Every metric should answer a real relationship question.

Communication taxonomy

A proprietary taxonomy spans communication acts, emotion, needs, boundaries, repair, risk, tactics and topics. Labels describe communication, not clinical diagnoses.

Relationship memory

Grey Mirror links repeated episodes, recurring loops and changes across the timeline, so a pattern that returns months later is recognised as the same pattern.

Report depth

The complete report brings together a cinematic explanation, timeline and participant comparisons, metric drivers, supporting evidence, confidence and limitations, plus available action-plan and conditional-forecast sections. Sparse or missing evidence can make a measure unavailable.

Relationship health score

A composite report-level read of stability, repair, reciprocity, timing, and risk signals when enough data exists.

Is the thread moving toward steadier connection or toward more strain?

Combines normalized metric families rather than treating one dramatic message as the whole answer.

Best read as a multi-signal score with supporting evidence windows and confidence labels.

Needs enough timeline coverage, both participant labels, and stable parsed timestamps.

Repair rate

How often conflict or tension is followed by useful de-escalation, accountability, or reconnection.

Do conflicts actually recover, or do they restart with the same unresolved pattern?

Looks for rupture windows, follow-up windows, apology or accountability language, and return-to-baseline signals.

Best with complete exchanges around disagreement and aftermath; offline repair lowers evidence density.

Best with complete exchanges around disagreement and aftermath.

Rupture vs repair

The balance between moments that damage trust and moments that restore it.

Is the relationship absorbing stress or accumulating unresolved damage?

Contrasts tension markers, escalation, blame, invalidation, repair bids, and follow-through windows.

Best when rupture and repair both appear in the text record with clear before, during, and after context.

Needs enough before, during, and after context around conflict.

Response time

Reply latency patterns by participant, time period, and context.

Who tends to wait, who tends to chase, and whether the pace has shifted.

Uses timestamp gaps after normalization and excludes gaps that are not countable.

Best with reliable timestamps, intact sequence, and enough repeated windows to separate ordinary schedule from drift.

Requires reliable timestamps and correct participant mapping.

Timing drift

Whether reply patterns get faster, slower, more selective, or more asymmetric over time.

Did the relationship cool gradually, or did one person suddenly change cadence?

Compares timeline windows and participant-level latency distributions.

Best with long histories and consistent timestamp coverage so slower pacing can be compared across phases.

Best with long histories and consistent timestamp coverage.

Turn-taking

How evenly the conversation passes between participants instead of collapsing into monologues or pursuit.

Does one person carry the thread, explain more, or repeatedly reopen the same issue alone?

Compares message runs, response density, unanswered streaks, and follow-up burden.

Best when topic type, response quality, and repeated back-and-forth are visible.

Needs accurate speaker attribution and enough back-and-forth.

Effort imbalance

Differences in initiation, follow-up, emotional labor, planning, and repair attempts.

Is the thread reciprocal, or is one person maintaining the relationship by default?

Combines initiation, unanswered bids, follow-through, planning, repair, and engagement markers.

Best when ordinary days, conflict days, planning days, and follow-through all appear in the upload.

Best when the uploaded history includes ordinary days, conflict days, and planning days.

Positivity reciprocity

Whether warmth, affirmation, humor, appreciation, and bids for connection are returned or ignored.

Does positive energy circulate, or does it mostly travel one way?

Counts and compares positive bids, reciprocal responses, and ignored warmth across windows.

Best with surrounding message context around affectionate, humorous, sarcastic, or supportive exchanges.

Needs message-level context around affectionate or supportive exchanges.

Power dynamics

Signals of pressure, control, invalidation, apology avoidance, boundary testing, or one-sided decision control.

Does one person repeatedly shape the terms of the relationship while the other adapts?

Looks for repeated pressure patterns, dismissal, coercive wording, threats, blame loops, and boundary response.

Best with repeated pattern evidence, recurrence, intensity, and trust-page resources for safety-related concerns.

Requires repeated pattern evidence, not a single phrase.

Future planning

How often concrete future plans appear, who initiates them, and whether they become action.

Is future talk specific and mutual, or vague and noncommittal?

Separates concrete logistics, time-bound plans, vague someday language, and follow-through.

Best with enough planning conversations, follow-through windows, and platform coverage to compare.

Needs timestamps and enough planning conversations to compare.

Communication shifts

Changes in length, tone, response cadence, conflict style, affection, or planning across the timeline.

What changed, when did it change, and which signals changed together?

Compares windowed metric changes and participant-level trend direction.

Best with multiple months or distinct relationship phases so timeline shifts can be compared.

Best with multiple months or distinct relationship phases.

Emotional momentum

The direction of emotional movement across windows when sentiment, repair, warmth, and tension align.

Is the relationship gaining connection, flattening, or moving toward shutdown?

Uses multiple signal families together instead of one sentiment counter.

Best with enough messages per window and surrounding context for emotional language.

Needs enough messages per window for confidence.

Recurring loops

Repeated relationship episodes that follow similar sequences, such as tension, apology, temporary warmth, and relapse.

Are you solving a problem or replaying the same episode with new words?

Compares episode structure, triggers, sequence order, and aftermath when sequence intelligence is available.

Best with repeated episodes that have clear starts, follow-up, and aftermath windows.

Needs repeated episodes with clear starts and aftermath windows.

Evidence confidence

A label that explains whether a metric is strongly supported, partially supported, or too thin to trust.

How much should the user rely on this particular claim?

Considers message count, window coverage, parser quality, participant confidence, and supporting evidence density.

Low confidence is not a negative finding; it means the evidence needs more support.

Uses parser diagnostics, timeline coverage, evidence references, and metric support.