Analysis files are UTF-8 JSON, optionally wrapped in gzip when the filename ends in .json.gz. Current normal-chat
exports use aibrain.session-analysis.v2; Infinite-mode Analysis+ exports use
aibrain.infinite-analysis-plus.v2. Both schemas describe real-time inference telemetry.
Files are deliberately compact. They do not contain renderer arrays, per-node frames, model weights, source prompts, hidden states, attention matrices, MLP activations, or chain-of-thought.
| Field | Meaning | Type / units | Classification |
|---|---|---|---|
schema |
Export contract identifier. | string | status |
created_at |
Export creation time. | ISO-8601 UTC | measured |
conversation |
Visible chat or Infinite simulation turns. | array | measured |
graph.kind |
real_time_inference_telemetry_map. |
string | status |
graph.topology |
Always display_only; nodes and links are not model topology. |
string | status |
graph.nodes, graph.edges |
Size of the rendered display layout. | counts | display-only |
graph.measurement_channels |
Ordered names of the nine normalized telemetry inputs. | string array | status |
graph.cluster_colours |
Stable display colour for each channel. | hex colour map | display-only |
measurement_integrity |
Human-readable measurement source and graph limitation. | string | status |
integrity |
Analysis+ statement identifying autoencoder inputs and excluded model internals. | string | status |
recorded_frame_summary |
Compact findings over the retained telemetry frames. | object | derived |
frames |
Number of retained telemetry frames included in the result. | count | measured |
mean_novelty, peak_novelty |
Autoencoder embedding-distance/reconstruction findings. | unitless | derived |
mean_reconstruction_error, mean_coherence |
Autoencoder reconstruction quality for this export. | unitless | derived |
most_active_channel |
Channel with the highest mean normalized measurement across retained frames. | channel name | derived |
maturity |
Persisted training-health state, conditions, and readiness. | object | health/status |
smart_analysis |
Present only in aibrain.infinite-analysis-plus.v2. |
object | derived |
analysis_storage |
Paged/resident counts, cache limit and use, and data-loss status. | object | measured/status |
smart_analysis.neural_network |
Autoencoder architecture, feature schema, calibration, and lifetime health. | object | health/status |
session_findings |
Compact channel and temporal-pattern findings. | object | derived |
channel_profile |
Mean and peak normalized value for each measurement channel. | ratio 0–1 | normalized measurement |
key_events |
Up to 24 notable telemetry frames with raw metrics and learned findings. | object array | mixed |
Each key event identifies its generated step, visible output_text chunk, dominant channel, Real-time source, raw telemetry,
embedding novelty, reconstruction error, coherence, and compact embedding. Raw telemetry uses these fields when a live
logits snapshot is available:
| Telemetry field | Meaning | Units |
|---|---|---|
context_tokens, context_limit |
Evaluated context position and requested context size. | tokens |
output_tokens, chunk_tokens |
Retokenized visible count in this reply and current chunk. | tokens |
stream_latency_ms |
First-chunk latency or observed time since the previous visible chunk. | milliseconds |
retokenized_tokens_per_second |
Current visible chunk's retokenized count divided by its latency. | tokens/second |
vocabulary_size |
Number of raw logits inspected. | entries |
raw_logits_available |
Whether the chunk carried a valid raw-logit snapshot. | 0 or 1 |
raw_logit_entropy_bits |
Shannon entropy of the raw-logit softmax. | bits |
raw_top_probability |
Largest raw-logit softmax probability. | ratio 0–1 |
raw_top_five_mass |
Sum of the five largest raw-logit softmax probabilities. | ratio 0–1 |
raw_confidence_margin |
Difference between the largest and second-largest raw probabilities. | ratio 0–1 |
recent_output_occurrences |
Matches for this emitted text chunk in the previous 32 visible chunks. | count |
The raw-logit fields describe the distribution before penalties, grammar, top-k/top-p/min-p/typical filtering, temperature, and token selection. They are not final sampling probabilities.
maturity.state is Baby, Teen, Adult, or Elder. Baby requires at least 2,048 persisted frames and sustained
recent consistency before Teen. Adult additionally requires at least 32,768 frames and sustained consistency/learning
slowdown. Elder requires a sustained post-Adult overfitting signal and freezes training weights. This is not an
accuracy claim. Baby and Teen exports carry readiness: "caution"; Adult and Elder carry readiness: "ready".
The NPZ model stores feature_schema=aibrain.realtime-inference-telemetry.v1. An older NPZ without that identifier may
contain retired simulated features and is rejected instead of mixing those weights with real-time telemetry. Generate a
new response to create a compatible model. The standalone inspector reports the reason explicitly.