Systems Atlas/ AI, Work & Organizations

FIELD-01

Information × Incentives in Recurring Exception Resolution

The first proposed direct test of the v2.1 configuration architecture.

Version
v2.1
State date
2026-09-27
Research stage
PRIVATE OPERATIONAL DISCOVERY

Central question

When an AI-enabled recurring exception-resolution workflow stalls, is the causal limitation information, incentives, their interaction, or neither?

Before reading the design

No experiment is running and no result exists

Research object

Why recurring exception-resolution workflows?

A recurring exception-resolution workflow in which verified factual state and actor incentives can plausibly be manipulated separately.

A useful host can provide discrete recurring cases, auditable states, exceptions, real decision rights, measurable waiting and rework, and a downstream outcome that can be distinguished from mere process completion. Those properties must be verified in the eventual setting.

Conceptual design

Information × incentives

Conceptual 2 × 2. This is not the final randomization architecture.
Information conditionExisting incentiveAligned incentive
Existing informationControlIncentive only
Verified factual informationInformation onlyInformation + incentive

Treatment orthogonality

The factors only mean something if they remain distinct

Information treatment

Change factual state, not authority

  • changes verified factual state
  • does not add recommendation or approval guidance
  • does not change formal authority or practical control
  • does not change accountability or economic payoff

Incentive treatment

Change local payoff, not factual state

  • changes expected local benefit or effort cost
  • does not change factual information
  • does not create a new decision right or approval authority
  • does not silently change liability or accountability architecture

Mechanism check

A null result is interpretable only if the intended mechanism changed

1Treatment assigned
2Did the intended mechanism actually change?

YesInterpret the outcome against the preregistered prediction.

NoA null outcome does not falsify the mechanism.

Hard problem 01

Process completion ≠ true organizational outcome

A ticket can close, an ERP submission can start, or a workflow step can complete without the intended accepted downstream state actually occurring. FIELD-01 requires case-level linkage to the real outcome.

Hard problem 02

High case volume ≠ high independent experimental N

Thousands of cases handled by a few persistent workers, teams, or sites can still provide little independent leverage. The treatment level and interference boundary must match the mechanism.

Research stage

Where FIELD-01 actually is

Public status

PROPOSED / PRE-OPERATIONAL

Current phase

PRIVATE OPERATIONAL DISCOVERY
  1. ConceptCOMPLETE
  2. Host archetype screeningCOMPLETE
  3. Public screeningCOMPLETE

    Complete enough; broad candidate expansion is stopped.

  4. Private discoveryCURRENT
  5. Host selectionNOT YET
  6. ProtocolNOT YET
  7. PreregistrationNOT YET
  8. ExecutionNOT YET
  9. ResultNONE

Valid outcomes

The study is not designed to prove v2.1

informationincentiveinteractionneither

“Neither” is a valid result. If manipulation checks pass but neither treatment nor their interaction moves the preregistered outcome, the relevant binding-constraint claims should be weakened rather than repaired after the fact with a new bottleneck story.

Current milestone

First private host evidence

First private host evidence: resolve true-outcome linkage and independent treatment units for at least one eligible host × workflow configuration.

Workflow-specific falsifier

For the focal workflow, the state/incentive binding-constraint claims are materially weakened if both manipulation checks pass yet neither treatment nor their interaction moves the preregistered primary outcome by the minimum meaningful amount, robust to the preregistered interference and case-mix analyses.