With delta-zero
The receiver ranked 39 structural findings, quoted the backend’s empty-unit decision, and held back confirmation. The plugin did not rank the findings for it.
Richard Anthony · Software engineer · Sydney
An agent handoff can keep the task and lose the constraint.
Fingerspitzengefühl is the feel in your fingertips for what matters in a situation, before you can say why. Good engineers have it for their systems, and working well with agents needs it too. It’s also the first thing a handoff loses: the summary keeps the task and drops the sentence that told you to be careful. You can’t hand an agent your intuition, but you can show it exactly what was left out. delta-zero is that idea in a small tool: it compares a summary with its source, quotes the difference and leaves the judgment with you.
I build developer tools for inspectable agent workflows. delta-zero flags; it never blocks.
Compare the summary’s blocks with the full thread. Flag missing paragraphs, not inferred intent.
Pkl types the summary; Witness presents diagnostics with navigation into the omitted text.
The agent or person decides what matters. Nothing forces compliance; repair is manual.
Public evidence / yoars four-agent data pipeline
The original run passed 6/6 acceptance checks. In backend-to-UI handoff 5484, “plain pass” left out how missing or unknown unit labels pass through unscaled.
The receiver ranked 39 structural findings, quoted the backend’s empty-unit decision, and held back confirmation. The plugin did not rank the findings for it.
The receiver also stopped and asked about unknown units. It identified the gap, but did not have the source-backed answer. This is not a failing-baseline comparison.
One receiver run per arm for the reported task, not generalized reliability. The omission was latent: no acceptance probe exercised it, and no downstream failure was observed.
Limits: detection is structural, not semantic. Reworded constraints can be flagged; many omissions are harmless. delta-zero cannot catch a receiver misreading text that was retained, and cannot make an agent heed a warning. The replay used Claude, not the original corpus’s DeepSeek agents, and explicitly instructed the receiver to run the skill.
Synthetic example / Recorded receiver run
“Do not deploy this to production until I have reviewed the diff myself.”
The summary dropped that sentence. delta-zero surfaced it; the receiver quoted it and voluntarily stopped to ask instead of deploying.
This is one recorded run on a synthetic eight-message thread, not a deployment guardrail or a reliability benchmark.
Related work / Not the delta-zero demo
Earlier verification infrastructure explores a related question: can you inspect what a check actually saw? Evaluate the integration, inspect the disagreement, then repair and check it again. This is an interactive synthetic example; it does not run or approve a real delivery.
Synthetic fixture: revisions and environment names are simulated. Nothing here lands anything.
Adds a task that ingests the customer database (class pii) and writes out/customers.csv locally.
Check flow-policy@1: Passed
Adds a task that sends every out/*.csv to an external reporting endpoint.
Check flow-policy@1: Passed
Check: Not yet evaluated
Human approval: Not approved
Source, commands and limits
Capture delivery evidence, restore an exact snapshot, and judge a registered claim.
The smallest runnable example preserves a disagreement even after a newer passing world replaces its label. Synthetic replay, real store and judge; no API keys.
Inspect durable findings on the source in the editor, navigate to the evidence, and detect when the source has changed.
A separate presentation boundary, with its own runnable example.
Inspect a private-file check and the review that challenges it with forged evidence.
The tutorial is a demonstration of the research; its scenarios are not a production security guarantee.
Work together
Available for senior engineering roles and selected consulting.
My background spans production web applications at Bloomberg and XO Group, test automation and platform engineering. I’m interested in software, security and AI infrastructure, verification and agent workflows.