Delta Analysis
A delta names one way run B differs from run A. Each tile says whether its delta fired, stayed quiet or was withheld, and Why gives the rule, this pair’s numbers and the parameters.
Inference: three deltas
Section titled “Inference: three deltas”These are 3.x’s rules. Where they answer differently from 2.0, the spec records the change and the evidence for it.
| Delta | What it asks | Fires when |
|---|---|---|
| ΔF | Did B gain runtime anomalies? | B has more steady samples beyond 5 robust deviations (1.4826 × MAD) from the median than chance allows: one-sided exact binomial test, p < 0.05 |
| ΔO | Did the variability change? | The 95% block-bootstrap interval on B’s relative spread, (p90 − p10) / p50 of the steady samples, over A’s excludes 1 |
| ΔTc | Did B settle at a different point? | Both runs settle, and their settle ranges do not overlap |
Notes on the rules.
- Warmup. Warmup samples are startup cost, so ΔF counts steady samples only.
- The 3-sigma rule. 2.0’s 3-sigma anomaly rule is a choice in Settings, and the caption names the rule a review used.
- Sample counts. ΔF and ΔO need 20 steady samples per side. So do the ratios in the headline.
- ΔTc and run shape. Each run gets a shape: flat, warmup, slowdown or no steady state. Only a flat run or one that warms up has a settle point.
- Where a run settles is a range, from the change points found at three penalties.
- When a RunTrace states its steady-state step, that step is used as written.
Kept off the inference page. ΔTd and ΔĀ describe training dynamics, so they do not appear for inference.
Warnings. The page warns when warmup takes more than half the run, when no steady state is found, and when a file holds only aggregated statistics.
Geometry: five deltas
Section titled “Geometry: five deltas”For two ASPIRE training runs. These are 2.0’s rules with 2.0’s defaults, ported and checked against 2.0’s results, with the corrections ruled on 2026-10-06. On exports whose steps are checkpoints × items, the rules ruled on 2026-10-07 apply instead (see Withheld, below, and Geometry).
| Delta | What it asks | Defaults |
|---|---|---|
| ΔF | Did a run fail persistently? | A failure that persists for 3 recorded failures or steps. The time and the kind come from the same failure. |
| ΔTc | Did the runs converge at different steps? | Velocity stays inside an epsilon band for 5 steps; the effective epsilon is max(0.02, 0.5 × robust sigma). Fires at 3 steps apart or more. |
| ΔTd | Did a dominant direction emerge at a different time? | The first eigenvalue exceeds 1.5 × the next, sustained or recurring |
| ΔĀ | Is one run’s evaluator spectrum more concentrated? | λ1 / Σλ per step, compared over the longest sustained stretch; on checkpoint × item exports, per checkpoint, with the same sign at every one. 2.0 called it evaluator agreement. |
| ΔO | Did one run show sustained instability? | Curvature above an adaptive threshold for 4 steps or more; floor 0.05 between runs, 0.1 within a run |
Withheld
Section titled “Withheld”A delta is withheld when the data cannot carry it. The tile then says why.
- On a geometry export that states its layout (the export contract):
- ΔTc, ΔTd, ΔO and ΔF are withheld when the steps are checkpoints × items, not time.
- ΔF is withheld when the scores were replayed or are fixed per item.
- ΔĀ is withheld when the two runs’ checkpoints do not pair, such as three against four.
- The headline speaks only from the deltas that stand.