The agent writes.
The evidence decides.
PRAMAN runs authored browser and command checks against real software, applies strict proof gates, and returns honest, evidence-backed verdicts. It is not another coding agent.
A green build is not evidence. Evidence is evidence.
AI agents now write code faster than anyone can inspect it. A passing test suite proves the tests pass; it does not prove the feature works when a browser, a process, or an API is actually used. PRAMAN closes that gap.
What a passing run proves
- Declared behavior was observed in a real browser or process
- The same requirements passed in fresh repeated sessions
- Checks, screenshots and outputs were collected and retained
- The execution identity was stable across the proof
What it does not prove
- Whole-application correctness
- That unstated requirements are met
- Permission to skip human review
Three steps. No self-reported success.
Declare the checks
The agent or author writes a scenario or command declaration: steps, checks, expected values, repeats and limits.
Run the real software
PRAMAN drives headless Chromium or trusted local processes, 2–10 fresh repeats, and records what actually happened.
Return an honest verdict
Passed requires every declared check, complete evidence and a stable identity. Missing evidence never turns green.
The loop closes when the next run disagrees less.
Claimed, then checked.
The first fix removed the crash. It did not make the feature work.
Headlamp called the Kubernetes object getter on the wrong receiver. Correcting that stopped the exception; the tooltip still did not appear. The link wrappers also had to forward the DOM reference. PRAMAN exposed the difference between “no error” and “works”.
tooltipText = (props as LinkObjectProps).getName()
Real Headlamp UI, live Kubernetes namespace: props.getName is not a function.
const kubeObject = props.kubeObject;
tooltipText = kubeObject.getName()
Same action, real tooltip, five acceptance checks across three fresh sessions and a replay.
Local checkout repair, not an upstream submission. Attribution and limits.
Run it before you trust it.
Public source, real execution, no credentials.
Open the repository in Codespaces, or clone it and run the documented route. The expected result is exit 0 with twelve CLI and MCP outcomes across pass, mismatch and unavailable.
No hosted execution endpoint. PRAMAN runs trusted commands with your authority and has no sandbox, so exposing it publicly would be unsafe.
What this is. What it is not.
Is
- Authored browser and command verification with evidence
- Three explicit modes: acceptance, repair, comparison
- CLI and stdio MCP for existing coding agents
- Validated on macOS plus Ubuntu and macOS CI
Is not
- A sandbox for hostile code
- Autonomous test design or whole-app correctness
- A hosted service or published package
- Independently human-validated yet
Trust the run, not the claim. Current status · Roadmap
Kiro was used. It did not build the core.
An AI coding assistant built PRAMAN’s original core before 5 November 2026. From 7 November, Kiro CLI agents built PRAMAN’s Kiro integration (a custom agent, a finish gate on the Stop hook, steering and an installable Power) and then fixed and extended PRAMAN itself through thirteen specs in recorded sessions: Ctrl-C cleanup, failure diagnosis, new browser actions, saved starting worlds, network faults, many users at once, recorded outside calls and a one-page report. Each change was shown first on a real failing run and proven with PRAMAN. The write-up gives a session ID for every claim, including what Kiro did not do.
Evidence before you ship.
The agent writes the code. PRAMAN proves what actually happens when it runs.