Skip to content

Proposal 5: a reviewer for the gate's specification, so it can be called verified #8

Open

Description

@abu-aska

Closes: element R2 of docs/STRUCTURE.md and item 1 of the Proven Commons verified bar (in #2).

The gap

The gate's proofs pass every mechanical item of the Proven Commons bar (formal/BAR-REPORT.md in #2):

Bar item Status
1. Specification approved by a separate reviewer OPEN
2. Each public claim names its theorem PASS
3. Proofs match the statements, in one command PASS: lake comparator --paranoid, six kernels
4. The code that runs is the code that was proved PASS
5. Differential tests, with the record in the repo PASS: 300,000 inputs, 0 failures
6. The page lists what is outside the proof PASS
7. The check runs in CI workflow added in #2

Item 1 is open for two reasons. An AI agent wrote the specification, and no person has reviewed it. And the reviewer list in Proven-Commons/mission (bar/reviewers.toml) is empty, so no approval could be recorded even if someone reviewed it today.

Until item 1 passes, nothing in optifab may be called verified, and it does not say so anywhere.

What to do

Nothing to build. A person who did not write it:

  1. reads formal/SPEC.md and formal/PlantGate/Spec.lean;
  2. tries to write a wrong gate that still meets the eight statements (the bar's own test of a specification);
  3. checks that formal/README.md says truthfully what is and is not proved;
  4. records the approval in formal/bar.toml with the spec hash.

Done when bar/check.py reports item 1 as PASS.

Things a reviewer should push on

  • run_safe assumes the start is within the limit. Is that the right hypothesis, and is it checked where the loop starts?
  • The gate takes the radius as the headroom of the baseline. A baseline with a margin exactly on its limit then admits nothing. Is that intended? (optifab treats it as a finding.)
  • A refused step spends nothing. Weft's own gate reports a deficit on underflow. The two differ on purpose; is the difference stated clearly enough?
  • The statements say nothing about whether the margins are true. Does the page say that plainly enough?

Follow-on

Weft's open Milestone I.3: a statement relating Weft's floating-point gate to this exact one.

Filed by an AI agent (Claude Code) working for the owner of aska-physical.

Activity

  1. abu-aska commented on Nov 17, 2026

    @abu-aska
    MemberAuthor

    Three updates for whoever takes this review.

    1. What CI reports on main. After #2 merged I read the log of the bar / bar job. The Proven Commons check ran in full: items 2 to 6 pass, item 1 is open (no approval recorded), and item 7 is open with this detail: ".github/workflows/bar.yml calls the verified-bar workflow, but aska-physical/optifab's default branch main does not require the bar / bar check". Closing item 7 needs a branch rule on main that requires that check. Only a repository admin can add it. I cannot.

    2. The specification grows in #13. It adds the plant model and six statements about designs to the eight about the gate, so the hash a reviewer would approve changes. The reviewer should wait for #13 to settle, then review the specification as it stands on main after that. The current hash on the #13 branch is 5a71fafcf86394ecd730d713e65a07ea6fb659946141e1e26afc8745d68492d8.

    3. Where a reviewer should push hardest. The gate statements are short and the first review found nothing in them. The new part is a model, and a model can be exact and still say the wrong thing. The places I would attack:

    • loop_keeps_physics: are DriveHeatSuffices, SelfCoolsAtFloor and RidesThrough the inequalities the prose says, in the right units?
    • objective: it subtracts the utility power of the shed load during an outage, including its constant part. This is carried over from the earlier floating-point model and may be a modelling slip.
    • The bounds (Plant.lo, Plant.hi) are inputs. loop_in_bounds says the design stays inside them. It does not say they are sensible.
    • step_complete is the only statement that rules out a loop that refuses everything.

    A separate agent is reading the new statements against the prose now. I will post its findings on #13.


    Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions