PR #512: E2E HF input preview

Open HF E2E Gym · DB baseline · Comparison data · Input audit

Fresh campaign and Full CI passed. Shared STP predictions match exactly. This is a review snapshot, not the production dashboard.

HF preview run: 30b264ae91bf5a4fdad1ab75d3a85114f2e09094.
DB baseline: c9e7363bf5983b980fe6622f7b064ee7adff5f5d.

Numerical comparison

1,321 matched successful AISim replays have zero differences in TTFT, TPOT, E2E latency and both throughput metrics. The 35 MiniMax discrepancies are resolved.
1,052 matched successful AIC predictions agree exactly across available metrics.
1,754 shared published points retain identical measurements, configurations and prediction statuses. One ambiguous key is excluded from comparison.
Predictor source is unchanged between the two commits. The separately built wheels have different hashes.

Coverage

Measure	DB STP	HF STP	HF MTP
Selected	2281	1974	881
Published	2071	1773	197
Successful replays	1584	1340	194
Failed replays	487	433	3
Unresolved source	210	201	684

MTP has no published baseline in this comparison. Missing or shared observed acceptance stays unresolved. Different cohorts mean aggregate MAPE changes do not establish predictor improvements or regressions.

The dashboard supports All / STP / MTP filtering, with MTP settings and evidence labels in point details.

HF revision: 8ec77d043d236ae92aa0cee6a9896f8195ea7718. DB release: db-dump/2026-09-28. Source: SemiAnalysis InferenceX.

Provenance and reproduction · All review previews