You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I read all three documents and built on them in #2. What was taken from them, what was not, and where it led:
What they set. The idea of one loop with two halves: the intelligence drives the structure through a gate, and "the other part of the loop is what the facility enables for the weft by providing a full end to end control loop for iteration, adaptation and evolution" (terrafab.md:86). optifab is now organised around that loop: docs/STRUCTURE.md places 32 of 41 repositories from the four organisations in one physical structure and marks which interfaces of the loop are checked (4 of 10).
The direct line to code. The header in terrafab-calc.md is where Weft's gate comes from: invariants.py says it was "repaired from weft/terrafab-calc.md:50-100". Weft's own audit of that header found it bounded "only abstract scalars, not actuator commands". In optifab the gate gets physical inputs: a budget in kWh of thermal storage and a radius measured in certified structural margins. So the gate is back at the job it was first written for.
How far that went. An exact integer version of the gate, and the loop that runs on it, are proved in Lean (formal/): whatever is proposed, no margin passes its limit and the budget is never overdrawn, and the number of changes is at most the budget a person set. A loop then changed hardware and software parameters of the campus together under that gate, with every decision replayed in Lean. It runs on a model. No sensor, load or machine is involved yet.
What was not taken. None of the documents' figures is used as evidence. The corpus README asks for that, and Weft's digest (docs/corpus/sources/B-top1.md) records why: the latency figures are asserted, the benchmark is rigged, the Lean proofs they cite did not exist. I tested one figure against this building's own dynamics: 15 ms against 0.5 ms loop latency. The direction holds for the plant's four vibration tones, but the campus as designed does not need a fast loop.
What is still an idea. The machines that make the structure's own parts (#10) and the chip (terraweft.md). Both are in the structure as ideas, with nothing behind them.
The distance to the full vision is in docs/VISION.md and tracked in #3.
Activity
abu-aska commentedon Nov 17, 2026
I read all three documents and built on them in #2. What was taken from them, what was not, and where it led:
What they set. The idea of one loop with two halves: the intelligence drives the structure through a gate, and "the other part of the loop is what the facility enables for the weft by providing a full end to end control loop for iteration, adaptation and evolution" (
terrafab.md:86). optifab is now organised around that loop:docs/STRUCTURE.mdplaces 32 of 41 repositories from the four organisations in one physical structure and marks which interfaces of the loop are checked (4 of 10).The direct line to code. The header in
terrafab-calc.mdis where Weft's gate comes from:invariants.pysays it was "repaired fromweft/terrafab-calc.md:50-100". Weft's own audit of that header found it bounded "only abstract scalars, not actuator commands". In optifab the gate gets physical inputs: a budget in kWh of thermal storage and a radius measured in certified structural margins. So the gate is back at the job it was first written for.How far that went. An exact integer version of the gate, and the loop that runs on it, are proved in Lean (
formal/): whatever is proposed, no margin passes its limit and the budget is never overdrawn, and the number of changes is at most the budget a person set. A loop then changed hardware and software parameters of the campus together under that gate, with every decision replayed in Lean. It runs on a model. No sensor, load or machine is involved yet.What was not taken. None of the documents' figures is used as evidence. The corpus README asks for that, and Weft's digest (
docs/corpus/sources/B-top1.md) records why: the latency figures are asserted, the benchmark is rigged, the Lean proofs they cite did not exist. I tested one figure against this building's own dynamics: 15 ms against 0.5 ms loop latency. The direction holds for the plant's four vibration tones, but the campus as designed does not need a fast loop.What is still an idea. The machines that make the structure's own parts (#10) and the chip (
terraweft.md). Both are in the structure as ideas, with nothing behind them.The distance to the full vision is in
docs/VISION.mdand tracked in #3.Generated by Claude Code