DRIFT — Project resources¶

Development line: project:drift · thread project-resources
Last event: 2026-06-04 · 1 dated since 2026-06-04 · Researched: 2026-09-05 · confidence: high

What it is¶

DRIFT is an open claim-centric framework for deep-research agent trajectories. It localizes spans where an unverified or contradicting claim becomes harmful.

It includes Claim Keeper, Support Seeker, and Dependency Tracer. TELBench provides 1 000 expert-verified trajectory tasks from 2 790 collected runs.

This is a reproducible tool to compare trajectory audits and investigate failures, not a ready monitoring layer for a running agent.

Development line¶

• 2026-06-04 — DRIFT project page and repository were linked. On 2026-06-04, a dated message linked to the DRIFT project page and its GitHub repository. That message places both resources in the development line. It does not establish a release, version, technical change, or project launch.

What changed¶

2026-06-04 — DRIFT and TELBench were presented as open artifacts for span-level error localization in deep-research agent trajectories.

How to use this¶

From 2026-06-04, evaluate DRIFT using the linked project page and GitHub repository. The event does not establish an installation or workflow change.

1. Clone the repository and install the package in editable mode. — https://github.com/NJU-LINK/DRIFT/blob/main/docs/USAGE.md
2. Download encrypted TELBench from Hugging Face and decrypt it with the provided script; the script checks the SHA-256 hash. — https://github.com/NJU-LINK/DRIFT/blob/main/docs/USAGE.md
3. Configure OpenAI-compatible Chat Completions or Responses API via arguments or an env file. Run drift --setting drift on JSONL trajectories. — https://github.com/NJU-LINK/DRIFT/blob/main/docs/USAGE.md
4. Pass summary.json to drift-eval. The report outputs macro/micro precision, recall, F1, first-error accuracy, and missing or extra identifiers. — https://github.com/NJU-LINK/DRIFT/blob/main/docs/USAGE.md

Best practices¶

• Run bare first on the same inputs and model, then run drift. This author-intended comparison separates the claim-centric pipeline effect from the base model. — https://github.com/NJU-LINK/DRIFT/blob/main/docs/USAGE.md
• Check the checksum of decrypted TELBench. Do not commit downloaded artifacts, the key, or decrypted JSONL to Git. — https://github.com/NJU-LINK/DRIFT/blob/main/docs/USAGE.md
• Do not pass gold labels, annotations, metadata, judge outputs, or manual notes to the model. Evaluation accepts only the question and ordered raw span text. — https://huggingface.co/datasets/NJU-LINK/TELBench
• Leave --reasoning-effort low for Responses API unless you have a measured reason to change it. That is the project default for JSON-only prompts. — https://github.com/NJU-LINK/DRIFT/blob/main/docs/USAGE.md

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• Public primary sources confirm the paper submission on 1 June 2026 and its revision v2 on 2 June. They do not record a separate DRIFT release on 4 June. The date 4 June remains the event date, not a confirmed release date.
• There is no primary dated changelog after 4 June 2026 to name later code or dataset changes.

Sources¶

Agent brief¶

• Subject: project:drift, thread project-resources, 1 dated events 2026-06-04 → 2026-06-04.
• Practical note: From 2026-06-04, practitioners evaluating DRIFT should use the linked project page and GitHub repository as the available primary resources; this evidence does not establish an installation or workflow change.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.