Evidence & Methods
How observation becomes a public claim
Miralucis studies a research system from inside the system as it continues to operate.
That creates unusual access to longitudinal evidence — and unusual methodological responsibilities.
The working record includes contemporaneous conversations, handwritten material, research documents, engineering files, code, experiment logs, deployment traces, failures, corrections, migrations, publications, and later source-linked reconstruction.
Different questions require different forms of evidence.
The central rule is:
Say fully what the evidence supports.
Do not claim what the evidence does not support.
From Source to Claim
Miralucis separates four stages that are easy to collapse:
SOURCE
→ OBSERVATION
→ CANDIDATE INTERPRETATION
→ PUBLIC CLAIM
They are related.
They are not equivalent.
Source
A source is the material from which an observation can be examined.
Depending on the research question, sources may include:
- contemporaneous conversations;
- handwritten notes;
- engineering documents;
- source files;
- experiment records;
- system logs;
- deployment results;
- preserved artifacts;
- publications;
- dated research notes;
- participant recollection;
- or later source-linked reconstruction, which remains secondary to the records it cites.
A source being preserved does not by itself establish the interpretation later attached to it.
Its value depends on provenance, context, completeness, timing, and relevance to the question being asked.
Observation
An observation states what the available material supports at the level being examined.
Examples include:
- a handoff omitted material (in P001, 31 of 97 material items under one structured workflow);
- an authority statement changed (in P001, a STOP boundary became permission for some read-only checks);
- an AI research role behaved differently across two periods;
- a particular term appeared in a dated record (research-system formation, in notes from June 7, 2026);
- a research loop involved several differentiated contributors (TR → Jia Shen → CA → Jia Shen → TR, June 11–12, 2026);
- an implementation produced a recorded result.
The aim at this stage is to describe what happened before deciding why it happened.
Candidate Interpretation
An interpretation connects observations into a possible explanation.
For example:
a changed working pattern may indicate role drift;
a newly available term may have helped participants recognize an older structure;
a preserved external environment may have contributed to later continuity;
or repeated changes across human and AI participants may suggest reciprocal adaptation.
These interpretations can be scientifically useful before they are established.
Their status remains visible while competing explanations are examined.
Public Claim
A public claim is a statement that has enough evidence and qualification to be stated externally at a defined level
The claim should not exceed the evidence that supports it.
Miralucis therefore distinguishes:
what was recorded
what was observed
what was interpreted
what was tested
what was published
and what remains open
Evidence Status
Public pages use a small set of evidence-status labels.
Published
A finding or argument has entered a stable public scholarly record.
Its publication date, version, and source can be identified.
Later research may challenge or revise the interpretation, but it does not silently change the earlier record.
Documented case
A specific event or pattern is traceable in the research record and can be studied as a case.
The label does not imply that the case establishes a general law.
Under reconstruction
Relevant historical material exists, but chronology, provenance, comparison, coding, or interpretation is still being rebuilt.
This is common when present research questions return to material created months earlier for a different purpose.
Hypothesis
A candidate explanation has become explicit enough to operationalize, compare against alternatives, or test prospectively.
A hypothesis is a research object to be tested, not a conclusion already reached.
How We Know
No single evidence type carries the entire argument.
The method is to connect different forms of evidence while preserving their provenance.
Historical Reconstruction
Historical reconstruction asks:
What happened?
When did it happen?
What did participants understand at the time?
What language existed then?
What changed later?
It traces formation across contemporaneous records rather than treating a later summary as the original event.
A current interpretation may reveal something important about the past.
It does not retroactively become what participants believed at the time.
Miralucis therefore distinguishes:
formation date
≠ recognition date
≠ naming date
≠ later research qualification
This distinction is central to the study of Historical Recursive Formation and Research Language.
Participant Evidence
Jia Shen is both the Human Operator and a longitudinal participant-observer inside Miralucis.
Her observations can provide information that external records alone may not contain:
- when a change first became noticeable;
- why an event was treated as significant;
- what real-world constraint affected a decision;
- which historical context made an anomaly recognizable;
- and how the collaboration was experienced while it was happening.
Participant evidence is preserved as participant evidence.
Where possible, it is compared with contemporaneous conversations, files, engineering records, later outcomes, and competing explanations.
Engineering Evidence
Some questions can be examined through artifacts created by the systems themselves.
Engineering evidence includes:
- code;
- files;
- package identities;
- logs;
- hashes;
- system behavior;
- test results;
- deployment traces;
- state records;
- and implementation outputs.
This evidence is especially useful when a research question can be translated into an observable technical event.
It also allows failures to be located at different stages rather than reduced to a single success or failure outcome.
Bounded Experiments
Miralucis uses bounded experiments when a question can be isolated enough to test directly.
A bounded experiment defines:
- the research question;
- the source object;
- the compared conditions;
- the evidence available to each condition;
- the evaluation procedure;
- the relevant metrics;
- stop conditions;
- and the claims the experiment is able to support.
The purpose of bounding an experiment is not to make the question smaller than it is.
It is to make clear what the result actually tested.
P001 is one example of this approach.
Cross-Model Challenge
Miralucis uses multiple AI systems for bounded forms of independent challenge.
External systems including Claude, Gemini, and others may be used for:
- competing explanations;
- literature and prior-art checking;
- adversarial criticism;
- evidence challenge;
- technical review;
- and manuscript review.
Agreement between models is not treated as a vote on truth.
Different systems may share sources, assumptions, or failure modes.
Their value comes from making it easier to expose weak reasoning, overlooked alternatives, unsupported novelty claims, or evidence that deserves further investigation.
Prospective Testing
Historical evidence can suggest patterns.
Prospective testing asks whether those patterns survive contact with future data.
The intended progression is:
historical reconstruction
→ measurement
→ competing explanations
→ prediction
→ prospective testing
→ independent replication
This is especially important for Miralucis's current work on human-side capability formation, AI-role recovery, and Human–AI Co-Evolution Formalization. For these programs, prospective testing is a planned stage, not a completed one.
Mathematics and formal models are used to make claims more testable.
They do not replace the underlying evidence.
Provenance Matters
A claim is stronger when the path back to its evidence remains visible.
Miralucis therefore tries to preserve:
where a record came from;
when it was created;
which version was used;
what was later added;
what remains missing;
and which interpretation belongs to which stage.
This matters because long-running research accumulates copies, summaries, migrations, revisions, and retrospective explanations.
A later document may be clearer than an earlier one.
It is not automatically a better source for what happened earlier.
Historical Is Not Current
Longitudinal research constantly encounters information that was once valid.
A historical instruction may have been authorized.
A research interpretation may have been accepted.
A system state may have been current.
A role may have held a particular responsibility.
Later evidence can change any of these.
Miralucis therefore separates:
historical state
from
current applicability
The historical record remains evidence of what existed then.
Current decisions require current qualification.
This distinction is also central to Miralucis Continuity.
Local Evidence and Larger Claims
A bounded observation supports a bounded conclusion.
A result from one case, one experimental source, one historical shard, one model configuration, or one time window does not automatically establish the state of the whole research system.
Miralucis learned this directly during historical archaeology, when a deliberately non-representative local sample could not answer a global question about the full corpus.
The working principle is:
Local evidence should not become a global verdict without an appropriate evidentiary bridge.
Broader questions therefore require broader mapping, comparison, replication, or explicit limits on the claim.
Failures Are Evidence
Miralucis preserves failed experiments, incorrect interpretations, contradictory findings, abandoned paths, and corrections.
A failure may show:
- that a protocol was inadequate;
- that a variable was not isolated;
- that a historical interpretation was too strong;
- that a tool did not behave as expected;
- that evidence was missing;
- or that the research question itself needs to change.
Negative evidence is not automatically decisive.
But it remains part of the research record.
A later successful result does not erase the earlier failure.
Correction Without Rewriting History
Miralucis distinguishes correction from replacement.
When an earlier interpretation changes, the preferred pattern is:
preserve the earlier record
→ identify what changed
→ add the later evidence
→ state the revised interpretation
This principle applies to research notes, historical reconstruction, engineering records, and publications.
The goal is to make intellectual change inspectable.
Preserve what was actually claimed at the time.
Add what was learned later.
Do not make the past say something new.
Competing Explanations
Miralucis does not assume that an observed change was caused by the collaboration merely because it occurred during the collaboration.
Depending on the question, competing explanations may include:
- ordinary human learning;
- accumulated project experience;
- model upgrades;
- prompt differences;
- improved tools;
- context availability;
- retrieval effects;
- changes in task difficulty;
- environmental changes;
- sampling;
- retrospective reinterpretation;
- or ordinary response variability.
A stronger research claim requires these alternatives to be separated where possible.
Where they cannot yet be separated, they remain part of the stated uncertainty.
Studying a System From Inside
Miralucis has points of contact with several established research traditions.
Its use of a researcher participating within the setting being studied relates to participant observation.
Its interest in problem-solving distributed across humans, AI systems, records, tools, and working environments connects with questions studied in distributed cognition.
Its attention to the observer as part of the system being observed has a point of contact with second-order cybernetics.
These are scholarly coordinates.
They do not define Miralucis as simply an instance or extension of any one tradition.
The specific research object is the longitudinal formation of a working human–AI research system.
What “AI Role Development” Means Here
Miralucis studies AI roles longitudinally.
A role can accumulate:
- responsibilities;
- historical materials;
- protocols;
- working patterns;
- tools;
- evidence;
- relationships to other roles;
- and recognizable professional contributions.
Its later behavior can then be compared with earlier behavior.
When Miralucis describes a role as developing, drifting, contracting, or recovering, the object being studied is this observable working configuration and its performance across time.
The claim is about the research role as it operates within the collaborative system.
From Archive to Discovery
The archive preserves where the research came from.
The research program asks what can now be learned from it.
Observe what happened.
Recover how it formed.
Separate fact from interpretation.
Form competing explanations.
Test what can be tested.
Preserve failures and corrections.
Let new evidence change the model.
Its value lies in making change visible.
The archive allows Miralucis to study not only what the system knows now, but how its questions, methods, roles, capabilities, and mistakes developed over time.
Evidence Is Part of the Result
For Miralucis, a research result is not only a conclusion.
Its status, source, provenance, limits, corrections, and unresolved alternatives are part of what the result means.
That is why the public site distinguishes Published, Documented case, Under reconstruction, and Hypothesis.
The purpose is not to weaken a finding.
It is to make clear what kind of finding it is.