Skip to content

optifab v0.2.0: one structure for four organisations' work, with a gate proved in Lean and a loop that runs on it - #2

Merged
ahrazzle merged 5 commits into
mainfrom
claude/intelligence-network-convergence-q5ti6q
Nov 17, 2026
Merged

ahrazzle merged 5 commits into
mainfrom
claude/intelligence-network-convergence-q5ti6q

Conversation

@abu-aska

@abu-aska abu-aska commented Nov 17, 2026 •

Copy link
Copy Markdown
Member

What this is

First content for optifab: the design of one physical structure that holds the work of all four organisations (aska-physical, Foundation-0, Moissanite-Labs, Proven-Commons), so that the digital and the physical infrastructure are a single unit. The idea comes from the Terrafab documents named in #1. The reference structure is funi-frame's 10 MW campus.

Status: claimed. Nothing here is measured. Every number is derived from other repositories' certificates, which are themselves models. An AI agent prepared this. Nothing in it may be called verified: the proofs are machine-checked, but no person has approved their specification.

Refs #1. The distance to the vision is tracked in #3, with eight proposals as sub-issues (#4 to #11).

How close to the vision

docs/VISION.md takes the vision apart: a fully end-to-end formally verified, recursively improving intelligence system that evolves its hardware and software together.

Part Exists Missing
Formally verified Eight theorems about the gate and the loop, proved in Lean, accepted by lake comparator --paranoid under six kernels. Proven Commons bar items 2 to 6 pass. A reviewer's approval of the specification (bar item 1, #8). Proofs about anything but the gate.
End to end One command runs certificates → margins → gate → ledger → replay of every decision in Lean (50 decisions, 0 mismatches). Both ends: no sensor feeds it (#4) and nothing obeys it (#5). 4 of 10 loop interfaces are checked.
Recursively improving Each admitted change becomes the next baseline; the proposer's step scale adapts. 45 of 50 proposals admitted, objective 0.914228 → 0.919009. The proposer is a seeded search. Nothing learns (#6).
Hardware and software One genome: 3 hardware genes, 2 software genes, one gate, one ledger. 30 hardware and 15 software changes admitted. The hardware is three numbers in a thermal model (#7). Nothing is built (#10).
One unit One predicate, one 64-byte frame, one margin vector, one record. 32 of 41 repositories placed in the structure. No shared interface in code between the organisations yet.

The two theorems that carry the weight, both in formal/:

  • run_safe: whatever is proposed, in whatever order, every margin stays at or below its limit and the budget is never overdrawn. No condition is put on the proposer.
  • run_bounded: the number of changes the loop admits is at most the budget a person set.

Both are about the gate's arithmetic. Neither says the margins fed to it are true.

Findings

Each is derived in docs/MODEL.md and recomputed by verify.py.

  1. The 3.06 MW of utility power that funi-frame's plant frees carries 2.85 to 2.96 MW of IT load. If funi-frame's dollar value is per MW of IT load, its revenue figures are 3.6% to 7.4% high. Fixed in v0.1.1: report unlocked compute in IT power; add the sorption re-cooling limit funi-frame#4.
  2. The sorption chiller's claimed COP of 0.62 requires heat rejection below 41.5 °C at best. The plant certificate carried no recooling temperature. Fixed in v0.1.1: report unlocked compute in IT power; add the sorption re-cooling limit funi-frame#4.
  3. Below about 0.56 MW of IT load the captured heat can no longer drive the chiller, so the workload has a load floor.
  4. The PCM skid is a thermal budget of 2,500 kWh_th: 15 minutes at full load, 30 minutes with half the load shed.
  5. The tightest certified margin in the building is the floor punching collar, at 80.1%.
  6. Weft's gate predicate, rewritten here, passes all 15 of Weft's weight-update vectors and agrees with Weft's own function on 100,000 generated cases.
  7. Six values copied by hand across funi-frame, funi-floors and venturi-collars agree, and are now under test.
  8. A cited map of 41 repositories and 43 links across the four organisations.
  9. On the Foundation-0 Ledger scale, the physical stack's 13 repositories are 10 claimed, 1 preprint, 2 with no result claimed.
  10. The contract's whole state fits the 64-byte frame Weft uses for telemetry.
  11. The Terrafab latency figures (15 ms against 0.5 ms) hold in direction for the plant's vibration tones, and the campus as designed does not need a fast loop.
  12. 32 of 41 repositories are placed in the structure; 4 of 10 loop interfaces are checked.
  13. An exact integer gate and the loop built on it are proved in Lean. formal/ follows the Proven Commons layout and has no dependencies.
  14. Hardware and software improve together under that gate, on a model, with every decision in a hash-chained ledger and 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 (126 of 186 proposals refused). Headroom is what makes change possible.

What is in the diff

  • verify.py: derivations, both gates, the plant frame, the loop, graph and unit checks, 24 must-fail controls. --formal rebuilds the proofs and replays the ledger in Lean.
  • loop.py: the improvement loop. formal/: the Lean project, with SPEC.md, bar.toml, differential tests and BAR-REPORT.md.
  • results/: the certificate, the ledger and the record of the last formal run.
  • data/: vendored inputs with hashes, 48 pins, the graph, the unit, and the commits of 36 repositories.
  • docs/: VISION.md, PROPOSALS.md, STRUCTURE.md, MODEL.md, CONVERGENCE.md (generated).
  • LICENSE (PolyForm Noncommercial 1.0.0, code), LICENSE-docs (CC BY-NC 4.0, text and data), NOTICE, HANDOFF.md, CHANGELOG.md.
  • Two workflows: the Proven Commons bar check on formal/, and a rerun of verify.py. Both are green on this head.

How to check

pip install sympy
python3 verify.py
# [PASS] optifab v0.2.0: 24/24 controls, Weft vectors 15/15, plant radius 0.1989 (floor_punching_collar),
#        differential test 100000 cases, extra IT 2.854 to 2.959 MW, recooling limit 41.46 C,
#        lineage depth 5 of 12, loop 45 changes admitted of 50 proposed, Lean replay 0 mismatches,
#        unit 32 repositories placed, 4 of 10 interfaces checked

python3 verify.py --formal        # needs elan and bubblewrap
# [FORMAL] comparator accepts 8 theorems under 6 kernels; 300000 differential tests, 0 failures;
#          ledger replay 50 decisions, 0 mismatches

python3 tools/fetch_upstream.py ../upstream && python3 verify.py --upstream ../upstream

What a reviewer should look at first

  • formal/SPEC.md and formal/PlantGate/Spec.lean. An AI agent wrote the specification. Until a person who did not write it approves it, the gate is proved against a specification nobody has checked (Proposal 5: a reviewer for the gate's specification, so it can be called verified #8).
  • The loop's plant model, objective and scenario in loop.py are choices. The 0.5% gain is an illustration. The behaviours in finding 15 are properties of the gate.
  • The placement and statuses in data/unit.json were assigned here.
  • The licence. Noncommercial use is open and commercial use is reserved. The copyright line in NOTICE ("Aska Physical and the optifab contributors") needs the owner's confirmation.
  • The gate only refuses. Nothing here commands a valve, a load or a machine.

Not done

Disclosure

Prepared by an AI agent (Claude Code) working for the owner of aska-physical. A second AI agent reviewed the first version, given the code and not the conclusions; HANDOFF.md section 7 lists what it found and what was fixed. Under the Foundation-0 rule that a person signs, this does not count until a named person has reviewed it.

🤖 Generated with Claude Code

https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K

optifab is the design of one physical structure that holds the work of
aska-physical, Foundation-0, Moissanite-Labs and Proven-Commons, so that
digital and physical infrastructure are a single unit. The idea comes from
the Terrafab documents named in issue 1. The reference structure is
funi-frame's 10 MW campus. Status: claimed. Nothing here is measured.

- verify.py recomputes every figure from pinned inputs and runs 19
  must-fail controls: extra IT load at a fixed supply (2.85 to 2.96 MW,
  against 3.06 MW of freed utility power), the recooling limit of the
  sorption stage (41.5 C at best), the self-cooling load floor (0.56 MW),
  the thermal budget (2,500 kWh_th) and a margin vector of 12 limit states.
- plant_gate re-implements Weft's gate predicate. It passes Weft's 15
  weight-update vectors and agrees with Weft's own function on 100,000
  generated cases. Its budget and radius are now physical quantities.
- The contract's state fits the 64-byte layout of Weft's telemetry frame.
- data/unit.json places 32 of 41 repositories in the structure and lists
  the loop's 9 interfaces; 3 are checked and the loop is open.
- data/graph.json maps 41 repositories and 43 cited links; every cited
  line is pinned to a commit and re-read by verify.py --upstream.
- HANDOFF.md records negative findings, notes for upstream maintainers
  and a queue of eleven proposed issues.

Prepared by an AI agent and reviewed by a second AI agent. Not reviewed
by a person.

Refs #1

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K
… text and data

Noncommercial use is open; commercial use is reserved to the copyright
holder. Third-party files keep their own terms, listed in NOTICE.
The owner should confirm the copyright line in NOTICE.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K
…s bar

formal/ is a Lean 4 project with no dependencies. It defines an exact
integer gate (margins in parts per million, no floating point), the
generation step that makes an admitted candidate the new baseline, and
the loop over any list of proposals. Eight theorems are proved, among
them:

- run_safe: whatever is proposed, in whatever order, every margin stays
  at or below its limit and the budget is never overdrawn;
- run_bounded: the number of changes the loop admits is at most the
  budget it was given.

`lake comparator --paranoid` accepts all eight under every kernel shipped
with Lean, with the standard axioms only. 300,000 differential tests
against the Python oracle agree. The Proven Commons bar check
(Proven-Commons/mission at 7b22e54) reports items 2 to 6 PASS and item 1
OPEN: no reviewer has approved the specification, and none can yet,
because the reviewer list is empty. The project is not verified and does
not say it is.

Also adds two workflows: the bar check on formal/, and a rerun of
verify.py.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K
…r the proved gate

loop.py holds a genome of three hardware genes (thermal store, liquid
capture, wall U-value), two software genes (scheduler load floor, load
shed in an outage) and the proposer's own step scale. Each generation a
change is proposed, the plant model turns it into 14 margins, and the
exact gate decides. Admitted candidates become the next baseline. Every
decision is recorded in a hash-chained ledger (results/ledger.json).

On the scenario defined in loop.py: 50 proposals, 45 admitted (30
hardware, 15 software), 5 refused by the gate; the objective rose from
0.914228 to 0.919009 and never fell; no admitted state passed a limit.

verify.py --formal rebuilds the proofs, runs the comparator and the
differential tests, and replays every ledger decision through the Lean
executable: 50 decisions, 0 mismatches (results/formal.json). The plain
verifier checks that this record matches the current Lean sources and
ledger. Five must-fail controls added (24 in all). Version 0.2.0.

The plant model, the objective and the proposer are not proved. The
scenario is a choice, so the figures are illustrations.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K
… the structure

- docs/VISION.md takes the vision apart (formally verified, end to end,
  recursively improving, intelligence, hardware and software, cohesive)
  and says for each part what exists and what does not.
- docs/PROPOSALS.md proposes eight systems for the open gaps, each with a
  first slice and a test that can fail.
- The unit now lists the proved gate and the loop (30 elements, 10 loop
  interfaces, 4 checked). README, MODEL, STRUCTURE, HANDOFF and CHANGELOG
  updated for v0.2.0.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012tWK9W6CVMEDEnj5eQTe6K

@ahrazzle ahrazzle left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

lfg

@ahrazzle
ahrazzle merged commit 74d68ea into main Nov 17, 2026
5 checks passed
@abu-aska abu-aska changed the title optifab v0.1.0: one structure for four organisations' work, as a checked stack optifab v0.2.0: one structure for four organisations' work, with a gate proved in Lean and a loop that runs on it Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants