dsh-models-radar

Usage

Capability Page

1. Open Settings in the lower-left corner of the DSH Web GUI.
2. Select Model Capability.
3. Choose DeepSWE or Pompeii at the top.
4. Select a model tier from the overview or tier selector.

The page refreshes on open within the freshness windows: data served from cache costs no upstream request, and only expired datasets are refetched. Clicking a row in the overview changes the model tier used by the metrics, trend, and task diagnostics below it.

Capability Overview

Each base model is represented by its strongest tier. Expand a row to inspect all available reasoning-effort tiers. IQ bars use an absolute 0–110 scale:

IQ	Band
< 70	Needs improvement
70–84.9	General
85–94.9	Steady
95–99.9	Excellent
≥ 100	Leading

Task Diagnostics

DeepSWE uses the benchmark's real majority-vote result. Pompeii keeps its continuous F1 semantics. Filters and counts are computed locally; no extra API request is made when changing filters.

Composer Readout

The compact pill below the composer follows the model selected for the session's next request, not merely the most recently completed response. It displays:

SWE IQ 90.2   ↑ +1.4   [48h sparkline]

Matching behavior:

1. Exact model@reasoningEffort match
2. Best tier of the same base model, prefixed with ≈
3. Hidden when the base model does not exist in the DeepSWE leaderboard

The readout updates immediately when the composer model changes. The readout polls the host every 15 minutes (the shortest freshness window), so each tick costs one local request and an upstream fetch at most once per window per channel; a failed refresh keeps the last successful value.