Conversation
- Full visual redesign in a clean academic tech-report style (Inter, cobalt-blue accent, soft cards), replacing the earlier layout. - Bilingual EN/中文 with a persistent language toggle. - Foreground the self-evolution narrative, fused with the GDP-valued framing; stateful kept as the enabling mechanism. - Sections: hero + stat band, key contributions, abstract, three eval targets, task-group anatomy + 12-group table, evolution methods, results (bar chart + green heatmap leaderboard), auto-eval pipeline, BibTeX. - Responsive fixes: clip page-level horizontal overflow, scrollable leaderboard/BibTeX, tightened task table on mobile. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- experiments/eval_workspace/: consolidate the four eval-workspace templates (codex, codex_zh, claude_code, claude_code_zh) under one folder. Codex guides drop time recording (token only). Claude Code guides document the Anthropic four-bucket token accounting: per-response usage deduplicated by message.id, output as the per-response max, and the Opus 4.8 cost formula. - experiments/claude_code_opus_4_8_xhigh/: released Claude Code run - config, 12 task-group reports, and generated skills, parallel to codex_gpt5_5_xhigh/. - experiments/EXPERIMENT_BOARD(.zh).md: add the Claude Code results table (12 task groups x 3 conditions) with the Opus 4.8 cost formula. Claude Code (Opus 4.8 xhigh, 12 groups): skill-based evolution improves accuracy by +20.31 pp on average and reduces token cost by 8.69% on average, reproducing the Codex run's direction on a second harness. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…de Code harness Restructure the page to follow the new five-part story (GDP-worthy tasks → self-evolution → dataset construction → rule-based grading → cost-with-accuracy results → invitation), replacing the earlier contributions/capabilities framing. Adopt the "in one line" callout device and align the bilingual copy and tone. The results section now carries both harnesses: Codex GPT-5.5 and Claude Code Opus 4.8 summary tables side by side, the Codex per-group bar chart, and a new Claude Code per-group heatmap. Construction is shown as a four-step agent pipeline (task factory → calibration → 6-reviewer panel → ship) instead of a figure that is not in the repo. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- index.html: slim, results-first landing in the spirit of a benchmark leaderboard — hero, the two-harness summary tables, per-group bar chart and heatmap, and the task-group table. Long-form narrative removed. - blog.html: new page holding the full essay (GDP-worthy tasks, self-evolution, dataset construction, rule-based grading, cost, invitation), reached from a distinct Blog tab in the top-right nav. - Add a light/dark theme toggle (system-preference default, persisted), with a full dark palette. - Drop the Results/Tasks links from the top-right nav; keep them reachable via the hero and in-page scroll. - Mark the two aggregate accuracy/cost ranges as tentative (TBD); concrete per-harness numbers and per-group tables stay. - Tidy the results note into separate lines and give reference links the accent color; blog call-to-action reads "Read the blog". Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
De-italicize <em> in the hero subtitle and blog header lead (matching the .prose convention), and render both questions in the same accent blue — the dismissed question in regular weight, the real question in bold — so the pair reads as a tidy two-tone contrast instead of italic + black bold. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request rebrands the benchmark from 'Agent Learn Bench' to 'GDPevo', updating the documentation, data boards, and experiment boards in both English and Chinese. It standardizes task-specific notes across all 12 task groups by removing construction metadata, and introduces a new experiment configuration for Claude Code (Opus 4.8, xhigh) along with detailed domain-specific 'SKILL.md' playbooks. Feedback on the changes identifies a critical issue in the new config.yaml where the model name contains a trailing '[1m]' ANSI escape code remnant that must be corrected to prevent parsing or API failures.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
| @@ -0,0 +1,24 @@ | |||
| run_id: claude_code_opus_4_8_xhigh | |||
| model: claude-opus-4-8[1m] | |||
There was a problem hiding this comment.
The model name contains a trailing [1m], which appears to be a remnant of an ANSI escape code (e.g., \x1b[1m for bold text) from a terminal copy-paste. This should be corrected to claude-opus-4-8 to ensure API calls or scripts parsing this configuration do not fail due to an invalid model identifier.
model: claude-opus-4-8
No description provided.