In 17 real Harness scenarios, a naïve ‘done’ label was wrong 10 times. An agent can finish a command, print a success message, and still leave the intended target unchanged. These two bounded studies ask what evidence is needed before a system may describe work as having had an effect.
10
wrong terminal labelsacross 17 real Harness scenarios
4/9
stdout / exit codeeach correctly indicated the observed state in 4 of 9 operations
9/9
Effect contract + readbackthe pair reached 9 of 9 in the selected operations
The denominators come from separate studies and are not pooled.
THE CENTRAL FINDING
An executor can report its attempt. Only the relevant target can support a claim that its state changed.
TWO BOUNDED STUDIES
The terminal knew less than it said.
One study checks whether state labels told the truth. The other checks whether common process signals could establish a target effect. Their samples, methods, and denominators are separate.
R5.217 real Harness scenarios
A naive terminal label was wrong in 10 of 17 cases.
The comparison started with the label a terminal would naturally emit: done, passed, or successful. It then reconstructed what happened from the authority source for each scenario. The two views disagreed in 10 cases. In this selected set, authority-source reconstruction matched the recorded determination in all 17 cases.
10wrong labels
7correct labels
The baseline used one coarse terminal state. Reconstruction separated Attempt, Effect, Adoption, and Acceptance, then returned to the relevant authority source. For this selected set, reconstruction agreed with the recorded authority determination in all 17 cases.
R59 selected real software operations
A zero exit code is not a target fact.
For each of the nine operations, stdout correctly indicated the observed state in 4 of 9 cases, and the outer exit status did so in 4 of 9 cases. Fresh target-native readback, paired with a minimal Effect contract, reached 9 of 9 in this selected set.
stdout / exit code
4/9
each correctly indicated the observed state in 4 of 9 cases
minimal Effect contract + target-native readback
9/9
reached 9 of 9 in this selected set
An outer exit code of zero can only mean that a carrier completed a transfer. A worker that wrote bytes can only mean that an intermediate artefact appeared. The target repository, service, or state still has to answer for itself.
HOW WE CHECKED
Leave the log and walk to the target.
The method is simple in principle: decide who owns the relevant fact, then make the conclusion travel through the evidence relationship rather than a convenient status field.
01
Freeze the scenario
Keep the operation, version, environment, intended outcome, and authority surface fixed.
02
Record the naive judgment
Use ordinary baseline signals such as terminal labels, stdout, and exit code.
03
Locate the fact owner
A target repository, service, file system, or responsible party owns a different fact.
04
Perform fresh readback
Observe the target after the attempt without reusing the producer's acknowledgement.
05
Reconstruct the semantics
Compare Attempt, Effect, Adoption, and Acceptance with their corresponding authorities.
DESIGN CONSEQUENCE
Make the stages visible.
Each stage carries a different evidence responsibility. Combining them into one success field hides the handoff.
01
evidence stage
Attempt
Was the action actually initiated?
A request, dispatch, process start, or message send only shows that an attempt entered an execution path.
02
evidence stage
Effect
Did the authoritative target change?
A minimal Effect contract names what to observe. The target then supports the claim through its own readback, not a caller declaration.
03
evidence stage
Adoption
Did the change enter the intended workflow?
A written file is not necessarily used by a service, and an existing commit is not necessarily on the working path.
04
evidence stage
Acceptance
Did the responsible party accept this exact result?
Acceptance binds the current authority, original objective, exact version, and obligations to an explicit decision. A producer cannot sign for an acceptor.
DESIGN CONSEQUENCE
Effect cannot be self-certified by the action producer. Adoption cannot be inferred from an artefact. Acceptance cannot be inferred from a passing test.
A minimal Effect contract makes a target-native check concrete. It does not turn Effect into business completion, legal sufficiency, or human acceptance. Those decisions retain their own authorities and evidence.
Product-team check: do not show one green tick. Label Attempt, Effect, Adoption, and Acceptance separately, and link each label to its request, target-readback, consumption, or authorized-acceptance evidence.
A COMPOSABLE RESEARCH DIRECTION
Evidence continuity can travel with the work.
Flowness
Keep work facts governed within a working environment.
Flowness explores a Harness for work, evidence, and drift across steps. Its direction is to keep a writer's report from replacing the target fact.
Closer to work execution and collaboration Harnesses
shared ruleA producer cannot prove reality for its consumer.
ToWow
Keep refusal and acceptance across independent authorities.
ToWow explores how agents coordinate across authorities without flattening the distinction between a request, an observed effect, and a responsible party's acceptance.
Closer to cross-authority relationship research
Flowness work governance and ToWow cross-authority research may be composed in future work. This is a composable research direction, not a deployed ToWow and Flowness integration.
CLAIM BOUNDARY
What these results do and do not establish.
Within the studies
Naive terminal labels were wrong in 10 of 17 real Harness scenarios.
Authority-source reconstruction agreed with the recorded determination in all 17 cases.
In nine selected operations, stdout and outer exit status each correctly indicated the observed state in 4 cases.
A minimal Effect contract plus target-native readback reached 9 of 9 in the selected set.
Outside the studies
They do not estimate a general failure rate for agent work.
They do not prove business completion, legal sufficiency, or human acceptance.
They do not establish real participant authorization, production security, long-term value, or a complete product episode.
They do not demonstrate a deployed ToWow and Flowness integration.
The quieter conclusion is useful too: when one authority can safely close a simple task, use the shorter path. ToWow does not need to appear in every call.
SOURCES AND RECORDS
Trace the claim back to its record.
The figures on this page come from real engineering records rather than a synthetic benchmark or public user telemetry. The records include scenario notes, operation observations, target-native readback, and reconstruction tables.
R5
Target-native Effect readback
Nine selected real software operations with a minimal Effect contract and fresh authority readback.
R5.2
Terminal-label reconstruction
Seventeen real Harness scenarios compared with reconstruction from the authority source.
CURRENT
Current research synthesis
The current synthesis keeps the incomplete product episode and real-world limits in scope.
Archive status: the underlying R5 and R5.2 records remain versioned internal material, not public downloads. This page is a bounded research report, not an independent reproduction package or production validation.