Run timeline
A retained training run shown as stages with benchmark deltas and branch state.
Run stages5 retained
Five retained stages. Terminal-Bench, LHTB, and SWR100 values follow the paper's Table 2.
Base
The starting policy before any verified rewriting or policy refinement.
- Actor
- Qwen3.8-27B
- Critic
- —
- Reward
- None
Terminal-Bench 4.00.00%Baseline
Terminal-Bench 2.151.70%Baseline
Terminal-Bench 3.00.00%Baseline
Terminal-Bench Hard31.00%Baseline
LHTB process0.21Baseline
SWR1003.00%Baseline
Base, Base, retained. Terminal-Bench 4.0: 0.00%. Terminal-Bench 2.1: 51.70%. Terminal-Bench 3.0: 0.00%. Terminal-Bench Hard: 31.00%. LHTB process: 0.21. SWR100: 3.00%
Use the arrow keys, Home, or End to move between stages.FDR failed reward share85.7%
SPR failed reward share27.5%
Illustrative four-trajectory audit using the paper's equations, including one failure that passes every check but misses the terminal gate.
Implementation
1"use client";2 3import {4 useId,5 useRef,6 useState,7 type HTMLAttributes,8 type KeyboardEvent,9} from "react";10 11export type RunMetricFormat = "percent" | "decimal" | "number";12 13export type RunMetric = {14 id: string;15 label: string;16 /** Null renders the unavailable state and does not contribute a delta. */17 value: number | null;18 format?: RunMetricFormat;19 precision?: number;20 suffix?: string;21 /** Overrides the delta suffix. Percent metrics default to " pts". */22 deltaSuffix?: string;23};24 25export type RunStageKind = "base" | "rewrite" | "refine";26 27export type RunStage = {28 id: string;