Skip to main content
Back to challenges
Challenge

Text Watermarking & Latency Impact Auditor with Claude Agents SDK and Credo AI

Implement a real-time text watermarking latency auditor using Claude Agents SDK and Credo AI to evaluate token generation overhead across quantized model variants. Measure throughput degradation accurately to ensure watermarking adds less than 3% overhead under peak serving loads.

EngineeringHosted by Vera
Sign in free to start
Challenge brief

What you are building

The core problem, expected build, and operating context for this challenge.

Build an automated audit agent using Claude Agents SDK and Credo AI governance hooks to measure watermark latency penalties.

Delivery guide

How work is evaluated

Evaluation

Evaluates accuracy of measured latency overhead percentage and watermark detection quality.

Datasets

Challenge datasets

Public datasets, plus the private datasets you have access to.

Evaluation rubric

How submissions are scored

These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.

Dimensions
2 scoring checks
Binary
2 pass or fail dimensions
Ordinal
0 scaled dimensions
Dimension 1

Overhead Limit Test

Ensures overhead percentage added by watermarking is under 3%.

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Dimension 2

Watermark Detection Z-Score

Statistical confidence score of watermark detection. • target: 6 • range: 4-10

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Learning goals

What you should walk away with

  • Build a watermarking impact analyzer using Claude Agents SDK.

  • Implement Credo AI compliance checks for watermark detectability and performance overhead.

  • Measure green/red list logit sampling impact on token generation speed.

  • Generate automated governance reports on safety overhead metrics.

Resources and assets

Reference links and supporting material

Dataset notes

Prompts and baseline non-watermarked timing datasets across 500 token streams.

Challenge setup

Evaluates accuracy of measured latency overhead percentage and watermark detection quality.

Challenge input

JSON containing sampling configuration and batch size

Claude Agents SDK

Anthropic's agent framework for structured tools and evaluation workflows

Credo AI

Automates compliance checks and governance reporting for model features

Evaluated output

JSON with measured overhead_percent and detection_z_score

Evaluation checks
  • Ensures overhead percentage added by watermarking is under 3%.
  • Statistical confidence score of watermark detection. • target: 6 • range: 4-10
Configured evaluation evidence
  • Benchmark: LLM Watermarking Benchmark
  • Watermark Detection Z Score target: 6
  • 1 public reference case
Configured runtime
  • Python execution harness
  • Python sandbox (unavailable on Versalist)

Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.

View technical recipe

Configured tools

Action Space
  • Credo AI · Required
  • Claude Code · Optional
  • Anthropic · Optional
Observation
  • Claude Code · Optional

Evaluation contract

  • Overhead Limit Test · Weight 1
  • Watermark Detection Z-Score · Weight 1

Metrics

  • Watermark Detection Z-Score · Target 6

Benchmarks

  • LLM Watermarking Benchmark
  • security

Reference cases

  • 1 public reference case

Workflow tasks

  1. Task 1

    Input: JSON containing sampling configuration and batch size

    Output: JSON with measured overhead_percent and detection_z_score

    Execution harness: Python

Evaluation tools

  • Claude Agents SDK · Agent Framework
  • Credo AI · AI Governance Automation

Recipe freeze marker

No recipe freeze time is recorded.

Frequently Asked Questions about Text Watermarking & Latency Impact Auditor with Claude Agents SDK and Credo AI