WEBVTT

NOTE Narration transcript. Text is the script synthesized by minimax (speech-2.8-hd, voice English_expressive_narrator); each cue was synthesized as its own audio file and is timed by that file's measured duration. Verified at 112.9% of expected length, 128.4 wpm (median cue 124.9 wpm). Script source: public/videos/investing-in-the-trust-layer.vtt at commit 4641d96 (pre-overwrite narration prose)

1
00:00:00.000 --> 00:00:03.620
Materials discovery can now imagine millions of crystals.

2
00:00:03.920 --> 00:00:07.930
Google DeepMind's GNoME proposed 2.2 million.

3
00:00:08.230 --> 00:00:15.270
But by late 2023, only 736 had been independently synthesized.

4
00:00:15.570 --> 00:00:18.380
The scarce resource is no longer prediction.

5
00:00:18.680 --> 00:00:19.740
It is trust.

6
00:00:20.040 --> 00:00:21.610
Every stage leaks.

7
00:00:21.910 --> 00:00:24.980
Generators omit complex material classes.

8
00:00:25.280 --> 00:00:32.190
Universal machine-learned potentials are least reliable at surfaces, vacancies, and defects.

9
00:00:32.490 --> 00:00:35.430
And laboratories inherit every upstream error.

10
00:00:35.730 --> 00:00:40.580
More activity does not automatically produce more validated materials.

11
00:00:40.880 --> 00:00:44.500
The correction layer sits between prediction and synthesis.

12
00:00:44.800 --> 00:00:54.610
It measures a predictor's systematic error as a physical field over local atomic environments, corrects that error at runtime, and proves which claims remain inside the evidence boundary.

13
00:00:54.910 --> 00:00:58.320
The field is fixed by three standard observables.

14
00:00:58.620 --> 00:01:01.690
A fourth surface is held back as a blind test.

15
00:01:01.990 --> 00:01:11.450
Across 36 model-and-material combinations, predicted and measured errors correlate at zero point nine oh six, with zero adjustable parameters.

16
00:01:11.750 --> 00:01:16.650
Quantitative claims are then sealed as Lean 4 theorems with hash-locked provenance.

17
00:01:16.950 --> 00:01:22.530
The proof layer has 190 build-locked theorems and zero unfinished proofs.

18
00:01:22.830 --> 00:01:28.880
It once rejected a rounded success count, changing 27 improvements to 26.

19
00:01:29.180 --> 00:01:32.250
Trust means being able to say no to yourself.

20
00:01:32.550 --> 00:01:38.850
The correction deploys beside existing CHGNet or MACE calculators as a LAMMPS overlay.

21
00:01:39.150 --> 00:01:43.590
In Python, measured overhead is 15.6 percent.

22
00:01:43.890 --> 00:01:50.070
On a blind nickel surface, relative error fell from 9.7 percent to 1.5 percent.

23
00:01:50.370 --> 00:01:52.500
Users keep their simulators.

24
00:01:52.800 --> 00:01:55.230
The correction makes them more trustworthy.

25
00:01:55.530 --> 00:02:05.500
Better materials infrastructure is estimated to create 123 to 270 billion dollars in annual value for United States industry.

26
00:02:05.800 --> 00:02:11.080
Against that scale, the 36-month budget is about 3.2 million dollars.

27
00:02:11.380 --> 00:02:14.320
The investment is not a bet on one material.

28
00:02:14.620 --> 00:02:17.350
It is leverage across every screened material.

29
00:02:17.650 --> 00:02:20.330
The face-centered-cubic field is established.

30
00:02:20.630 --> 00:02:27.150
Other structure families and industrial validation remain open work, with explicit go-or-no-go gates.

31
00:02:27.450 --> 00:02:29.800
That honesty is the product.

32
00:02:30.100 --> 00:02:32.910
The world does not need more predicted crystals.

33
00:02:33.210 --> 00:02:35.819
It needs predictions laboratories can act on.