Delta Analysis

Every comparison produces a set of canonical deltas. Each delta fires only when the difference is statistically meaningful — no noise, no false signals, no manual threshold tuning.

Five canonical delta types

Deltas are ordered by causal salience, not mathematical complexity:

Delta	Full Name	Category	What It Measures	Fires When
ΔF	Failure Rate	Event	Anomaly frequency	Failure frequency or kind differs between runs
ΔTc	Convergence Time	Timing	Steps to reach stable latency	Steady-state reached at different steps (3+ step separation)
ΔTd	Total Duration	Timing	Wall-clock time / structural emergence	Dominance onset differs (suppressed in TFRT preset)
ΔĀ	Average Latency	Behavior	Mean metric value	Mean differs meaningfully (suppressed in TFRT preset)
ΔO	Output Variability	Behavior	Oscillation / runtime instability	Area-above-threshold score differs beyond noise floor

Each delta has three possible statuses:

• Present — the difference is real and meaningful
• Suppressed — the difference is below threshold or irrelevant for the active preset
• Indeterminate — cannot determine (insufficient data)

How deltas are computed

Each delta type has its own detector with configurable thresholds. Every delta includes:

• Confidence score (0.0 to 1.0) — how certain the difference is meaningful
• Anchors — specific data points and view targets that triggered the delta
• Trigger type — what kind of signal caused the detection (e.g., sustained, recurrence, area episode, persistence-weighted)
• Human-readable explanation — auto-generated text describing the finding (target: 12 words or fewer)
• Summary sentence — neutral, descriptive sentence for export summaries

Convergence detection (ΔTc)

The convergence detector looks for when a signal stays within an epsilon band for a sustained window:

• Window: number of consecutive stable steps required (default: 5, minimum: 3)
• Epsilon: base tolerance band; effective epsilon = max(base epsilon, 0.5 * robust sigma)
• Resolution: minimum 3-step separation between runs to count as meaningful
• Confidence is a heuristic based on tail stability and noise level — it affects visual intensity but never suppresses the delta

Stability detection (ΔO)

The stability detector uses area-above-threshold scoring with an adaptive threshold:

• Threshold adapts based on both median and sigma of curvature magnitudes
• Episodes must sustain for at least 4 steps to avoid flicker false positives
• Between-run suppression applies a delta floor of 0.05
• Within-run noise floor of 0.1 filters out trivial episodes

Failure detection (ΔF)

Detects anomalies using a persistence window (default: 3 steps). Can trigger on norm violations, loss explosions, or other failure kinds. Reports which run failed, at what step, and what kind of failure occurred.

Emergence detection (ΔTd)

Detects structural emergence through eigenvalue dominance. Fires when one eigenvalue exceeds k times the next (default k = 1.5) for a sustained window, or through a recurrence rule (repeated dominance segments within a rolling window).

Runtime presets

TFRT (TensorFlow-TRT)

The built-in TensorFlow-TRT preset (tensorflowrt-runtime-v1) is designed for inference comparison:

• Maps inference-specific signals: latency, throughput, memory, CPU/GPU load
• Suppresses ΔĀ and ΔTd — these have no meaning for inference workloads
• Active deltas use latency as the primary signal (ΔTc for stabilization time, ΔO for oscillation, ΔF for outliers)

TFRT guardrails

The preset raises warnings when:

• No steady-state milestone is found in the trace
• Warmup exceeds 50% of the run duration
• Only aggregated stats are available (disables time-based deltas entirely)

These guardrails appear as inline warnings in the Compare tab so you know when results may be limited.