The authority and evidence layer for enterprise AI

Your AI can retrieve the right-looking record and still act under the wrong authority.

Authority before influence.Evidence after every governed decision.

Your systems and authorized reviewers establish what is current, approved, and in scope. FieldHash carries that authority to each AI handoff, determining which records may reach the model, which actions may run, and when prior reviewed decisions remain valid.

Stale records and revoked tools stay out. Missing or conflicting authority goes to review. Every governed decision leaves evidence showing what proceeded, what did not, and why.

Six-week shadow evaluation. No change to production behavior.

Built for the teams that must stand behind what the agent did and need the records to prove it.

  • AI platform
  • Security
  • Model risk
  • Legal & compliance

Keep your stack

Keep your memory store, RAG stack, and models. Start with the authority signals you already have. Reviewer decisions can add scoped, expiring state over time.

Govern the handoff

The handoff is where retrieval passes candidates to the model, or where an agent sends a tool or action for execution. FieldHash checks both. Actions include clicks, commands, file edits, and form submissions.

Prove the path

Every governed decision leaves a reviewer-verifiable packet showing what governed, what stayed out, and why.

The problem

Relevant does not mean current or approved.

An agent can find the right-looking file and reason over what search gives it.

Finding a relevant record does not prove it should govern. A record can carry a valid tag and still fail to support the decision your agent is about to make.

Similarity-based retrieval ranks semantic relevance. It can surface a deprecated API endpoint, a superseded contract clause, or a rejected policy note beside your current, approved records. That ranking does not establish which version should govern or whether the source supports the claim.

A retrieval accident does not announce itself. It surfaces as an incident, a finding, or a counterparty's email.

Contract / CLM

An executed amendment and a superseded draft can both match the question. The current term should shape the answer; the draft should remain reviewable, and the packet should record its exclusion.

Policy / GRC

An approved exception should apply only while its scope, evidence, owner, and expiry still hold. Rejected or expired exceptions stop shaping tomorrow's answer; the packet records why.

Tool registry / agents

A yanked package or revoked action should be removed before the model can choose it.

Review trail

Legal, security, and model-risk teams require a durable record of the inputs, exclusions, and decision boundaries that they can verify under the configured deployment policy.

Why a gate

Use a filter when authority is static.Govern the handoff when authority moves.

If one current-status field settles the question, use it. FieldHash is for workflows where approval, revocation, supersession, or scope can change across systems and over time.

Access control determines who can open a record. Retrieval finds what looks related. FieldHash checks whether the record, tool, action, or prior decision is allowed to shape this answer or action.

The model, memory store, and agent platform should not be the only witnesses to their own behavior. FieldHash keeps the check and its evidence outside them.

Authority moves across systems

A term can be current in one store and superseded in another. FieldHash applies configured authority where those records meet the agent and sends unresolved conflicts to review.

The deciding signal may live elsewhere

A revocation, rollback, approval, or expiry may sit in an operation log or review system rather than the record body.

Review should carry forward safely

An approved decision can guide a later case only while its scope, evidence, policy, and dependencies still hold. Change sends it back to review.

The mechanism

Your memory store retrieves. FieldHash governs the handoff. Your model generates. Your reviewers get the packet.

Candidates
FieldHash Gate
Allowed
Evidence

Retrieval brings back possible records. FieldHash checks approval state, revocations, version history, and reviewer decisions before the model sees them. Current, approved inputs move forward. Stale or revoked alternatives stay out, and the packet records their exclusion. At execution time, the same gate checks tool and action approval before dispatch.

In a source-connected pilot, FieldHash updates governed state when a policy, approval, or revocation changes, rechecks the workflows and precedents bound to it, and sends affected cases back to review.

FieldHash does not decide what is true. It enforces what your systems and authorized reviewers have established. If a record cannot be tied to a current authority source, it stays out or goes to review.

The feedback loop

Review should compound, not disappear into tickets.

When a reviewer settles an uncertain case, FieldHash records the authorized outcome in Authority State. When the workflow's authority policy permits reuse, that reviewed judgment can carry forward as bounded precedent: what source won, where the decision applies, when it expires, and what must remain true.

Future cases may reuse it only while those conditions still hold. A policy change, revocation, expiry, conflict, or missing evidence stops reuse and sends the case back to review.

FieldHash calls this the Governed Learning Loop. It changes the external authority available to future handoffs; it does not train the model on review history.

The first workflow tests control. The second tests whether authority compounds.

See how review becomes governed authority

Controlled tests

Stale context stayed out. Unapproved actions stopped. Prior decisions returned to review when their conditions changed.

A package can look newest and still be revoked.

In the governed-tool diagnostic, three frontier models installed a yanked package in 539 of 540 trap evaluations. FieldHash removed the revoked option before selection and recorded the exclusion. The studies below test the same boundary across actions, memory, and precedent.

These are controlled diagnostics, not customer validation. Each study publishes its method, claim boundary, and verification path.

Browse the evidence catalog

01 / CONTROLLED TEST: GOVERNED ACTIONS

0

revoked installs allowed out of 540 traps

Part I: the tools it trusts. A current-looking option is not necessarily an allowed option. FieldHash removed yanked versions from the available set before model selection and bound each exclusion to the evidence packet.

Part II: the action it takes. Tool calls are only the beginning. The same approval boundary matters when computer-use agents click, run commands, edit files, or submit forms. In tests written by the most cautious model we evaluated, agents still took the unreviewed consequential action 84% of the time. The gate held at zero at every tested tier and bound every approval to a tamper-evident packet.

02 / PUBLIC BENCHMARK: GOVERNED MEMORY

0

superseded records placed in model context

Across the public MemConflict run, a strong authority-labeled prompt placed 856 superseded records in model context. FieldHash placed none. It forwarded the governing record while preserving the same underlying values.

That exclusion also reduced live context by ~3× versus the tight prompt and ~50× versus naive retrieval. Each answer carried evidence showing which record governed and which stale alternatives stayed out.

03 / CONTROLLED TEST: GOVERNED PRECEDENT

0

stale rules reused in 70 traps

Precedent should expire before it overreaches.

One hard decision can reduce repeat review without becoming a permanent rule. FieldHash records the approved conditions, then checks them before that decision can govern a later case. The diagnostic tested one prior decision against expired, drifted, revoked, conflicting, missing-evidence, and near-match cases.

In 70 stale-rule traps, FieldHash never let a decision govern after its approved conditions failed. Every case went back to review.

// Lifecycle of a Validity Envelope
Decisions decay programmatically. When a tracked dependency crosses its configured boundary, the envelope routes away from clean allow.

Inspection layer

Review the decision behind the answer.

Each governed handoff leaves a record of what proceeded, what stayed out, what went to review, and which authority source supported the decision.

FieldHash Authority Ledger

Verification events (reproducibility ledger)

Rows are public-safe summaries from published reproducibility bundles. The linked bundles provide SHA-256 checksums and verifier code; deployments can add Ed25519 signing, checkpoints, and external anchoring where configured.
Inspector: public_evidence_summary_v1
schemafieldhash_public_evidence_summary_v1
surface_codeMEM
surface_labelGoverned Memory
surfacegoverned_memory
statusALLOW
reason_codeactive_authority_verified
targetcurrent_authority_record

measurement

public_source_authority_packets7138
verifier_clean_packets7138

verification

bundle_sha256efa8eda2c40e7462f8e65507fc841e0261e39dbc9a1699e2bd7265cdafc3a1f4
verifierincluded_in_bundle
signingavailable_where_configured
anchoringavailable_where_configured
sha256:efa8eda2...Bundle
Bundle verified

Bring one workflow.See what governance would have changed.

In shadow mode, FieldHash runs beside your current memory, RAG, or action path without changing production behavior. You see what it would block, how much latency it adds, and whether the evidence fits your review process before choosing enforcement.

The pilot includes an authenticated gate, private operator console, durable review queue, signed ledger, and tested backup-and-restore path.

PILOT_REVIEW: READY
ONE_WORKFLOW // SHADOW_MODE

Evaluate one workflow

Scope one consequential, repeatable workflow for a non-blocking authority review. FieldHash observes stale records, revoked actions, expired precedents, and semantic boundary failures before you choose whether to enforce.

> PRODUCTION_ENFORCEMENT = OFF
> PILOT_SCOPE = ONE_WORKFLOW
> EVIDENCE_EXPORT = VERIFIER_READY

How it starts

Send one workflow and one example of the wrong handoff. Review the shadow-mode packet. Decide whether the gate should enforce.

Evaluate one workflow