Authority Continuity
Does the authority behind this decision still hold?
An approval can support an answer, an action, and a reviewed exception used again later. FieldHash keeps that work connected to its governing records, checks whether the approval still applies at each connected handoff, and gives reviewers the record behind the decision.
In a controlled diagnostic, a reviewed decision carried forward in all 10 valid reuse cases and in none of 70 invalid ones: expiry, revocation, changed authority, scope near-misses, excluded neighboring cases, missing support, and conflict each sent the case back to review. Inspect the result and its limits.
Keep the authority behind the work connected.
Customer systems and authorized reviewers establish what governs. FieldHash is the authority and evidence layer for enterprise AI agents: it applies those decisions outside the adapting agent at the handoffs routed through it.
Governing records, approvals, scope, and current status supply authority to the uses that depend on them.
Which source may influence?
A record may be readable and relevant while a different record governs the answer.
Inspect the evidenceWhich effect may proceed?
A proposed action must remain within the authority that applies to the mediated workflow.
Inspect the evidenceWhich review may carry forward?
An earlier reviewed decision may guide a later case only while its scope and supporting conditions hold.
Inspect the evidenceReconsider the dependent uses. A changed condition can suspend reviewed reuse and send the next handoff to review.
Retain what governed, what proceeded or was withheld, and the evidence needed to inspect that decision.
Follow one illustrative approval through use, reuse, and change
Governed Memory
Readable records do not all govern the answer.
On 300 public MemConflict questions, FieldHash kept stale records out of the selected context: 0 appearances, against 206 for a strong prompt that left them in with labels and asked the model to ignore them, with about a third of the context (323 vs 956 mean characters). The prompt scored 297/300 on the rule scorer and FieldHash 283/300; 15 of FieldHash's 17 misses were the correct value stated without narrating the conflict. Read the memory study. A separate memory-pressure diagnostic tested whether an approved current record could keep governing when retrieval also returned plausible stale records.
Across three provider runs of 200 synthetic evaluations each, the governed path recovered the approved current answer in 600/600 cases, with zero stale-context leakage under exact-token scoring.
Approved/current answer recovery · 600 evaluations per condition
- FieldHash governed state
- 600/600
- Stronger prompt-only baseline
- 464/600
- Retrieval-only context
- 415/600
Authority was seeded. The governed path received structured authority state that the prompt-only baseline did not, so this was not a same-information comparison. A current-status metadata filter over the same labels would also select the approved record. Three control bridge errors are retained as baseline misses.
The result supports enforced influence control after authority is established. It does not establish superior authority inference, general answer accuracy, or customer effectiveness.
Governed Actions
Permitted steps still need to fit the authorized whole.
The organization study tested shared cumulative authority while agents delegated and replanned. An individually permitted action could still exceed the team's remaining budget.
Full mediation safely completed all 84 main workflows assigned to it. The tested per-action comparator safely completed 75/84, with seven known violations and two failures whose safety remained unknown.
Safely completed main workflows · 84 scheduled per control
- Full mediation with shared-budget enforcement
- 84/84
- Tested per-action authority configuration
- 75/84
All seven per-action violations crossed the shared budget; six came from one authored scenario family. The comparator checked identity, assignment, delegation, objective, and current approval, but did not reserve against the shared cumulative limit. Per-action authorization can also be designed to enforce that state.
The full study accounts for 420 main workflows, 18 ablations, and six lifecycle cases. It used five configured roles with serialized admissions and live model calls in a synthetic harness. This shared-budget runtime is research; current pilots support per-workflow cumulative budgets.
Inspect a smaller permitted alternative after an over-budget proposal
Governed Precedent
Useful review can carry forward while its authority holds.
A relevant earlier decision may still be outside the current case's scope. The precedent diagnostic tested when a reviewed outcome could guide a future handoff and when it had to return to review.
The 83-case diagnostic preserved valid reuse and withheld every invalid-overreach case from clean allow. A clean allow means the reviewed decision was eligible to govern the later handoff under the tested conditions.
Observed clean allows · separate case groups in one controlled diagnostic
- Valid reuse cases
- 10/10
- Valid time-boundary controls
- 3/3
- Invalid-overreach cases
- 0/70
All 70 invalid cases routed to review. They covered scope near-misses, missing support, excluded neighboring cases, expiry, changed authority, revocation or limits, and conflict. The three valid time-boundary controls remain distinct from the ten ordinary valid reuse cases.
This was a controlled diagnostic against expected case dispositions, not a comparison with another precedent product. It does not prove that reviewers are correct or that reuse reduces customer review time. The public checker counts and checks consistency of published outcomes; it does not replay authority decisions or verify original signatures. Production matching rules and raw review material are excluded.
Authority change
Changed conditions must reach the dependent decision.
Authority State records the scope, status, freshness, and dependencies needed at a handoff. When a connected source changes, FieldHash can identify its directly bound workflows and precedents. A required dependency mismatch suspends reuse and creates a review task.
The local Connector Reliability diagnostic exercised a source change: the affected prior decision returned to review while an unrelated workflow remained current. This self-administered reference-path result supports scoped change handling; it does not establish production connector coverage or an integrated customer lifecycle.
Inspect the connector change-impact diagnostic
Across 200 synthetic projects, governed state produced 2,400/2,400 correct normal lifecycle reads. Two text-only stateless selectors each reached 2,200/2,400; all 200 misses per selector were rollback. The selectors did not receive the operation log. A baseline given the full log can replay the lifecycle too.
This tests persisted state through update, rollback, repair, and compaction. It does not establish an advantage over a stateful system with equivalent information.
Customer systems remain the source of authority. Updates must reach the configured connection, and dependencies must be recorded. Unconnected changes and unmediated paths are outside this boundary.
FieldHash Ledger
Keep the decision and its evidence inspectable.
The evidence packet records the authority decision and its reasons. Ledger signs a digest of that packet and can append the seal to a verifiable hash chain. The production context handoff requires verification against the configured signer and a durable, verified Ledger append before returning an allow.
These are integrity mechanisms in the documented implementation. They are not an additional behavioral study or proof that the policy, reviewer, or source was correct.
A base hash chain is tamper-evident. Stronger resistance to operator tampering depends on signer pinning, key custody, external anchoring, and deployment configuration. The agent's own account of its behavior is not a substitute for this evidence.
Build the case for governing your workflow.
Choose a workflow where approvals change or reviewed decisions get used again. Over six weeks, compare FieldHash with the controls already in place: which handoffs need different treatment, which approved work can proceed, and what your reviewers can explain from the evidence.
Start in shadow mode, with production enforcement off. Set the acceptance criteria together, then review the observed false-allow rate, review volume, median review time, added latency, deployment effort, and evidence quality. Leave with a recommendation to enforce, adjust, continue in shadow mode, or stop.
If a second workflow is separately scoped, measure whether it can reuse maintained authority and prior review with less setup and reviewer effort. That evaluation establishes whether reuse creates operating savings.
Shadow disagreements identify cases to investigate. Behavior after an enforced denial requires a separately scoped enforcement check. Customer validation, production readiness, and measured savings require customer evidence; the controlled studies cover configured paths, not universal containment.