Data Labs I Corpus report · 2026-10-20

Data Labs I

Produced by Shippr, Phazur Labs' AI-native studio, with the Phazur Labs engineering team as the human in the loop: they write the briefs, approve the plans and steer the work. The system was engineered from AI design challenges and custom design prompts by the team, then tuned on real use cases from consented challenge participants, volunteer users and engineers. Agents: Claude Code (Anthropic), Codex (OpenAI) and Cursor. Model-written text stays out of any training sale until those providers authorize it.

Phazur Labs · ConfidentialPlanning estimates, not an appraisal

Phazur Labs · Shippr studio · Human in the loop · Snapshot 2026-10-20

Data Labs I

Ten months of Shippr, our AI-native studio, at work. Every task starts from a real person's need. The Phazur Labs engineering team writes the brief, approves the plan and steers the work, while agents build, test and ship it. This is the record of that human-centered loop, and what it's worth.

See the value

How Shippr was built. The system behind the studio was engineered by the Phazur Labs team, then tuned on real work.

AI design challengesStructured design challenges set the tasks the system learned to solve.

Custom design promptsPrompts, skills and workflows written by the engineering team, including the Sumi design system.

Consented real-world useAgents tuned on real use cases from consented challenge participants, volunteer users and engineers.

Loading the corpus summary…

What we have

The full record of the task loop: the Phazur Labs team's briefs and decisions, agent logs from every coding tool, and the code those tasks produced.

–agent tool calls logged

Agent logs by tool

Actions are tool calls, shell commands and edits. Diffs are full file patches.

–

Tool	Units	Actions	Diffs	Covers

Claude Code tool calls

What the agent did, by tool.

Repos with test suites

Tests turn finished tasks into graded tasks, the format buyers pay most for.

–

OwnedNeeds confirmation or release

What it's worth

Graded tasks built from owned repos are worth far more than raw logs. Ranges come from published per-task, per-token and per-codebase prices.

–top opportunity

Products

One-time values share one scale. Recurring and hourly lines are listed as stated.

Product	Range	On one scale	Status

Session corpus valuation

Three approaches, weighted.

–Blended range

Approach	Weight	Range

SONAR score

Gulp Data's five factors, each out of 5.

–

What has to clear first

No number in this report is final until every gate below is done.

–gates cleared

Provider terms are the hard limit. Shippr's tasks ran on Claude Code, Codex and Cursor, and Anthropic and OpenAI forbid training on their models' outputs without written permission. Until that's granted, only the human side sells: the Phazur Labs engineering team's briefs, approvals and decisions, plus owned code, tests and lessons.

Gates

Work top to bottom. Each one is tracked in the DataLabs I legal folder.

Who buys

Task and environment buyers pay the most and fit graded tasks. Marketplaces suit the human-only corpus.

–leads