Measured

Build logBuilt in public

Proofline

An open-source, multi-agent content pipeline built with LangGraph. Nothing ships without passing the evals and a human sign-off.

Give Proofline a topic and an audience. It plans an angle, researches it, writes a short technical script with citations, and a critic scores the draft before it goes to a person for approval. Proofline also runs Measured Ops: a supervisor with specialist agents, where a human approves anything public, any deploy, any spend and any message to a real person.

Everything in the repository is demo configuration and sample data. Nothing in it publishes, deploys, spends or messages anyone.

Open the Proofline repository on GitHub (opens another site)

How it works

The content pipeline

  1. Step 1: plannerPicks one angle for this audience and the questions to answer.
  2. Step 2: researcherGathers notes and gives each source an id.
  3. Step 3: writerWrites the script; every claim cites a source.
  4. Step 4: criticScores the draft against a rubric, with a quote as evidence and a concrete fix.
  5. Step 5: finalizerRenders the script with its cited sources.

The critic sends the draft back to the writer until the score passes or the revisions are used up.

The supervisor and the approval gate

  1. Step 1: call_agentRuns one agent, which proposes actions.
  2. Step 2: evaluateRuns the checks on each proposed action. A missing or failed check sends the action to a human.
  3. Step 3: triageApplies the policy: some actions are ready, some wait for a human, some are never allowed.
  4. Step 4: gateA person decides herePauses until a person approves or rejects.

A rejection with feedback becomes a new task for the agent.

What works today

Each line below is in the repository's main branch and covered by its fast tests.

  • Content pipeline with a critic loopWorks
  • Supervisor and approval gateWorks
  • Checks without a model: quality, prompt injection, personal informationWorks
  • Approval inbox sized for a phoneWorks
  • Golden set of 20 topicsTopics approved by the owner. No real output has been run against them.Topics approved
  • Deterministic eval tier, in CIIt runs in CI with a fake writer. It measures no model.Works
  • Regression baselineA placeholder: reported, never enforced.Placeholder

Built, never run

The model-graded eval tier exists and has never been run against a model. No critic calibration has been run. So this page shows no score, and makes no claim about how good the pipeline's output is.

Not built yet

  • Tools over MCP (Jira and a calendar). Until then, an approved action is a recorded decision: nothing executes.
  • Scheduled runs with budget caps and a kill switch.
  • More agents.

Log entries

There are no log entries here yet. For now the record is the repository itself: its commits and its written decision records.

Open source under the Apache License 2.0, copyright SanoVinci Inc. Built by Hassan Mansour on his own time.

Stated as the repository's README has it at commit 8fdc21c.

Open the Proofline repository on GitHub (opens another site)