Skip to content
Paradigm

Work  /  Live demos

Provenance

A flight recorder for agent fleets. Provenance is the studio's own product, running on the studio's own work; the demo is a real ledger from a real run, not a mockup. Open the live demo and click the red row; no account needed.

Provenance trail explorer: hash-chained event timeline of an agent run with a kill scar mid-decision, resume markers, and a decision detail pane showing the weighed options, the agent's pick, its rationale, and the human verdict.

The problem domain

Agents make hundreds of discretionary choices per task: which framing to lead with, which claim is safe to publish, when to escalate to a human. Today those choices evaporate the moment they are made. Tracing tools record what an agent did, but not what it almost did. Checkpointing systems resume the work, but a checkpoint cannot testify. And every human review of an agent decision is a labeled preference about how the organization decides, discarded at the exact moment it becomes valuable.

What we built

One append-only record whose unit is the decision: the options the agent actually weighed, with full contents and arguments, the choice, the rationale, the autonomy level of the gate it passed through, and the eventual human verdict, hash-chained with write-ahead durability so a single command verifies the whole file offline. Zero runtime dependencies, one JSONL file, 33 tests including a kill matrix that murders the process at every event boundary and inside every decision window, then proves the resumed output byte-identical under two hash seeds.

The kill and the resume

The demo's red row is the scar where we killed the process between a decision's open and its commit. Resume folded the ledger, pinned the three committed decisions, replayed the cached model calls, re-entered exactly the interrupted decision, and produced a launch post byte-identical to an uninterrupted run. The scar stays in the file on purpose: an audit record that hides its crashes is marketing, and the chain would catch the edit anyway. There is a button on the demo that flips one byte so you can watch verification fail.

From verdict to training row

Review verdicts turn the log into labels. A human rewrite of the agent's headline becomes a preference pair; an overruled risk call becomes a correction pair; a 900 millisecond rubber-stamp approval gets filtered out, with the receipt printed. Every exported row carries the decision id and event hash it came from, so a training example traces back to the audited chain. The demo's alignment panel runs on a synthetic corpus and is stamped SYNTHETIC on its face; the mechanism, the durability, and the leakage guards are the product. Take the pen yourself in the demo: overrule the headline and watch your verdict become a pair and move a policy.

The you-hold-the-pen panel: the reviewer overrules the agent's headline pick, the verdict appears as an event, the DPO preference pair materializes, and a policy replay shows held-out picks moving toward the reviewer's taste.

The production path

Provenance instruments the studio's own agent fleet through an MCP capture server, and the empirical question it exists to answer, whether an organization's accumulated verdicts measurably align a model to its judgment, is written into a registered study protocol with the test frozen before the data arrives. The design and reference implementation are described in a preprint prepared for arXiv. This is the studio's standing claim applied to itself: audit trails and human deferral mechanisms, used in anger, on our own work first.