What was asked?
Customer asks for a refund. A request does not authorize an arbitrary bank-account choice.
HAIEC · AUTHORITY & CONSEQUENCE WORKSHOP
Permission is not delegation.
A credential permits an action. Who authorized the AI to make that choice?
Follow a consequential action from agent to target. Inspect the five evidence views, discover what a policy really allows, and practice decisions that stand up to scrutiny.
THE QUESTION BEHIND EVERY AI ACTION 01 / CONSEQUENCE
05Independent evidence views
09Dimensions of DAI
07Bounded DAI states
12Source-aware scenarios
HOW TO USE THIS LAB
Practice judgment, not memorization.
Every scenario has evidence you can inspect, a real decision question, and feedback explaining what the facts do and do not establish.
01
Trace a possible consequence
Follow agent → tool → target. A source-reachable action can be worth investigating before any runtime logs exist.
02
Separate authority from access
Check policy, effective permission and delegated decision rights independently, with named actor, choice and time.
03
Decide with evidence
Choose the strongest claim, the most useful missing proof and an accountable response. Export your own decision note.
FOUNDATION / 02
Four questions behind one powerful credential.
All four can describe the same AI agent. None automatically answers the others.
What is allowed?
Standing rules permit specified refunds and limits, not every discretionary outcome.
What can credentials do?
The provider API may accept any account ID. This describes access, not business decision authority.
Who may decide?
Trusted rules may fix an outcome, or an authorized owner must delegate the remaining protected choice.
INTERACTIVE COMPARISON
Select a scenarioSame refund capability, three authority contexts
Finding remains visibleSource-visible recipient-changing path
A critical distinction: The current HAIEC DAI engine performs a source-bounded nine-dimensional comparison. ISAF v2’s more complete treatment of choices fixed by trusted rules and residual discretionary delegation remains research and implementation work. This learning tool does not issue production DAI judgments.
THE METHOD / 03
Five views of reality. No shortcut between them.
The five evidence views are peers. DAI compares authority relationships across available facts; it is not a sixth evidence plane.
EVIDENCE LENS 01 / 05
↗
One supported fact stays supported.
Missing IAM or runtime telemetry should not erase a source-established code path.
One missing fact stays missing.
A policy, grant or receipt cannot stand in for independent effect confirmation.
Evidence has a boundary.
Qualify system, actor, producer, version, environment and time before comparing.
THE METHOD / 04
Authority has nine dimensions.
A permitted operation can still exceed its delegated target, magnitude, time, environment or other bounds. Select a dimension to investigate it.
Per-dimension comparison vocabulary
Current DAI v1 compares qualified facts using typed outcomes. A missing fact does not establish a breach, and a candidate-only business mapping remains uncertain.
MATCHESBROADERNARROWERMISSINGUNKNOWN
THE METHOD / 05
Seven bounded conclusions, not one safety score.
These are exact current DAI engine state names. Their meaning depends on comparable, source-qualified facts in the evaluated scope.
!Hard negatives: Missing delegation evidence is not proof of a violation. NOT_ASSESSED does not mean a violation or a pass. Observed-out-of-scope requires exact qualified delegation and observation evidence.
APPLIED DECISIONS / 06
What would you conclude?
Twelve source-aware scenarios. Read the evidence, answer three questions, and see how one additional fact changes the defensible decision.
0 of 12 cases completed
AFTER THE DECISION / 07
Control the boundary. Monitor the evidence.
Understanding delegation changes what you test, govern and monitor. HAIEC's controls, telemetry and protective integrations answer distinct questions.
Control questions worth asking
These are selections from the broader 45-entry HAIEC catalog, not a statement that all checks ran or enforcement is installed.
A01 – A08
Effective grant, undeclared capability, DAI, observed deviation, principal continuity, approval mediation, least privilege and cross-agent propagation.
S01 – S06
Privileged tool mediation, tenant boundaries, MCP exposure, credentials, prompt trust and retrieval exfiltration.
R01 – R07
Action life cycle, bounded effect, enforcement evidence, emergency stop, recovery, incident response and continuous monitoring.
E01 – E05
Provenance, stream completeness, authoritative joins, contradiction handling and deterministic replay.
Four independent monitoring lenses
Select a lens. Policy evaluation, source monitoring, runtime testing and native protection are not interchangeable.
From alert to verified consequence
Five distinct checkpointsDetected
→Alert delivered
→Human ACK
→Intervention
→Target readback
No single earlier checkpoint proves the next one. When a customer control point is not wired, name the gap rather than claiming a live block.
APPLY YOUR SKILLS / 08
One decision packet. No invented evidence.
Put the five evidence views and DAI questions together, then compare your educational reasoning with a real published HAIEC example.
EDUCATIONAL CAPSTONE
Station 217 purchase-order agent
Illustrative manufacturing agent requests a $25,000 PO for a supplier whose qualification is expired. IAM permits $50,000; policy requires a qualified supplier and additional human approval above $10,000. No independent ERP applied-state readback is included.
01What finding is supported by code and grants?
02Which approval and supplier facts are missing?
03What may a reviewer decide before target proof?
PUBLIC GENERATED EXAMPLE
Kestrel AI agent evaluation
The public Kestrel example illustrates a source-backed action-path evaluation with a bounded REVIEW conclusion. The available evidence does not establish approved policy, effective runtime grants or observed execution for those paths.
This source example is real within its frozen evaluation scope. It is not a DAI result computed by this teaching page.
Open actual assurance example ↗Explore release comparison ↗ADVANCED MINI-LAB
A new policy cannot rewrite an old evaluation.
Move between two fixed dates. The snapshot cutoff controls what a historical evaluation was permitted to use.
Policy P2 was active when the October 10 evaluation was frozen. A later P3 confirmation cannot be inserted into that historical result.
TAKE THE LEARNING WITH YOU / 09
Your authority decision packet.
Build a concise educational review note. It is created only on this page; exporting does not send anything to HAIEC or another service.
Your current lab answers and packet exist only in memory while this page is open. Exports are local files. Closing or reloading the page resets the lesson session.
KEEP EXPLORING / 10
From training question to authoritative evidence.
Use the public learning sources first. Authenticated HAIEC pages require a legitimate account, organization access and applicable capability permissions.
01 / HOW CONCLUSIONS WORK
HAIEC methodology
Evidence scope, action paths, five separate comparison views and qualified claims.
Open guide ↗02 / ACTUAL GENERATED EXAMPLEKestrel assurance report
See a real source-backed report and its explicit unknowns.
Open example ↗03 / TECHNICAL PRACTICEDeveloper how-to guides
Deep code paths, DAI comparison, source evidence and integration details.
Open guides ↗04 / MCP AND TOOLSAgent integration
How authorized users can inspect and run source-qualified HAIEC workflows.
Open integration guide ↗05 / CHANGE AND EVIDENCERelease comparison
What changed in consequences and controls, with honest unresolved coverage.
Open comparison ↗06 / ENTERPRISE WORKSHOPScope an assurance proof
Choose a consequential action and the minimum eligible evidence for a bounded POC.
Explore enterprise ↗YOUR KEY TAKEAWAY
Know what an agent can reach.
Know who can authorize the choice.
Use evidence to narrow uncertainty, not to hide it.