AI answerable to evidence On-Prem · Recorder

Fieldheld

Fieldheld makes AI answerable to evidence.

Grounded answers from your own documents. Inspectable records of autonomous work. Fail-closed when the evidence isn't there.

One brand, two product surfaces: Fieldheld On-Prem and Fieldheld Recorder.

Two Product Surfaces

Fieldheld On-Prem

Private document intelligence for firms whose documents cannot leave the building — built for law and accounting practices.

  • Answers are verbatim spans copied from your own documents, cited to page and coordinates — never generated prose.
  • Fails closed: when no grounded evidence exists, the system abstains instead of guessing.
  • Multi-document matters: agreements, addenda, and amendments composed with per-span verification.
  • Native DOCX and XLSX intake, per-persona access control with privileged-document isolation, and an append-only audit trail.
  • Installs fully offline from gpg-signed bundles for air-gapped deployment; client data never leaves the premises.

Gated by an adversarial evaluation suite of synthetic casefiles — privilege traps, date-conflict traps, and not-affected traps — behind a fail-closed acceptance gate. On-Prem has no public repository, so it is presented here without a link.

Fieldheld Recorder

Evidence gates for autonomous work, so work can be inspected before it is trusted.

  • Run records: a structured account of intent, actions, and produced artifacts.
  • Evidence packets and verification reports connecting commands, diffs, checks, and review notes.
  • Policy gates with accepted, escalated, and blocked outcomes and explicit decisions.
  • Review packets and bounded rollback surfaces.
  • PR review bundles built from public GitHub PR metadata.

First Public Artifact

Fieldheld Recorder is a small synthetic CLI that creates evidence bundles and policy-gate reports for replay, review, and bounded rollback.

The public repo includes the demo, quickstart, Artifact Walkthrough, Agent Run Record Spec, run review checklist, and sample packets for accepted, rejected, escalated, and blocked runs. It can also create a public-safe review bundle from GitHub PR metadata. It is not a live agent adapter yet. Review Evidence Gate . Open Reviewer Packet Index . Inspect Blocked Run . Read Run Review Checklist . Compare Traces And Evidence Gates . Generate PR Review Bundle .

Current Boundary

Fieldheld Recorder ships with synthetic and public-safe examples so autonomous work can be inspected before it is trusted; it is not a live agent adapter. Fieldheld On-Prem has no public repository and is described here without one. Production safety, certification, compliance approval, customer adoption, live agent integration, and broader rollback remain outside the current public surfaces.