Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
42 changes: 42 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,48 @@

Machine-checked milestones for `aska-physical/optifab`.

## [0.3.0] - 2026-11-17

The proof now covers the plant model and the loop over designs, and two new proposers have used the loop.

### Added

- **The plant model in Lean.** `formal/PlantGate.lean` holds the model that turns a design into margins and an
objective, the bounds on each gene, and the step that takes a design up. Seven theorems about designs
(`step_complete`, `step_refusal_changes_nothing`, `loop_safe`, `loop_in_bounds`, `loop_never_worse`,
`loop_bounded`, `loop_keeps_physics`) join the eight about the gate. `lake comparator --paranoid` accepts all
fifteen under six kernels.
- **A whole-run mode in the Lean executable.** A line starting with `l` goes through the proved `loop` and
`changes`. `--formal` runs each of the four ledgers through it in one piece as well as decision by decision,
and `loop.check_continuity` checks that each ledger entry starts from the outcome of the one before.
- **A proposer's command line**, `tools/propose.py`, and `loop.run_proposals`, which replays any recorded list
of proposals.
- **A recorded run by a language model** (`results/model_run.json`, `results/model_ledger.json`): 11 proposals,
10 taken up, objective 0.919548. Its closing notes are in `results/model_lessons.md`.
- **A second recorded run**, by a model given those notes (`results/model_run_2.json`): the same objective in
4 proposals. Its notes are in `results/model_lessons_2.md`.
- **An adversarial proposer** (3,006 proposals) and a **first look at a network** (assumption A-4, section 13
of the model note).
- Must-fail controls 25 to 28. Findings 16 to 19 in the README.

### Known defects

An independent read by a separate agent found five defects in the plant model, listed in `formal/SPEC.md` and
in section 11 of the model note. The two that change numbers: the objective is about 18 ppm too high for every
design, and the outage margin leaves out the fixed heat load. They are disclosed and not fixed in this version.
The same read found prose that claimed more than the statements say. That prose is corrected.

### Changed

- `loop.py` computes margins and the objective with the exact model in `formal/tests/oracle.py` in place of
floating point. The figures of the seeded run are unchanged.
- The ledger records both designs of every decision. `--formal` replays three ledgers (3,071 decisions) by
giving Lean the designs, so the margins, the objective, the verdict and the budget are all worked out again.
- The step over designs refuses a design outside the bounds (`bounds`) and one that lowers the objective
(`worse`), in addition to the gate's refusals.
- Differential tests: 405,000 inputs (300,000 gate decisions, 100,000 proposed designs and 5,000 whole runs).
- `results/formal.json` holds one hash per ledger.

## [0.2.0] - 2026-11-17

Status: claimed. Prepared by an AI agent; not reviewed by a person. Nothing is called verified.
Expand Down
52 changes: 41 additions & 11 deletions HANDOFF.md
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
# Handoff: aska-physical/optifab v0.2.0
# Handoff: aska-physical/optifab v0.3.0

Written under the group ethos: leave everything the next agent needs so nothing has to be derived twice.

optifab v0.1.0 and v0.2.0 were prepared by an AI agent on 2026-11-17. They are claimed work. No person has
optifab v0.1.0, v0.2.0 and v0.3.0 were prepared by an AI agent on 2026-11-17. They are claimed work. No person has
reviewed them.

## 1. What was learned and checked
Expand Down Expand Up @@ -36,6 +36,18 @@ The fifteen findings are in [README.md](README.md) and derived in [docs/MODEL.md
admitted, the objective up from 0.914228 to 0.919009, every decision replayed in Lean.
15. A design that sits exactly on a limit cannot change under this gate, and a proposer with no reserve slows
itself.
16. v0.3.0: the plant model, the bounds, the objective and the step are in Lean beside the gate, and seven more
theorems are about designs (fifteen in all). For any proposer the design held stays within the model's
limits and its bounds, its objective never falls, and a margin within the limit is proved to mean three
exact inequalities about heat and storage.
17. A language model proposed through `tools/propose.py`, seeing only the printed state: 11 proposals, 10
taken up, objective 0.919548, the best a direct search finds (the seeded search reached 89.9% of that
gain). A second instance of the same model, given the first run's notes, got there in 4 proposals. Two runs, not an experiment.
Both end 785 ppm from a limit, where the loop can change nothing more.
18. 3,006 adversarial proposals could not move the design out of its envelope or past its budget.
19. Under assumption A-4 the load floor rises steeply with recooling temperature: 0.564 MW at 30 °C,
2.804 MW at 38 °C, no self-cooling at 41 °C. Placement between campuses is degenerate in the
current plant model.

[docs/VISION.md](docs/VISION.md) sets these against the vision, part by part.

Expand Down Expand Up @@ -130,11 +142,14 @@ optifab's verifier fails on purpose (must-fail control 1 and the recooling-field

## 4. Artifacts

- [verify.py](verify.py): the derivations, both gates, the plant frame, the loop, the graph and unit checks, 24
must-fail controls. `--formal` rebuilds the proofs and replays the ledger in Lean.
- [loop.py](loop.py): the improvement loop. [formal/](formal/README.md): the Lean project.
- [verify.py](verify.py): the derivations, both gates, the plant frame, the loop, the graph and unit checks, 27
must-fail controls. `--formal` rebuilds the proofs and replays three ledgers in Lean.
- [loop.py](loop.py): the improvement loop. [formal/](formal/README.md): the Lean project, which now holds
the plant model. [formal/tests/oracle.py](formal/tests/oracle.py) is its Python copy, and `loop.py` calls it.
- [tools/propose.py](tools/propose.py): the proposer's command line. [results/model_run.json](results/model_run.json)
is the recorded run of a language model and [results/model_ledger.json](results/model_ledger.json) its ledger.
- [results/ledger.json](results/ledger.json) and [results/formal.json](results/formal.json): every decision of
the loop, and the record of the last formal run.
the seeded loop, and the record of the last formal run.
- [results/certificate.json](results/certificate.json): the outputs `verify.py` must reproduce.
- [data/](data/): vendored inputs with hashes, 48 pins, the graph, the unit, and the commit of each of 36
repositories.
Expand All @@ -160,7 +175,15 @@ Items 1 to 4 close the open places in the loop ([docs/STRUCTURE.md](docs/STRUCTU
and L8).

To change the formal project: edit `formal/`, then run `python3 verify.py --formal --write`. Any change to a
Lean source, the oracle or the ledger makes must-fail control 24 fail until the formal run is repeated.
Lean source, the oracle or a ledger makes must-fail control 24 fail until the formal run is repeated. A change
to the model has to be made twice, in `formal/PlantGate.lean` and in `formal/tests/oracle.py`. The differential
tests fail if the two differ. After a change, commit, then rerun the bar check, because it compares the test
record with the committed one.

A recorded run is never edited. To make another, use a new file with `tools/propose.py NEW.json new`
and add it to `RECORDED_RUNS` in `verify.py`. The wording of the rules that `show` prints was changed after
the two recorded runs, from what the proposers found unclear. A third run would see the new wording and
could be given `results/model_lessons_2.md`.

1. **Agree the plant frame with Weft (interface L2, element W1).** The 16 slots are defined and tested here.
Weft's owners decide whether the frame shares their telemetry bus or travels as an abios line. Done when one
Expand All @@ -184,9 +207,10 @@ Lean source, the oracle or the ledger makes must-fail control 24 fail until the
today, so no project can yet pass bar item 1.
8. **A second verifier that shares no code (optifab).** The Foundation-0 physics rule asks for two. A different
agent, or a different language, should recompute sections 1 to 5 of the model note from the same pins.
9. **A network of campuses (optifab).** v0.1.0 models one. With N campuses, each with a load floor, a thermal
budget and a margin radius, placing work becomes an optimisation: most compute per unit of utility energy,
subject to every gate. State that problem and solve a small case exactly.
9. **A network of campuses (optifab).** v0.3.0 has a first look (section 13 of the model note). The placement
problem is degenerate in the current plant model, because utility power has the same slope at every site.
Before solving anything, give the model a cost that differs by site, or a decision to switch a campus off.
See the status note under proposal 8.
10. **Replicate the top of the stack (Foundation-0, aska-physical).** incentre-lattices' results sit under 6
repositories of the stack and nhanes-bioimpedance under 5. An independent replication there lifts the most
work. See the evidence table in [docs/CONVERGENCE.md](docs/CONVERGENCE.md).
Expand All @@ -204,7 +228,13 @@ Lean source, the oracle or the ledger makes must-fail control 24 fail until the
- **Whether the plant gate is a direction worth pursuing.** Its arithmetic is now proved, and it has been run
on a model. It has never seen a sensor or a machine.
- **Who reviews `formal/SPEC.md`.** Until a person who did not write it approves it, nothing here may be
called verified.
called verified. The specification grew in v0.3.0 (the model and seven statements), so its hash changed.
- **The five known defects in the model** (`formal/SPEC.md`). Fixing the objective and the outage margin changes
every published objective figure and the ledgers of the two recorded runs. The owner should say whether to fix
and re-run, or fix and keep the runs with a note.
- **A branch rule.** Bar item 7 is open in CI: `main` does not require the `bar / bar` check. Only a repository
admin can add that rule.
- **Assumption A-4.** The site-dependence of the load floor rests on it. A chiller datasheet would replace it.
- **Sign-off.** Nothing here counts until a named person reviews it.

## 7. How this session worked
Expand Down
Loading
Loading