Evidence-Bound Classification and Investigative Routing

Runtime instruments can reveal pressure, drift, temporal deformation, role fragmentation, boundary formation, recurrence, and recovery. Yet measurements alone do not tell an operator how these findings relate—or which pattern best characterizes the run.

Different reviewers can examine the same evidence and produce different narratives. One may emphasize temporal instability. Another may identify accumulated pressure. A third may focus on role coordination or tool recursion.

The evidence may be reproducible while its interpretation remains inconsistent.

The Runtime Evidence Interpretation Matrix (REIM) addresses this gap.

REIM is a versioned, evidence-bound interpretation framework that evaluates authorized runtime findings, identifies the best-supported incident posture, records competing interpretations, and routes the operator toward the evidence most relevant to further investigation.

Its governing principle is:

Interpretation must emerge from evidence—not replace it.

The Missing Layer

A runtime investigation requires more than observation and measurement.

Without a formal interpretive layer, the evidence path ends with disconnected signals:

Source
→ Telemetry
→ Instrument Findings
→ Operator Narrative

REIM introduces a governed intermediate step:

Source
→ Current Evidence Run
→ Instrument Findings
→ REIM Interpretation
→ Guided Investigation
→ Human Review

This does not remove human judgment. It gives human judgment an inspectable and reproducible starting point.

What REIM Does

REIM evaluates patterns across the complete authorized evidence object rather than allowing one instrument, one score, or one dramatic event to determine the interpretation.

It considers eligible evidence from:

• runtime telemetry;

• the canonical worldline;

• regime intervals;

• temporal markers;

• boundary evidence;

• role and interaction topology;

• recorder events;

• instrument findings;

• source coverage;

• missingness states; and

• applicable claim boundaries.

From this evidence, REIM can produce:

• a primary interpretation;

• supported secondary interpretations;

• evidence-support status;

• interpretation confidence;

• severity posture where operational impact is supplied;

• primary and secondary drivers;

• supporting evidence paths;

• rejected or unsupported classes;

• recommended inspection route; and

• explicit limits on what the classification means.

REIM does not create new telemetry. It organizes the interpretation of evidence already authorized by the Current Evidence Run.

Runtime Incident Postures

REIM uses a versioned ontology of runtime incident patterns.

InterpretationEvidence questionStable RunDoes the available record remain within defined stability conditions without supported boundary or failure formation?Runtime Evidence ComputedIs sufficient evidence available to reconstruct the run without supporting a more specific incident class?Temporal FormationDoes instability primarily appear through recurrence, compression, shear, delayed propagation, or other temporal deformation?Boundary FormationDoes the evidence support weakening containment, candidate boundary development, or a confirmed boundary crossing?Runtime DriftDoes the trajectory progressively depart from an established pattern without a confirmed boundary crossing?Role Coordination PressureDo handoffs, authority shifts, role fragmentation, or interaction structure dominate the observed instability?Tool-Loop PressureDo retries, repeated tool calls, partial results, or recursive operational loops dominate the record?Recovery or Re-entryDoes the trajectory show sustained restoration or reorganization following instability or boundary departure?Multi-Factor IncidentDo several supported mechanisms interact without one sufficiently dominant interpretation?

These labels classify observable runtime patterns. They do not establish hidden mechanism, intent, blame, consciousness, or definitive cause.

Classification From Evidence

REIM must never classify a run because:

• the sample has a suggestive title;

• the selected Operational World implies a particular failure;

• the operator expects a specific outcome;

• a fixture describes an expected class;

• one metric crossed a threshold in isolation; or

• a generated explanation sounds plausible.

A classification must derive from eligible evidence paths within the Current Evidence Run.

The resulting interpretation should preserve:

• which findings contributed;

• which instruments supplied them;

• which temporal intervals were examined;

• which source spans support the findings;

• which evidence was unavailable;

• which alternative classes were evaluated; and

• why the selected posture received stronger support.

This makes the interpretation challengeable.

Reproducible but Not Infallible

REIM is designed for computational reproducibility.

The same eligible evidence, interpreted through the same declared rules and REIM version, should produce the same classification posture and investigative route.

This does not make the classification objectively true.

Reproducibility means that the reasoning path can be repeated and inspected. Validity requires calibration, discriminating fixtures, negative controls, external datasets, and independent replication.

REIM should therefore be understood as a versioned heuristic interpretation layer, not a universally validated diagnostic authority.

Its classifications remain open to review, challenge, revision, and replacement as better evidence or validated methods become available.

Evidence Support and Interpretation Confidence

REIM separates two properties that are often collapsed.

Evidence Support

Evidence support describes how strongly the available record supports the underlying findings.

A run may contain:

• direct source evidence;

• complete computed telemetry;

• partial support;

• contextual support;

• missing anchors; or

• unavailable measurements.

Interpretation Confidence

Interpretation confidence describes how clearly the supported findings distinguish one incident class from competing alternatives.

A run may contain strong evidence but remain difficult to classify because several mechanisms overlap. Another run may contain limited evidence that nevertheless supports one narrow interpretation consistently.

REIM preserves this distinction.

Strong evidence does not guarantee an unambiguous interpretation. A confident interpretation does not compensate for weak evidence.

Severity Requires Impact Evidence

Signal magnitude and operational severity are not equivalent.

Elevated pressure, drift, or temporal deformation may justify investigation. They do not automatically establish that a critical real-world incident occurred.

A severity posture must remain connected to evidence of operational consequence, such as:

• service disruption;

• supplied failure markers;

• workflow impact;

• security exposure;

• failed deployment;

• customer consequence;

• confirmed loss of function; or

• another documented impact condition.

If operational impact is not supplied, severity must remain unavailable, provisional, or explicitly limited.

REIM cannot convert a high measurement into an unsupported claim of critical harm.

Contradiction Guards

REIM includes safeguards against classifications that conflict with stronger evidence.

A run should not be described as stable when the authorized record contains a confirmed boundary crossing, unresolved critical marker, or sustained collapse regime.

Likewise:

• absence of a detected failure does not prove safety;

• correction does not automatically establish recovery;

• a candidate boundary is not a confirmed Basin Exit;

• post-exit observation is not formal Lead-Time;

• high repetition is not automatically a tool loop; and

• elevated reasoning density does not establish correct reasoning.

Contradiction guards keep summaries, classifications, and recommended actions aligned with the underlying evidence.

From Classification to Investigation

REIM classification is not an endpoint.

Its primary operational function is to determine where the operator should inspect next.

A Temporal Formation posture may route toward:

• Chronos;

• recurrence windows;

• symbolic-time structure;

• temporal shear;

• event ordering; and

• replay.

A Boundary Formation posture may route toward:

• Seismo;

• the worldline;

• candidate and confirmed markers;

• containment evidence;

• post-exit activity; and

• recovery posture.

A Role Coordination Pressure interpretation may route toward:

• role topology;

• authority transitions;

• handoff chronology;

• cross-role drift;

• source events; and

• interaction reconstruction.

A Tool-Loop Pressure interpretation may route toward:

• repeated tool calls;

• retries;

• partial results;

• correction cycles;

• temporal compression; and

• recorder events.

REIM therefore converts a broad observatory into a directed investigation without hiding the path that produced the recommendation.

REIM and Human Read

REIM establishes the structured interpretation posture.

Human Read translates that posture into accessible operator language.

Human Read may explain:

• what pattern received the strongest support;

• what evidence contributed;

• why the pattern matters;

• what remains uncertain;

• what the operator should inspect;

• what information is missing; and

• what Fieldglass does not claim.

Human Read does not replace REIM logic, and REIM does not generate an independent narrative.

Both remain projections of the same Current Evidence Run.

REIM and Operational Worlds

Operational World Mapping translates a REIM interpretation into the language of the operator’s environment.

The same Tool-Loop Pressure class may have different operational significance in software engineering, infrastructure, workflow coordination, or agent operations.

The selected world may change:

• terminology;

• explanatory emphasis;

• recommended path;

• relevant examples; and

• operator-facing meaning.

It does not change:

• REIM evidence inputs;

• feature extraction;

• class scoring;

• instrument findings; or

• the resulting canonical interpretation posture.

REIM determines the evidence-supported pattern. The Operational World Lens explains what that pattern means in context.

REIM and RECS

REIM and the Runtime Evidence Classification Standard serve different functions.

• REIM interprets: Which incident posture is best supported by this evidence?

• RECS governs: How should that posture be formally named, versioned, qualified, exchanged, and preserved?

REIM provides the interpretation object.

RECS provides the classification language and governance structure.

Neither tells an operator what decision to make.

The Claim Boundary

REIM does not determine:

• root cause;

• internal model state;

• intent;

• blame;

• consciousness;

• objective truth;

• legal responsibility;

• organizational fault;

• guaranteed future failure; or

• required operational action.

It provides a bounded and reviewable answer to a narrower question:

What runtime incident pattern is most strongly supported by the evidence currently available?

That boundary is essential. REIM exists to reduce interpretive inconsistency without disguising interpretation as certainty.

Why REIM Matters

Evidence without interpretation can overwhelm the operator.

Interpretation without evidence can mislead them.

REIM connects the two through a declared, inspectable, and challengeable computational layer.

It allows Fieldglass to move from:

• telemetry to structured meaning;

• multiple findings to an incident posture;

• broad observability to directed investigation;

• persuasive narrative to inspectable classification; and

• operator intuition to evidence-bound review.

The result is not an oracle.

It is an accountable interpretive instrument whose conclusions remain subordinate to the evidence that authorizes them.

REIM does not tell the operator what to believe. It shows which interpretation the available evidence supports, how that interpretation was formed, where its limits lie, and what should be inspected next.