Copyright © 2026 Saccadic LLC · Patent pending.

Browse chapters
  1. Release NotesChanges by version
  2. Chapter 1What IRIS Is and Is Not
  3. Chapter 2The IRIS Data Model
  4. Chapter 3How IRIS Measures
  5. Chapter 4Safe Interpretation and the First 15 Minutes
  6. Chapter 5Dashboard Overview
  7. Chapter 6Installation
  8. Chapter 7Configuration
  9. Chapter 8Build the Model Spec and Probe Set
  10. Chapter 9Minimal Tracker Integration
  11. Chapter 10Launch the Bench and Verify Capture
  12. Chapter 11Runtime Control, Signals, Metrics, and Logs
  13. Chapter 12Integration Patterns by Model Family
  14. Chapter 13Distributed Training
  15. Chapter 14Artifacts, Reports, Export, and Security
  16. Chapter 15Troubleshooting
  17. Chapter 16Whole Model and Layer Structure
  18. Chapter 17Domain Training and Affinity
  19. Chapter 18Attention-Head Diagnostics
  20. Chapter 19MoE Expert and Router Diagnostics
  21. Chapter 20Facets and Emergent Pathways
  22. Chapter 21Gradient Residuals, Stability, Curvature, and Update Efficiency
  23. Chapter 22Adaptive Cadence and Event Detection
  24. Chapter 23CLI Reference
  25. Chapter 24Configuration-Key Reference
  26. Chapter 25API Reference
  27. Chapter 26File and Schema Reference
  28. Chapter 27Glossary and Concept Index
  29. Chapter 28The Flow Map
  30. Appendix ADiagnostic Availability Matrix
  31. Appendix BLLM Prompt Library
On this page

Chapter 27 — Glossary and Concept Index

Terms

These definitions describe what the public readings mean and where their interpretation stops. Keep the chart’s units, scope, uncertainty, and availability statement with a value. Related explanations are in How IRIS™ Measures, Adaptive Cadence, and the Flow Map.

Adaptive cadence. A variable diagnostic sampling schedule. Denser capture is an instrumentation choice, not proof of a particular learning event.

WD penalty. A retired weight-decay-penalty channel. Do not treat a retired field as an active measurement in a current run.

ANISOTROPIC. A spectral-signature label emphasizing uneven concentration in measured structure. It is not a model-quality verdict.

BASIS_ROTATING. A spectral-signature label drawing attention to changing representation orientation. It does not establish the cause of that motion.

BBP edge. An academic signal/noise reference from random-matrix theory, named for Baik, Ben Arous, and Péché. Its presence in a discussion does not establish a universal training threshold.

b, spectral exponent. The slope of a measured eigenvalue spectrum on log-log axes. Compare its trajectory with k95, entropy, and the spectrum itself.

Bench. The browser dashboard that opens captured diagnostic bundles and presents their measurements, notes, and exports.

BOUNCE. An event label in IRIS telemetry. Read it as a recorded event marker with its run context, not proof of convergence, failure, or a particular mechanism.

BROAD_SPECTRUM. A spectral-signature label for broadly distributed measured structure. Broadness is descriptive, not automatically desirable.

Bundle, .iris. The primary diagnostic record written by the tracker and opened by a compatible Bench. Treat it as a confidential model-inspection artifact.

C-space. Covariance space: a representation-space view distinct from parameter motion and trainer-reported outcomes.

Category. A probe’s subdomain when one is declared, otherwise its domain. Categories only add a separate comparison when they differ from domain labels.

Ceilings. Resource settings that constrain capture coverage. Read the selected mode and any shed or skipped channels rather than inferring data that was not recorded.

CKA, Centered Kernel Alignment. A representation-similarity reading. Similarity under one measurement does not prove two layers have the same function.

Core energy. A spectral-variance reading associated with dominant structure. Read it with frontier and tail views and the measurement’s definition.

current_margin. A directional stability-margin reading. Under adaptive optimizers it is not a fully calibrated guarantee of convergence or safety.

Calibration. A channel’s reported qualification or availability state. Preserve uncalibrated, skipped, or unavailable states in interpretation.

CFL ratio. A curvature-related comparison with a measured directional boundary. Its qualification depends on optimizer and coverage; it is not a safety guarantee.

Cutback. The supported control for fast and steady-state bounds in adaptive capture. It changes sampling work and temporal resolution.

path_excess_cumulative. A cumulative C-space path-excess reading. It describes the recorded representation trajectory, not physical dissipated energy.

Dead neuron. A unit marked inactive under the available measurement and probes. That label does not establish permanent irrelevance to the training distribution.

Domain. A top-level human-assigned probe category, such as arithmetic or code. It is not automatically a model-internal grouping.

Domain composition. A view of how labeled probe domains are represented in the measurement. Read its reference, normalization, and coverage before interpreting a difference.

Domain stability. A temporal view of a domain’s representation consistency. Interpret it with measured steps, probe coverage, and capture cadence.

dθ/dt, basis rotation. A change in principal representation directions over captured training steps. It measures orientation change, not the cause of the change.

Principal-direction carry-over. A view of how representation directions persist between measurements. Read the actual frame interval and capture provenance.

Extraction. A diagnostic capture of the model at a training step. An already measured step is not written again as a new observation.

Facet. An observed grouping of probes with similar representation paths. It is distinct from a supplied domain and is not an identified causal circuit.

FEA, Finite Element Analysis. An engineering analogy used in discussing training dynamics. Analogous terminology does not establish material properties or safety certification for a network.

FFN, Feed-Forward Network. The per-position feed-forward component of a transformer layer. Architecture and capture determine which of its readings are available.

Frame. One captured diagnostic snapshot at a recorded training step. Display transitions between frames are not new measurements.

Frontier energy. A spectral-variance reading near the boundary between dominant structure and weaker directions. It is not a standalone generalization predictor.

HBM, High-Bandwidth Memory. Device memory relevant to capture resources. Read device and host resource limits in the context of the actual machine.

HF, Hugging Face. The model/tokenizer ecosystem used by supported integrations. Match the model wrapper, tokenizer, and version to the integration.

encoder_module. A model-spec attribute identifying the encoder in an encoder-decoder integration. It must match the actual model structure.

HVP, Hessian-Vector Product. A standard operation used in curvature measurements. A directional or restricted reading is not the full optimizer geometry.

lambda_max_proxy. A legacy proxy name. IRIS does not produce a diagnostic by temporarily moving the customer’s parameters; read the actual available curvature channel instead.

Gradient residuals. Training-dynamics readings made available through optimizer registration and, for some channels, compatible evaluation callbacks.

IRIS. Intrinsic Rank Inspection System, Saccadic’s training instrumentation tool. It reports measurements for researchers to interpret with independent evidence.

iris.cmd. The run-directory text file for supported live tracker commands. The training loop must poll the instrument for commands to take effect.

iris.log. The tracker-side log of run state, capture outcomes, and applied controls. It is distinct from the Bench process log.

JL, Johnson–Lindenstrauss. An academic random-projection reference. A projected reading must retain the uncertainty and coverage stated by its figure.

k95. The number of leading eigenvalues needed to account for 95% of measured activation variance. It is an effective-dimensionality reading, not a quality score.

Representation velocity². A concept-space representation-motion reading. It is different from weight motion, physical kinetic energy, and trainer-reported accuracy.

cov_velocity_sq. A covariance-space representation-motion reading. Compare its observed timing with actual outcomes and sampling resolution; it is not a universal predictor.

Velocity sensor. An instrumentation event/telemetry channel associated with representation motion. A marker does not establish what learning mechanism occurred.

Layer numbering. The indices used by the measured model layers. Check the spec and chart labels rather than inferring module identity from screen position.

LM head. The language-model output projection. Do not confuse a model’s readout with an attention head.

MDS, Multidimensional Scaling. A projection method for displaying relationships in fewer dimensions. A display projection does not expose every property of the original representation.

Memory mode. The capture-coverage state reported for a frame. Reduced, shed, and skipped data must remain explicit in later interpretation.

MoE, Mixture of Experts. An architecture with routed expert components. Expert traffic, representation structure, and parameter motion are different readings.

N_anti. A count of sampled probe-gradient pairs assigned an anti-alignment label. It does not establish that two tasks are inherently incompatible.

Observed Δθ. The model’s recorded parameter displacement. It is distinct from a learning-rate setting and from representation motion.

Petal chart. A polar display of measured domain or spectral structure. Read its legend and definition; the shape is not a circuit diagram.

PCA, Principal Component Analysis. The standard principal-direction analysis used to describe variance. Its measured spectrum depends on the data and representation being analyzed.

path_excess. A C-space trajectory reading. Despite engineering analogies, it is not a measurement of physical energy or heat.

cov_displacement_sq. A covariance-space displacement reading relative to the reference named by the measurement. Preserve that reference when comparing runs.

Pinned pool. A host-memory resource term. Resource availability can affect capture coverage, and the run’s reported mode and reasons remain part of the evidence.

PR, Participation Ratio. A threshold-free spectral effective-dimensionality measure. It is distinct from the k95 variance-coverage count.

Probe. A labeled input used to inspect the model. A fixed probe set is not automatically an independently sampled held-out test.

RANK_ONE. A descriptive spectral-signature label for strongly concentrated measured structure. It is not a verdict about capability or training quality.

Representation. The internal pattern of activity a model produces for an input. A projection or summary preserves only selected aspects of it.

RISING_K95. A temporal label drawing attention to increasing measured effective dimensionality. It does not establish improved generalization.

RMT, Random Matrix Theory. The academic study of random-matrix spectra. An initialization or noise comparison does not by itself validate a model-quality claim.

SE, Spectral Entropy. A reading of how evenly variance is distributed across measured spectral directions. Interpret it with the spectrum and probe coverage.

Spectral signature. A descriptive label summarizing measured spectrum/orientation behavior. It is not an automatic diagnosis or quality grade.

Subdomain. A finer human-assigned category within a domain. When declared, it is the probe’s category for applicable category-level views.

Subspace similarity. A comparison of measured representation subspaces. Similarity alone does not establish identical semantics or computation.

Tail energy. A spectral-variance reading associated with weaker directions. Its interpretation depends on the measurement, model, and probes.

τ, tau. A covariance-similarity telemetry reading. It is not a standalone event predictor and should be read with its sampling and coverage.

τ trigger. A legacy label: τ sampling is telemetry, not a standalone cadence trigger. Use the actual event and schedule information recorded by the run.

Tracker. The instrument running inside the training process, capturing diagnostic frames and receiving supported live controls.

Training affinity. A comparison of sampled probe-gradient directions. Agreement or opposition is conditional on the sampled inputs and measurement, not universal task compatibility.

cov_step_sq. A covariance-space step reading. It describes representation-space motion and should not be interpreted as physical work.