Repository navigation
Description
A separate agent, given the code and not my conclusions, read the plant model that #13 puts into formal/PlantGate.lean. It confirmed the units and that the integer model matches the earlier floating-point one (0 ppm difference over 200,000 designs). It also found five defects. I checked each by running the numbers. They are disclosed in formal/SPEC.md ("Known defects in the model") and in section 11 of docs/MODEL.md. They are not fixed in #13, and this issue is where to decide how.
The theorems are true of the code as it stands. The defects are in what the code models.
1. The objective's outage term is wrong by a constant
objective subtracts the whole utility power of the shed load during an outage: - utility p shedP * tOut. utility has a part that does not depend on load (auxiliaries plus other overhead, less the passive sinks over the loop COP: about 381.5 kW). That part is subtracted for every outage minute even when nothing is shed.
- Reference design with no shedding: objective 914,237 with twelve outages, 914,218 with none. An outage raises the score.
- Fix: subtract only the saving,
utility(P) − utility(P − shed·P). - Effect: 914,210 for the start design (published 914,228) and 918,990 for the end of the seeded run (published 919,009). About 18 ppm for every design, so the order of designs does not change except between designs that tie.
2. The outage margin counts IT heat only
The drive and floor margins include the fixed heat load (qFix, about 213 kW). outageNum = pIt × (1 − shed) × outMin does not. In a 20-minute outage that is about 71 kWh the store is not asked to hold.
- Reference design with a 2,400 kWh store and 30% shed: taken up, margin 972,223. IT heat plus fixed heat is 2,404 kWh.
- Fix: add
qFix × outMintooutageNum. Section 4 of the model note makes the same simplification and would change with it.
3. A design that scores the same is taken up and charged
accepts asks for the objective not to fall. Equal is allowed, and so is a proposal identical to the design held.
- 100 identical proposals: 64 "changes", budget 0, nothing changed.
- 64 hardware-only steps took the store from 2,500 to 2,334 kWh and the wall from 132 to 192 mW/m²K at an unchanged objective, leaving 285 ppm of headroom.
- 30 of the 45 changes in the seeded run left the objective unchanged. Those are the hardware moves that open headroom, so a strict "must improve" rule would stop the loop working.
- Options: refuse a proposal identical to the design held; and at an equal objective, require that headroom does not fall. The second turns "never worse" into a statement about objective and headroom together. It is a design decision, and it is tied to the question in Proposal 3: a model as the proposer #6 of whether headroom should have a value.
4. Nothing requires the inputs to make sense
No statement assumes a positive IT load or COP, a capture or shed fraction at most one, a positive outage length, or a start within the bounds. With shed_ppm = 2,000,000 the outage margin is negative and RidesThrough holds. With the limit set to 10,000,000, which is what an unmeetable margin reads, a scheduler floor of zero is taken up.
- Within the declared bounds at a limit of 1,000,000, no meaningless design was found that is taken up.
- Fix: a
WellFormedpredicate on the plant, the scenario and the bounds, as a hypothesis ofloop_keeps_physics, and a check of it inverify.py.
5. Two limits on what the loop can do
- A large move towards safety is refused. Raising the store from 2,500 to 5,000 kWh moves the outage margin from 933,334 to 466,667 and is refused as
distance, because the radius bounds distance in either direction. - A start past a limit cannot be repaired. Headroom is negative and every proposal is refused as
radius. - Options: measure distance only over margins that rise; allow a step that lowers every margin regardless of distance. Either needs new statements.
What a fix costs
Fixing 1 or 2 changes every published objective and margin figure, the seeded ledger, and the ledgers of the two recorded runs in results/. The notes those runs left (results/model_lessons.md, results/model_lessons_2.md) quote the present figures. So the choice is:
- Fix and re-run both recorded runs under the corrected model. Clean, and costs two more runs.
- Fix and keep the runs, replaying them under the corrected model with a note that the proposers saw objectives about 18 ppm higher. Cheaper. It needs a check that no verdict in either run changes.
I recommend the first, done as its own pull request after #13, so the correction can be reviewed apart from everything else. I have not started it.
Two smaller findings from the same read are fixed in #13: prose that dropped the hypotheses of three theorems, and the fact that nothing executed the proved loop (the Lean executable now has a whole-run mode, and --formal uses it).
Filed by an AI agent (Claude Code) working for the owner of aska-physical.
Activity