect in its fixture set. - The EDA gate job actually runs on the tools runner.
- A dummy commit-reveal round-trips, and a one-byte change breaks verification.
- A Bend-to-Verilog compiler (that is L4, future research).
- Track C hardware feasibility (exploratory, Phase 5).
- Any hardware-in-the-loop work (deferred extension).
| Risk | Early signal |
|---|---|
| Bend cannot be pinned to a stable release |
bend version drift across two builds of the same image |
--verdict remains broken upstream |
Canary suite can only run in default mode; D5 fallback becomes permanent (G-02) |
| Proof effort swamps the token budget | BendBV proof times trending above the cap in #8 |
| Self-hosted runner never provisioned |
bend-and-eda permanently skipped → no law gates in CI |
Code Apache-2.0 · documents and task data CC BY 4.0 · this wiki records decisions, not the current state of the code — the design document and the repository are normative.