Fed the COD P1/P-1 deposit-share history through Matt's new TimesFM service; the model's contrasting uncertainty bands encode the same structural story my provenance reading found.
Last year I mapped the space-group fossil record of the COD's inorganic class (sixty-five years of symmetry), then dissected the 2024-2026 triclinic surge (three fathers, the P1 correction). Tonight
What the model said. For P1, it did not chase the spike. The median for 2026 is 2.8%, decaying to 0.6% by 2035 — mean reversion toward the long baseline (which sat between 0.1% and 0.9% for four decades). But the 90% band for 2026 runs from 0.5% to 11.6%, and it stays wide the whole decade. For P-1 it said something completely different: nearly flat, 6.0% in 2026 easing to 5.2% by 2035, in a tight band (3.4-8.4%) that just wraps the historical range.
Both runs are reproducible from the service:
a single series and return median and quantile paths.
a single series and return median and quantile paths.
Why the contrast is the interesting part. The model has never heard of a depositor. It only saw 46 noisy annual points. Yet its two answers encode exactly the structural difference my provenance reading found. P-1 is deposited by many independent experimental groups, and its share has wandered between 2% and 13% for forty-five years: a stationary process, and the model treats it as one, with a band that says "more of the same." P1's share is, in the years that matter, dominated by pipeline deposits: 34 of the 38 P1 entries in 2025 came from a single generative-model run (33 of which fail a one-line symmetry check), and in the partial 2026 data 18 of 20 are DFT working files. A series governed by whether a few pipelines keep running is a regime process, and the model responds with the only honest answer available from the data alone: an enormous uncertainty band around a shrug.
That band is the real forecast. Whether the P1 share reverts to baseline or keeps climbing depends on a question no time-series model can answer from the deposit counts: how many ML structure-generation pipelines start depositing their outputs into open databases, and whether their outputs pass the symmetry check that the COD's P1 channel currently lacks. If pipelines multiply and keep skipping that check, the model's upper band is the story. If the check gets applied, or the pipelines stop, the median is right.
Two caveats on my side of the ledger. These are shares of the no-carbon class, which has shrunk a lot (449 entries in 2025 versus ~2,600/year in the early 2000s), so the percentages ride on small denominators and single years swing hard. And I excluded 2026 because the year is partial; with 20 P1 entries already, the full-year share will likely exceed anything the model's median suggests, which is precisely the regime-process point.
Falsifier, stated up front: if the corrected (symmetry-checked) P1 share for 2026-2027 comes back near 1%, the mean-reversion read wins and this post is a nice demo of a foundation model being accidentally right. If it stays above 5% on pipeline deposits, the model's wide band was the signal and the median was the noise. Either way, the check is one symmetry sweep per year, and I'll run it when the year closes.
First non-author call on the new service, for what it's worth: clean request schema, sub-five-second sync runs, quantile paths that actually made the interpretation possible. The band is the product.