1 · Tech watch, checked at the source
What the plan’s three resources actually say
A watch item is only useful once it has been read at the source. Each claim of the plan was checked against the paper’s PDF and code, the model cards and the GIFT-Eval result files (checked on 2026-10-30).
| Claim in the plan | What the source says | Verdict |
|---|---|---|
| IBM’s three forecasters emit quantiles natively | True (FlowState 9, TTM-r3 new multi-quantile head, PatchTST-FM 99) | holds |
| The GIFT-Eval leader has only 18.5M parameters | #1 on 12 April 2026 only with two filters (zero-shot, replication code); #13 today in that filter | partly, outdated |
| The Complexity Router: higher accuracy, much lower cost | The headline ranks series by realised test MASE (an oracle). The deployed rule, on a holdout: 0.981 vs 0.989, 33% fewer FM calls | partly |
| Specialists win on “physically constrained” systems | The paper’s energy regime is 7 series of transformer oil temperature, not load | does not transfer |
| The Substack classifies limits as workaround / not / open | Two groups only, published 24 April 2026 | partly |
Sorting series by their realised FM advantage and routing the best ones first gives the best possible curve, but it needs the test set’s answer. The deployed rule, with thresholds learned on a calibration split, matches the FM’s accuracy and saves a third of its calls. That is still useful, and much less than “70%”. Block 3 evaluates its own router out of sample from the start.
Chronos-2’s training list has no ENTSO-E or OPSD data. Chronos-Bolt’s corpus is not itemised. If it reuses the Chronos corpus, it contains Spanish national load 2015-2018, built from ENTSO-E: the same years and provider, another country. No overlap with French load is documented. All three weights are Apache-2.0; IBM’s TTM-r3 is non-commercial (CC-BY-NC-SA-4.0).
Sources and details: 01_tech_watch/README.md.