Ara arafatkatze

Candidate system prompts by problem

The system-role prompt was identical for all six problems. Each file below is an exact per-problem copy of that frozen prompt. The official problem statement was supplied separately as a user message and is intentionally not included here.

Cline IMO 2026: best observed traces

This is a targeted adaptive, model-profiled best-of-unequal-k capability ceiling. It is not a replacement symmetric benchmark and it is not pass@k. Claude Fable 5 and GLM 5.2 optimization candidates are excluded; their frozen parent baselines remain unchanged.

Model	Best observed score	Selected generation cost	Cumulative search spend
GPT-5.6 Sol	42/42	$3.2336	$3.2336
Claude Fable 5	41/42	$17.1956	$17.1956
Kimi K3	35/42	$5.1328	$19.5928
DeepSeek V4 Flash	30/42	$0.1215	$1.3122

Cline IMO 2026 — complete sdk-native-max-fast8-v1 trace index

Public evidence for the frozen 8 models × 6 IMO 2026 problems = 48 cells benchmark run through the Cline SDK native-max protocol.

Headline

Model	Score	Submitted	Total cost
GPT-5.6 Sol	42/42	6/6	$3.2336
Claude Fable 5	27/42	4/6	$32.2501

Recursive Self-Improvement Prompt for Agent Evals

You are conducting a long-running, evidence-driven improvement campaign for an AI agent harness.

Your goal is not to win a benchmark by any means necessary. Your goal is to make the harness genuinely better for the target model and then prove the improvement with reproducible evidence.

Objective

Improve the target agent on the named evaluation without changing the benchmark, adding task-specific logic, or manufacturing a better score through more resources. Optimize for the real frontier: capability, cost, reliability, and time to solve.

Cline Open Eval Ledger

This is the public index for the Terminal-Bench runs behind Cline's open-weight eval hill climb. It includes the selected model runs, exact job IDs, provider routes, scores, costs, token counts, Cline commits, checksums, and direct links to downloadable raw traces.

Start here

Interactive dashboard

Terminal-Bench 2.1 Kimi K3 Recursive Self-Improvement: Traces & Proof

Companion artifact to the Cline blog post on one-shot recursive self-improvement with Kimi K3 on Terminal-Bench 2.1. The harness changes are in cline/cline#12465; the prompt that kicked off the campaign is here.

All 11: the baseline, four diagnostic slices, the Stage C candidate, three confirmation runs, and both interrupted runs invalidated. 4.52 GB total. The failures are in there with the same care as the wins.

Start here

Complete trace manifest, links every run, including the interrupted ones, with sizes and SHA-256 checksums
[Archive README](https://uploading-sdk.s3.us-east-2.amazonaws.com/cline-builds/kimi-k3-terminal-bench-2-1-recursive-self-improvement

You are conducting a long-running recursive agent-harness improvement experiment.

The central research question is:

Can Kimi K3 inspect and improve the Cline CLI agent harness that runs Kimi K3, then demonstrate a real, reproducible improvement on Terminal-Bench 2.1 without reward hacking, benchmark-specific patches, increased resources, or misleading experimental comparisons?

You are both:

The engineering agent modifying Cline.
The target model whose performance is being improved through those Cline changes.

	#!/usr/bin/env bash

	set -u
	set -o pipefail

	export LC_ALL=C

	SCRIPT_NAME="$(basename "$0")"
	HOSTNAME_VALUE="$(hostname 2>/dev/null \|\| uname -n 2>/dev/null \|\| echo unknown-host)"
	TIMESTAMP_UTC="$(date -u +"%Y-%m-%dT%H:%M:%SZ" 2>/dev/null \|\| echo unknown-time)"

	{
	"model": "claude-opus-4-6",
	"max_tokens": 128000,
	"messages": [
	{
	"role": "user",
	"content": [
	{
	"type": "text",
	"text": "<system-reminder>\n\nUser system info (darwin 24.6.0)\nModel: Claude Opus 4.6\nToday's date: 2026-03-03\n\n# The commands below were executed at the start of all sessions to gather context about the environment.\n# You do not need to repeat them, unless you think the environment has changed.\n# Remember: They are not necessarily related to the current conversation, but may be useful for context.\n\n% pwd\n/Users/ara2/Desktop/cline\n\n% ls\nCHANGELOG.md\nCLAUDE.md\nCODE_OF_CONDUCT.md\nCONTRIBUTING.md\nLICENSE\nREADME.md\nSECURITY.md\nassets\nbiome.jsonc\nbuf.yaml\ncli\ncli-ts\ndist\ndist-standalone\ndocs\nesbuild.mjs\nevals\ngo.work.sum\nknip.json\nlocales\nnode_modules\nout\npackage-lock.json\npackage.json\nplaywright.config.ts\nproto\nscripts\nsrc\nstandalone\ntest-results\ntest-setup.js\

	<!DOCTYPE html>
	<!-- saved from url=(0053)https://console.anthropic.com/workspaces/default/logs -->
	<html class="h-screen antialiased [font-feature-settings:'ss01'] bg-bg-100 __variable_dcab32 __variable_820c23 __variable_e4195f" lang="en" data-theme="claude" data-mode="dark"><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8"><meta http-equiv="origin-trial" content="A7vZI3v+Gz7JfuRolKNM4Aff6zaGuT7X0mf3wtoZTnKv6497cVMnhy03KDqX7kBz/q/iidW7srW31oQbBt4VhgoAAACUeyJvcmlnaW4iOiJodHRwczovL3d3dy5nb29nbGUuY29tOjQ0MyIsImZlYXR1cmUiOiJEaXNhYmxlVGhpcmRQYXJ0eVN0b3JhZ2VQYXJ0aXRpb25pbmczIiwiZXhwaXJ5IjoxNzU3OTgwODAwLCJpc1N1YmRvbWFpbiI6dHJ1ZSwiaXNUaGlyZFBhcnR5Ijp0cnVlfQ=="><style>body {transition: opacity ease-in 0.2s; }
	body[unresolved] {opacity: 0; display: block; overflow: hidden; position: relative; }
	</style><meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1, viewport-fit=cover"><link rel="stylesheet" href="./Anthropic Console2_files/026a2e7f617cf4f9.css" non