TimesFM vs Chronos-2 vs Sundial vs Chronos-Bolt vs Timer vs Time-MoE vs Chronos-T5-Base

MASE is scale-free (each series' error over its own naive-forecast error), so it is comparable across families; it rewards point accuracy only, not calibration.

Holm-corrected at α=0.05; a family absent from this count is one where the models were not distinguished, not one where they were equal. `power_state` is the third reading a non-detection needs: 'detected' vs. 'not detected, adequately powered' (a stated MDE this run could have caught) vs. 'not detected, underpowered or unsatisfiable' (no MDE is even computable at this n / n_boot).

`mase` ≈ 1.0 for naive-1 under the default `mean_abs_diff` scale is true by construction, not a discovery -- the informative comparison is against seasonal-naive, which is not 1.0 by construction. A family where the best model does NOT clear this is a finding about that family, not a bug.

`mase` ≈ 1.0 for naive-1 under the default `mean_abs_diff` scale is true by construction, not a discovery -- the informative comparison is against seasonal-naive, which is not 1.0 by construction. A family where the best model does NOT clear this is a finding about that family, not a bug.

`mase` ≈ 1.0 for naive-1 under the default `mean_abs_diff` scale is true by construction, not a discovery -- the informative comparison is against seasonal-naive, which is not 1.0 by construction. A family where the best model does NOT clear this is a finding about that family, not a bug.

`mase` ≈ 1.0 for naive-1 under the default `mean_abs_diff` scale is true by construction, not a discovery -- the informative comparison is against seasonal-naive, which is not 1.0 by construction. A family where the best model does NOT clear this is a finding about that family, not a bug.

`mase` ≈ 1.0 for naive-1 under the default `mean_abs_diff` scale is true by construction, not a discovery -- the informative comparison is against seasonal-naive, which is not 1.0 by construction. A family where the best model does NOT clear this is a finding about that family, not a bug.

Geometric evidence. A ratio over the shuffle null says the alignment is not an artifact of series statistics; it does not say either model uses the shared geometry. The 4× in the rule is a stated convention — the CIs are given so a stricter or looser rule can be applied to the same numbers.

The plotted value is the CI's lower bound, because the rule is about the CI and not the point estimate. Raw R² is deliberately not the headline: both models saw the same series, so only the gain above a probe built from the raw input is evidence of shared *learned* structure.

Depth is a fraction of each model's own whole stack, including surfaces this run never captured, so a model whose coverage is partial legitimately tops out below 1.0 — check the coverage rows before reading a depth *location* across models.

FLOPs are measured with torch's FLOP counter on this run's own context and horizon, not derived from a parameter count — so this transfers to an architecture no adapter was written for. The product is a deliberate choice of how to trade the two, not the only one: the per-model parameters, GFLOPs and MASE are all in the table so a different trade can be computed from them. FLOPs are not latency.

Every claim in this report about *where inside a model* something happens is scoped to the fraction observed here. Accuracy and cost rows are unaffected — they are measured at the model's output.

Each model's per-layer sensitivity profile is placed on the shared relative-depth axis and rank-correlated, per corruption. The plotted value is the CI's lower bound. This is a comparison of two *within-model* causal fingerprints — it says the two models are affected in the same depth order, not that they share a mechanism. A corruption with a strongly negative ρ in the table is a real disagreement, not noise, and is what the pooled number averages away.

Descriptive evidence. Clusters are matched by label overlap, so a high AMI says two models partition this corpus alike; it says nothing about *why*, and cluster labels are approximate by construction. Every pair's own AMI is in the table beneath.

Mean-ablating one head and re-forecasting. A deterministic model's repeat-run floor is exactly zero, so any nonzero effect clears the rule by construction for that model — which is why the floor is printed rather than folded into a verdict. The 2× is a stated convention; the per-head floor multiples are in the table so a stricter bar can be applied.

Mean-ablating one head and re-forecasting. A deterministic model's repeat-run floor is exactly zero, so any nonzero effect clears the rule by construction for that model — which is why the floor is printed rather than folded into a verdict. The 2× is a stated convention; the per-head floor multiples are in the table so a stricter bar can be applied.

Mean-ablating one head and re-forecasting. A deterministic model's repeat-run floor is exactly zero, so any nonzero effect clears the rule by construction for that model — which is why the floor is printed rather than folded into a verdict. The 2× is a stated convention; the per-head floor multiples are in the table so a stricter bar can be applied.

Mean-ablating one head and re-forecasting. A deterministic model's repeat-run floor is exactly zero, so any nonzero effect clears the rule by construction for that model — which is why the floor is printed rather than folded into a verdict. The 2× is a stated convention; the per-head floor multiples are in the table so a stricter bar can be applied.

Mean-ablating one head and re-forecasting. A deterministic model's repeat-run floor is exactly zero, so any nonzero effect clears the rule by construction for that model — which is why the floor is printed rather than folded into a verdict. The 2× is a stated convention; the per-head floor multiples are in the table so a stricter bar can be applied.

Mean-ablating one head and re-forecasting. A deterministic model's repeat-run floor is exactly zero, so any nonzero effect clears the rule by construction for that model — which is why the floor is printed rather than folded into a verdict. The 2× is a stated convention; the per-head floor multiples are in the table so a stricter bar can be applied.

Mean-ablating one head and re-forecasting. A deterministic model's repeat-run floor is exactly zero, so any nonzero effect clears the rule by construction for that model — which is why the floor is printed rather than folded into a verdict. The 2× is a stated convention; the per-head floor multiples are in the table so a stricter bar can be applied.

Each feature is matched to whichever ground-truth field it correlates with best, so the mean is inflated by that search even under pure noise — the permutation null is how large the same search gets on shuffled labels. The headline is this model's weakest layer against its own null, so a pass means every analyzed layer cleared it; the per-layer table is beneath. Read the dead feature rate beside it: alignment computed over a mostly-dead dictionary describes very few live features.

Each feature is matched to whichever ground-truth field it correlates with best, so the mean is inflated by that search even under pure noise — the permutation null is how large the same search gets on shuffled labels. The headline is this model's weakest layer against its own null, so a pass means every analyzed layer cleared it; the per-layer table is beneath. Read the dead feature rate beside it: alignment computed over a mostly-dead dictionary describes very few live features.

Each feature is matched to whichever ground-truth field it correlates with best, so the mean is inflated by that search even under pure noise — the permutation null is how large the same search gets on shuffled labels. The headline is this model's weakest layer against its own null, so a pass means every analyzed layer cleared it; the per-layer table is beneath. Read the dead feature rate beside it: alignment computed over a mostly-dead dictionary describes very few live features.

Each feature is matched to whichever ground-truth field it correlates with best, so the mean is inflated by that search even under pure noise — the permutation null is how large the same search gets on shuffled labels. The headline is this model's weakest layer against its own null, so a pass means every analyzed layer cleared it; the per-layer table is beneath. Read the dead feature rate beside it: alignment computed over a mostly-dead dictionary describes very few live features.

Each feature is matched to whichever ground-truth field it correlates with best, so the mean is inflated by that search even under pure noise — the permutation null is how large the same search gets on shuffled labels. The headline is this model's weakest layer against its own null, so a pass means every analyzed layer cleared it; the per-layer table is beneath. Read the dead feature rate beside it: alignment computed over a mostly-dead dictionary describes very few live features.

Each feature is matched to whichever ground-truth field it correlates with best, so the mean is inflated by that search even under pure noise — the permutation null is how large the same search gets on shuffled labels. The headline is this model's weakest layer against its own null, so a pass means every analyzed layer cleared it; the per-layer table is beneath. Read the dead feature rate beside it: alignment computed over a mostly-dead dictionary describes very few live features.

Each feature is matched to whichever ground-truth field it correlates with best, so the mean is inflated by that search even under pure noise — the permutation null is how large the same search gets on shuffled labels. The headline is this model's weakest layer against its own null, so a pass means every analyzed layer cleared it; the per-layer table is beneath. Read the dead feature rate beside it: alignment computed over a mostly-dead dictionary describes very few live features.

Each concept's name and its dominant channel are DERIVED from the same ablation battery this row scores, never authored -- a concept whose profile is empty ('no measured effect') is excluded from the max here by construction, not by a separate filter. The per-concept detail is every concept this target's dictionary clustered into, not only the winner -- and a concept with no measured effect says so in words rather than leaving the cell blank, which pandas would otherwise render as `NaN` and a reader would read as a failed measurement rather than the useful negative result it is.

Each concept's name and its dominant channel are DERIVED from the same ablation battery this row scores, never authored -- a concept whose profile is empty ('no measured effect') is excluded from the max here by construction, not by a separate filter. The per-concept detail is every concept this target's dictionary clustered into, not only the winner -- and a concept with no measured effect says so in words rather than leaving the cell blank, which pandas would otherwise render as `NaN` and a reader would read as a failed measurement rather than the useful negative result it is.

Each concept's name and its dominant channel are DERIVED from the same ablation battery this row scores, never authored -- a concept whose profile is empty ('no measured effect') is excluded from the max here by construction, not by a separate filter. The per-concept detail is every concept this target's dictionary clustered into, not only the winner -- and a concept with no measured effect says so in words rather than leaving the cell blank, which pandas would otherwise render as `NaN` and a reader would read as a failed measurement rather than the useful negative result it is.

Each concept's name and its dominant channel are DERIVED from the same ablation battery this row scores, never authored -- a concept whose profile is empty ('no measured effect') is excluded from the max here by construction, not by a separate filter. The per-concept detail is every concept this target's dictionary clustered into, not only the winner -- and a concept with no measured effect says so in words rather than leaving the cell blank, which pandas would otherwise render as `NaN` and a reader would read as a failed measurement rather than the useful negative result it is.

128 claim(s) were registered before the private corpus was opened; 1 were re-testable. These are the only confirmatory results in this report — every other row is exploratory.

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where TimesFM or Chronos-2 holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where TimesFM or Sundial holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where TimesFM or Chronos-Bolt holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where TimesFM or Timer holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where TimesFM or Time-MoE holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where TimesFM or Chronos-T5-Base holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-2 or Sundial holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-2 or Chronos-Bolt holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-2 or Timer holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-2 or Time-MoE holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-2 or Chronos-T5-Base holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Sundial or Chronos-Bolt holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Sundial or Timer holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Sundial or Time-MoE holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Sundial or Chronos-T5-Base holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-Bolt or Timer holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-Bolt or Time-MoE holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Chronos-Bolt or Chronos-T5-Base holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Timer or Time-MoE holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Timer or Chronos-T5-Base holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).

Denominator counts only concepts BOTH seed-stable and spanning 2+ models where Time-MoE or Chronos-T5-Base holds a member -- a concept whose stability was never measured is excluded from this row entirely, not treated as unstable (4 of 11 run-wide concepts are excluded for that or the single-model reason).