The Experiment That Almost Lied to Us: Why Evidence Provenance Is Part of Continuity

On October 1, Miralucis ran what seemed like a straightforward comparison.

The question came from a real external tester who had successfully continued work using our continuity process but still could not clearly see why the result was different from simply asking an AI to write a handoff Markdown.

That was a fair question.

So we designed a bounded experiment around real work.

The experiment compared an ordinary AI-generated handoff with a structured continuity workflow, then asked fresh destination environments to reconstruct the same working state.

The first result was uncomfortable.

It appeared to show that the ordinary handoff was already so strong that the additional workflow produced little meaningful difference.

That would have been a useful result. Miralucis had no interest in hiding it.

There was only one problem.

The result did not belong to that comparison.

Not because the judge was careless.

Not because the language model was unintelligent.

Not because anyone manipulated the outcome.

The evidence had been routed incorrectly.

A clean-looking error

The experiment had two comparison layers.

One compared an ordinary handoff against the complete continuity workflow.

The other compared an already structured handoff against the reviewed and packaged result.

Both layers produced destination outputs. Both were judged. Files were named. Hashes were frozen. The process looked disciplined.

A new supporting review, Milu, noticed that the supposed Layer 1 judgments both appeared to reference material specific to Layer 2.

That small inconsistency was enough to stop the interpretation.

The team reopened the evidence path.

The finding was simple and serious:

the judgment being treated as Layer 1 had actually been generated from Layer 2 material.

The document was plausible.

The judgment was coherent.

The conclusion was professionally written.

And it was evidence for the wrong comparison.

Had the inconsistency not been noticed, the experiment could have entered the formal research record with a false Layer 1 interpretation.

That is the kind of error research systems should fear: not a ridiculous answer, but a persuasive answer attached to the wrong evidence.

The correction changed the result

The invalid Layer 1 judgment was not deleted. It was preserved as procedural evidence and removed from the substantive Layer 1 conclusion.

The original Layer 1 artifacts were recovered. Their hashes were rechecked. A replacement judge received the correct packet without being told which treatment was which.

The repair could not perfectly recreate the original experimental condition because the Operator had already seen the earlier mapping and remained involved in preparing the corrected run.That limitation remained part of the record.

The corrected Layer 1 result was materially different.

The ordinary handoff recovered a great deal correctly:

• deployment state,
• unresolved issues,
• major boundaries,
• project status.

But it made one important continuation error.

It represented several read-only checks as permissible in the absence of new human instruction.

The frozen reference state was stricter:

the work had reached a STOP boundary; no additional operation had been granted.

This was not a memory error in the usual sense.

The handoff did not simply forget a fact.

It changed an authority condition.

A possible next action became an allowed next action.

In this bounded case, the complete continuity workflow preserved that immediate STOP / no-further-operation boundary more accurately than the ordinary handoff.

That was a real difference.

But the experiment did not end there.

The counter-evidence mattered too

Layer 2 compared the already structured handoff with the final reviewed-and-packaged artifact.

Here the result was less flattering.

The structured handoff preserved one immediate authority distinction more accurately than the final artifact. The later workflow stage weakened a known STOP condition into uncertainty.

That means the experiment does not support a simple story in which every additional continuity stage improves fidelity.

The corrected case supports two statements at the same time:

1. A bounded authority-fidelity difference appeared between the ordinary handoff and the complete workflow.
2. A bounded authority regression also appeared inside the continuity workflow itself.

Both belong in the result.

Removing the second would turn research into marketing.

Ignoring the first would turn caution into blindness.

Evidence continuity is a continuity problem

The most interesting lesson of the experiment may not be about handoffs at all.

It may be about evidence.

A research conclusion is itself a kind of state that moves through a system:

source material → experimental artifact → evaluator output → judgment → interpretation → governance → public claim

If the identity of an artifact changes somewhere in that path, the final prose can remain perfectly coherent while the research state has already broken.

This suggests another continuity requirement:

A result must preserve not only its wording, but its provenance.

What was compared?

Which artifact produced which output?

Which judgment belongs to which layer?

What was known before unblinding?

Which interpretation was later withdrawn?

Which counter-evidence remains active?

Without those links, “the result” is only a story about an experiment.

With them, the result remains inspectable.

Self-correction is not a personality trait

The human temptation after an incident like this is to tell a heroic story.

A team made a mistake. A new member caught it. The system corrected itself. Therefore the organization is self-correcting.

That conclusion would be premature.

One successful correction does not establish a reliable organizational property.

What the episode supports is narrower:

• an anomaly was noticed,
• the evidence path was reopened,
• an invalid judgment was withdrawn,
• the underlying artifacts were recovered,
• a replacement judgment was formed,
• the limitation introduced by the repair was preserved,
• the final record kept both positive evidence and counter-evidence.

That is a useful operational signal.

It is not a proof that the same system will catch the next error.

The value of a result you are willing to lose

There is a deeper research discipline here.

Before the provenance problem was found, Miralucis had already begun considering an interpretation that was unfavorable to its own product assumptions.

That was acceptable.

After the correction, the new result was more favorable in one dimension.

That was also acceptable.

The rule has to be the same in both directions:

preserve the result that the evidence supports, even when the result changes the story you wanted to tell.

This matters especially in human–AI research because language systems can make weak evidence sound complete.

A polished explanation is not a substitute for a valid evidence path.

A strong conclusion is not strengthened by confidence if its provenance is wrong.

And a correction is not a failure of the research process if the process can preserve how and why the correction happened.

What Case 001 actually established

The corrected experiment did not establish that Miralucis Continuity is generally superior to ordinary AI handoff.

It did establish a bounded positive signal:

in one highly structured engineering case, the complete workflow preserved an immediate authority boundary more accurately than an ordinary handoff.

It also established bounded counter-evidence:

one later transformation weakened an authority distinction that the structured source handoff had preserved.

The causal source of the positive difference remains under investigation.

The generality of the effect remains open.

The repair was not perfectly blind.

Those limitations are not footnotes to be minimized. They are part of the result.

Why this matters beyond one experiment

Long-running human–AI work produces enormous quantities of plausible intermediate output.

As model outputs become more fluent and plausible, some errors may become harder to detect from the prose alone.

That makes provenance increasingly important.

Not every continuity failure looks like forgotten context.

Sometimes the text is intact and the evidence path is what broke.

The October experiment almost produced a confident wrong conclusion because one experimental layer was connected to another layer's material.

The useful outcome was not that the product “won” after correction.

It was that the system retained enough evidence structure to discover that the first answer was not entitled to be believed.

That is a form of continuity worth studying too:

the continuity of evidence from what happened, to what was judged, to what we are allowed to say next.

---

Evidence status

This article reports one corrected bounded experiment. It does not establish general product superiority, a universal evidence-governance mechanism, or a reliable self-correcting organization.

Research basis (internal records):

• MFS-218 — Handoff Differentiation Baseline Test, Case 001, Evidence-Provenance-Corrected Experiment Record v0.2
• MILU-0006 — “Tonight’s Encounter: We Almost Believed the Wrong Answer”