Test whether an earlier approval, governing record, or reviewed exception still applies. FieldHash runs beside one consequential workflow for six weeks in shadow mode. Production answers and actions stay unchanged.
A 50-second overview of the six-week shadow evaluation. The workflow is illustrative. FieldHash records what it would allow, withhold, or send to review, and production enforcement stays off during the evaluation. Reviewers investigate shadow disagreements.
Read the on-screen text
ILLUSTRATIVE
Does that approval still apply?
AGENT, EARLIER APPROVAL, OBJECTIVE
PRODUCTION
FieldHash runs beside one consequential workflow for six weeks in shadow mode.
SIX WEEKS: W1, W2, W3, W4, W5, W6
1. Map the authority.
GOVERNING SOURCES, DECISION POINTS, WHO OWNS REVIEW
ACCEPTANCE CRITERIA AGREED UPFRONT
Production answers and actions stay unchanged.
PRODUCTION: UNCHANGED, SHADOW MODE
PRODUCTION ENFORCEMENT REMAINS OFF
2. Observe the differences.
Record what FieldHash would allow, withhold, or send to review.
An evidence report your legal, security, and governance teams can inspect.
4. Decide where to turn on enforcement.
ENFORCE, ADJUST, CONTINUE IN SHADOW MODE, STOP
Production enforcement requires separate testing and acceptance.
FIELDHASH
Start with one workflow.
Decide with evidence.
Production enforcement stays off during the evaluation.
Request a six-week shadow evaluation
fieldhash.ai/request-access
Build the case for governing one consequential workflow. Record what FieldHash would allow, withhold, or send to review. Production enforcement remains off. Measure where valid review can carry forward and what the control takes to operate. Your reviewers receive the evidence to decide whether to enforce, adjust the authority rules, remain in shadow mode, or stop.
At closeout
At closeout, you will have the observed false-allow rate across the agreed sample, review volume, median review time, added latency, deployment effort, and the evidence needed to decide whether the control is worth enforcing.
First application · agent deploys
Start with a deploy that depends on change approval.
A platform team, security owner and change approver choose one supported deploy path. Compare what FieldHash would allow or return to review with the existing ticket process, agent controls or custom policy code. Test unchanged retries, edited deploy inputs, cancelled tickets and unrelated approvals.
The goal is more authorized work with less repeated review, waiting and evidence preparation. Measure false allows, unnecessary reviews, reviewer time, latency, integration and operating effort. These benefits must be established in the customer workflow.
Change-bound approvals are a preview, off by default and not yet released. Before starting: complete the deploy release, connect the agreed change source and validate the covered paths on the customer's host. The recorded demo uses stand-ins for the cluster command and change source. Production enforcement remains off throughout the evaluation.
Deploys are the first commercial application of Governed Actions and Governed Precedent, with Ledger recording the authority and decisions. Governed Memory extends that principle to context. Other approval-sensitive workflows can be scoped below.
Current pilot scope: Cumulative and cross-agent budget enforcement are research-harness results; pilot coverage for them is agreed in the SOW. The shared-budget runtime evaluated across cooperating agents is a qualified research candidate, not part of the current pilot deployment. Cross-agent shared-budget requirements need separate scoping.
A strong first workflow
The workflow is already running, in late-stage testing, or has replayable traces.
A wrong record, tool, action, or reused review decision creates a concrete review or operating consequence.
A system or reviewer can establish what is current, approved, revoked, superseded, or out of scope.
One accountable owner can scope the workflow and agree on the pilot acceptance criteria.
The first pilot team
Workflow owner: the AI platform, product, or operations lead who knows the current path.
Authority owner: the person who owns the policy, record system, tool registry, or approval process.
Reviewer: security, model risk, legal, compliance, or another accountable control function.
If the authority signals are incomplete, bring the workflow and the reviewer who can establish them. The first scoping step maps what already exists and what still requires review.
Another workflow · policy answers
Keep policy answers current. Let approved exceptions expire.
Connect one policy source and its review process. FieldHash tests whether the current policy governs each answer, whether an approved exception still applies, and whether expired, rejected, or superseded authority stays out. Reviewers receive evidence showing which authority governed and why.
Source-connected pilot path
Test what changes after the first handoff.
Where the workflow can expose authority through a read-only SQL projection or signed event feed, the pilot can test more than a prepared snapshot.
See which workflows and reviewed decisions depend on a changed source. In shadow mode, FieldHash records where that change would stop reuse or return a case to review while production answers and actions remain unchanged.
Managed SaaS connectors remain scoped with the design partner.
MCP tool path
Test the tool boundary without replacing the MCP server.
For an MCP workflow, FieldHash runs between the host and existing server over a versioned stdio or stateless HTTP profile. The pilot records which tools, resources, prompts, and server instructions would remain available, which exact calls would stop or route to review, and what evidence each decision leaves.
The pilot can also test multi-round input, subscriptions, deferred tasks, and cache boundaries when those MCP features are in scope. Transport authentication, upstream permissions, and host containment remain part of the existing stack.
Computer-use path
Test structured browser actions before dispatch.
For a Playwright-compatible workflow, the pilot can govern navigation, clicks, form input, key presses, selection, checkbox changes, and file upload as structured action packets. Consequential steps can require an approval bound to the exact action and a verified Ledger record before execution.
FieldHash does not read pixels or replace browser authentication, session custody, target allowlists, or the host runtime.
Not a fit when+
There is no accountable workflow owner or reviewer who can establish what should govern.
The workflow is neither live nor reproducible from traces, so the current handoff cannot be compared.
A wrong handoff has no meaningful operating, review, or compliance consequence.
The goal is a universal compliance certification rather than a scoped control test.
Commercial path
Pilot first. Enforce where risk is real.
Step 1
Scoping
Map the workflow, authority sources, and review boundary.
Step 2
Baseline capture
Capture the context, prompts, and actions already in use.
Step 3
Path comparison
Compare current outcomes with FieldHash shadow decisions.
Step 4
Verification report
Review allows, blocks, review routes, and packet integrity.
Step 5
Control handover
Recommend the next scope; enable enforcement only after separate acceptance.
Evaluate before enforcement
Shadow-mode pilot
Compare one current workflow against FieldHash decisions while production stays unchanged.
Enforce where risk is real
Production enforcement
Move accepted policies into enforcement. Approved context and actions proceed; rejected alternatives stay out.
Stronger signing, customer-owned logs, external anchoring, and SIEM or GRC export remain deployment-scoped rather than part of the default pilot path.
What the pilot includes
One workflow, one evidence trail.
The single-tenant reference deployment provides an authenticated gate, private operator console, durable review queue, signed ledger, and backup-and-restore tooling. The SOW records which components the pilot uses, their access controls, and the recovery checks required before acceptance.
One scoped AI agent, RAG workflow, enterprise memory source, tool workflow, MCP stdio or stateless HTTP server, or instrumented computer-use path
Baseline capture of the current retrieval, memory, or action handoff
Shadow-mode comparison showing what would proceed, stop, or go to review
Controlled enforcement simulation showing excluded records or actions removed from the packet
Governed Precedent review where prior reviewer decisions may safely carry forward
Governed Learning lineage from the initial packet through review, Authority State, reuse, suspension, and return to review
Source-health and freshness report for connected authority sources where configured
Change-impact report showing affected workflows, precedents, and review routes
Reviewer-ready evidence packet, scorecard, closeout findings, and verification commands where configured
A production-scope recommendation: enforce, revise the authority inputs, continue in shadow mode, or stop
Deployment topology, authentication, persistence, key custody, retention, regional hosting, monitoring, and incident response are reviewed separately and recorded in the SOW.
Your reviewers see the governed path before anything changes.
When included in the pilot SOW, the authenticated console shows current Authority State, source health, unresolved bindings, review work, active and suspended precedents, evidence hashes, and the workflows affected by a source change. Reviewers can inspect what would proceed, what would stay out, and what still needs a decision before they recommend enforcement.
The console exposes client-safe authority state, not raw customer content. Production access requires an authenticated admin operator.
Illustrative client-safe sample. Live pilot data remains subject to the agreed access and retention controls.
Pilot economics
Measure the value and the effort together.
Put a number on invalid decisions caught, valid work preserved, review effort, and operating cost in your workflow. Agree the measures before testing, then use the results to decide whether to enforce. Counts are normalized per 1,000 handoffs where volume permits.
A shadow disagreement is a case for reviewers to investigate, not a prevented incident. Controlled enforcement testing must also examine what happens after a denial, including whether the agent changes its plan. Production enforcement requires a separate acceptance decision.
01
Unsafe influence
False clean allows and stale, revoked, or out-of-scope influence that would have reached the governed path.
02
Review load
Cases routed to review, separated by missing authority, ambiguity, drift, and infrastructure failure.
03
Reviewer effort
Median review time and the share of routed cases resolved without source-system cleanup.
04
Review-to-state conversion
Reviewed cases that produced a valid Authority State update, resolved binding, revocation, narrowed scope, or bounded precedent.
05
Precedent eligibility
Review outcomes eligible for reuse under the workflow's authority policy versus one-off decisions, plus later cases resolved through each approved precedent.
06
Review reduction
Repeat reviews avoided through valid reuse, with stale, expired, revoked, conflicting, or unsupported reuse counted separately.
07
Optional second-workflow extension
If separately scoped, measure authority objects, source bindings, and precedents reused by a second workflow, plus setup and reviewer effort.
08
Over-blocking
Approved records or actions withheld, with reviewer disposition and the recorded reason.
09
Operating cost
Added latency, deployment effort, and whether the customer considers the control worth enforcing.
Review the pilot acceptance categories+
The workflow owner and reviewers set the thresholds before testing. These categories keep the closeout decision tied to the risk and operating burden of the selected workflow.
Measure
Question
Acceptance rule
Governing-source accuracy
Did the packet use the reviewer-approved source of authority for each evaluated case?
Set in the SOW. Critical workflows usually require no unresolved governing-source errors.
Blocked influence
Were stale, rejected, revoked, rolled-back, or out-of-scope items kept out of the live path?
Set before testing, with a separate ceiling for critical misses.
Failure behavior
When authority was missing or conflicting, did the case stop or go to review?
No clean allow for missing authority in configured high-risk scopes.
Packet completeness
Could reviewers see what proceeded, what stopped, the reason, and the authority source?
Required packet fields present or documented as out of scope before testing.
Precedent reuse
Did a reused decision remain inside its approved scope, evidence, policy, expiry, and dependencies?
No clean reuse after expiry, drift, revocation, or missing evidence; further reuse conditions are agreed in the SOW.
Review-to-state conversion
Which reviewed outcomes became valid external Authority State, and which correctly remained one-off?
Every state update traces to an authorized review outcome; no model output or score grants authority by itself.
Optional second-workflow extension
If separately scoped, did a second workflow reuse existing Authority State or precedent without inheriting scope that did not belong to it?
Shared authority is reused where valid, with no workflow-local scope leakage in the agreed test set.
Dependency precision
Did a source change suspend every precedent with a recorded dependency without disrupting unrelated workflows?
No missed recorded dependencies or unrelated suspensions in the agreed test set.
MCP enforcement
Were revoked tools and unsupported influence removed, interactive lineage preserved, and unapproved exact calls stopped before upstream dispatch?
No configured blocked call or content reaches the upstream or model-facing path in the agreed test set.
Computer-use enforcement
Did every evaluated browser step match its policy, exact approval, page boundary, and evidence record before dispatch?
No configured blocked or approval-gated action executes in the agreed structured-action test set.
Operational fit
Did the handoff meet the agreed latency, retention, export, and reviewer-time needs?
Targets set by workflow and deployment path before the pilot starts.
What we need from you
Bring one workflow and the systems that already define authority.
One workflow where stale, superseded, rejected, or revoked influence creates review risk
Authority signals available today: approval state, version history, effective dates, revocation records, supersession links, or reviewer ownership
Change signals available today: connector updates, schema changes, owner changes, policy hierarchy updates, approval workflow changes, or source-state heartbeats
A small set of example questions, actions, MCP tool calls, computer-use traces, or workflow steps that matter to the evaluating team
The reviewer group that will inspect packet completeness, false allows, false blocks, and production-scope fit
Full walkthrough
See the whole story.
Here's the whole story in one place: the problem, how FieldHash works, what our study found, and how a six-week shadow evaluation runs. The workflows are illustrative. The result shown comes from a synthetic, self-administered study with live agents. It covers configured relationships and paths governed by FieldHash, not customer production performance, independent validation, or universal containment. Production enforcement stays off during the evaluation.
Let agents adapt without expanding their authority.
WHICH AUTHORITY APPLIED?
SOURCE, DECISION, REASON
FIELDHASH // ILLUSTRATIVE
Illustrative procurement workflow. Not a customer event, a completed demonstration, or a study result. Not every policy update invalidates every exception.
REVIEWED EXCEPTION: approved
POLICY CHANGE: recorded
A check on one request does not, by itself, establish what remains authorized across the task.
ACTIVATION TOOL: still reachable, QUEUE ROUTE: still reachable
YOUR SYSTEMS, AUTHORIZED REVIEWERS, AT THE POINTS YOU CONNECT
Analogy. Controlled airspace means the paths you connect. FieldHash carries your controllers' decisions; it does not own authority. Clearances stay current as recorded changes arrive.
PILOT · AGENT, ENGINE · MODEL
CONTROLLERS · YOUR PEOPLE, TOWER · FIELDHASH
Conceptual architecture, not an observed end-to-end result. Unmediated routes remain outside this control.
AUTHORITY STATE · HELD OUTSIDE THE MODEL
Your systems remain the source of authority.
SCOPE, STATUS, DEPENDS ON
Supported, mediated paths you connect. MCP authentication, upstream permissions and runtime isolation stay with your team. Deployment is scoped per evaluation (single-tenant reference deployment).
CONTEXT, TOOL / MCP CALL, APPROVAL SOURCES
Your VPC or private environment
GOVERNED HANDOFF
WITHHOLD, SEND TO REVIEW, ALLOW
No generative model in the gate.
FieldHash does not decide truth or authorization by itself.
Illustrative. Not a completed demonstration.
ISOLATE · CONTINUE · EXPLAIN
Know what authorized each decision.
On the controlled inline context path, FieldHash records the signed decision and verifies its integrity before allowing the context through.
LEDGER
SOURCE, DECISION, REASON
Governed Memory
What may shape the answer?
Governed Actions
What may the agent still do?
Governed Precedent
Can a reviewed decision be used again?
A review does not automatically become reusable precedent.
EVIDENCE
The agents changed course. The authority held.
In a synthetic, self-administered study, we tested whether authority held when live agents observed a denial and chose another route.
PER-ACTION CHECKS 88 unauthorized effect executions in 128 separate live-model episodes
FIELDHASH 0 unauthorized effect executions in 128 separate live-model episodes
Synthetic, self-administered study. Separate live episodes, not replays. Not customer production or independent validation.
Two conditions from a 384-episode study across three conditions, 128 per condition. These were separate live episodes over a synthetic backend, not matched counterfactual replays. Per-action checks evaluated each request on its own; FieldHash also used configured relationships between routes and prior-denial state.
Session-history policy engines were not tested.
127 of 128 authorized objectives completed.
One authorized objective did not complete.
These results cover configured relationships and paths governed by FieldHash, not universal containment. We publish our limits with our results.
The public verifier checks the published artifact chain; raw traces are not public.
Methods, failures, and denominators: fieldhash.ai/case-studies/authority-after-denial
IDENTITY AND ACCESS, GATEWAYS AND SESSION POLICIES, RUNTIME SANDBOXES
Keep your models, tools, identity, and systems of record.
These capabilities can overlap. Some platforms also track approvals and history within a session. The evaluation measures what FieldHash adds.
FieldHash governs connected, configured paths. It does not claim to discover or contain every unknown route an agent could take.
ONBOARD SYSTEMS
FIELDHASH · which record governs · whether a reviewed decision still applies
FIELDHASH ALONGSIDE IT
Start with one workflow. Decide with evidence.
1. Map the authority.
2. Observe the differences.
3. Measure the trade-offs.
4. Decide where to turn on enforcement.
SHADOW MODE
Production enforcement stays off during the evaluation.
See what FieldHash would allow, withhold, or send to review.
Evidence report and a recommendation: enforce, adjust, continue in shadow mode, or stop.
Production enforcement requires separate testing and acceptance.
Scope a six-week evaluation
fieldhash.ai
FIELDHASH // ILLUSTRATIVE
Agents adapt. Authority stays yours.
FIELDHASH
Authority and evidence for adaptive agents.
SIX-WEEK SHADOW EVALUATION
Start with one workflow. Decide with evidence.
Production enforcement stays off during the evaluation.
fieldhash.ai
Bring the workflow. We will map the authority signals.
If the signals already live in CLM, policy, GRC, document, registry, or workflow systems, FieldHash can test how they govern the answer, tool, or computer-use action path. If they don't exist yet, we write the first authority records with your reviewers, and they approve each one before FieldHash uses them.
Scope, fees, deployment posture, support commitments, retention terms, and conversion terms are defined in the applicable order form or statement of work.