A distinction that matters
Language-model output can be fluent, confident and wrong. When a model proposes changing files, releasing information or completing a workflow, the operational question is different from the conversational one: what independently decides whether the action is permitted?
Consider a document-review assistant that notices a discrepancy in a synthetic invoice. Its proposal could be useful, but the underlying authority to alter the record, disclose a document or approve a payment cannot be established by that proposal alone. In a real organisation, policies, authenticated approvals, record integrity and auditability also matter.
A policy gate is one way to separate suggestions from action authority. A restricted, typed proposal can be evaluated against rules that are not supplied by the untrusted model response.
A small public experiment, not a commercial deployment
Unfire's BOLT Agent Control Plane is a separate MIT-licensed research demonstration. The current public repository applies local, deterministic rules to structured proposed actions and produces one of three classifications:
Allow: an eligible action type passes the local rule set.
Review required: a proposal is flagged as requiring separate human attention.
Deny: a proposal is disallowed by the local rule set.
These labels are read-only classifications. The experiment does not execute proposed actions, authenticate an approver, integrate with a customer's enterprise systems or enforce real-world permissions. It should not be confused with the separately distributed commercial Bolt macOS application or with Unfire Autonomy Convergence and Unfire Control Vision research.
A fixed research pre-release was published on 21 November 2026. The associated technical brief records 12 locally passing offline checks on a specified public commit. These are narrow software checks, not security certification, a formal audit or evidence of production reliability.
What a serious evaluation would still require
- Independent policy ownership: decide who can create and modify rules, rather than allowing model output to redefine them.
- Trusted identity and approval: bind review decisions to authenticated people, roles and an explicit action scope.
- Careful boundary testing: exercise malformed proposals, prompt injection, retries, race conditions and conflicting instructions.
- Reconstructible evidence: record what was proposed, the policy version used, the decision reached and any authorised downstream execution.
- Real-world validation before claims: measure limitations, operational risks and acceptance criteria with a prospective customer under agreed legal, technical and privacy conditions.
None of these higher-assurance capabilities is claimed to have been established by the public read-only experiment. They identify the gap between a useful demonstration and a deployable service.
Sources and verification
Readers can inspect the source, review the documented limitations and reproduce the supported local checks without relying on a promotional claim. Archived source is provided for persistence, not as third-party certification.
About this research
Unfire is an independent, founder-led technology R&D initiative based in Spain and operated by Omar Baró as an individual; it is not currently a separately incorporated company. The public code is an experimental research artefact. There is no claim here of enterprise customers, independent performance validation, awarded procurement contracts or endorsement by institutions whose public platforms link to Unfire.
For technical questions, corrections or media enquiries, contact hello@unfire.technology. Written correspondence is preferred.
This note is explanatory editorial material by the project author, not independent news coverage, a peer-reviewed paper, or legal/security advice.