Record format

The training record is diary.jsonl in the run directory: UTF-8, one JSON object per line. The authoritative list of event types and required fields is ormas_monitor/schema.py; this page mirrors it. A line may carry more fields than listed; it may not carry fewer.

Common fields

Every line has:

Field	Type	Meaning
seq	int	Position in the record, from 0, contiguous.
time	float	Unix time from the local clock, rounded to milliseconds.
event	string	The event type, from the table below.
prev	hex string	SHA-256 of the previous line (hash of line seq - 1), or 64 zeros for the first line.
hash	hex string	SHA-256 of prev followed by the canonical JSON of the line without prev and hash.

Canonical JSON means json.dumps(body, sort_keys=True, separators=(",", ":")). To verify one line by hand:

import hashlib, json
body = {k: v for k, v in line.items() if k not in ("prev", "hash")}
assert line["hash"] == hashlib.sha256(
    (line["prev"] + json.dumps(body, sort_keys=True, separators=(",", ":"))).encode()).hexdigest()

run_start.record_version (currently 1) changes when a field is removed or its meaning changes. Adding an event type or an optional field does not change it.

Event types

Event	Required fields beyond the common ones	Fixed values
run_start	name, tool, model, parameters, layers, architecture_sha256, optimizer, python, torch, platform, config, record_version
structure	step, pairs, relu_after (also graph: where the autograd walk started, "loss", "model outputs" or null)
manifest	step, check_every, layers, producers, channel_checks, repair, precision, distributed, notes
alarm	step, severity, layer, kind, diagnosis, grad_norm, baseline, sigma	severity = "warning"
alarm_cleared	step, layer, diagnosis, grad_norm, baseline
channels	step, severity, layer, batchnorm, channels, level, n_certified, n_flat, n_observed, n_decayed, certified, flat, observed
channels_cleared	same as channels
certificate	step, severity, layer, batchnorm, channels, count, statement, assumes	severity = "proof"
certificate_refused	step, severity, layer, batchnorm, count, reason	severity = "info"
no_gradients	step, severity, message	severity = "warning"
loss_reduced	step, severity, elements	severity = "info"
repair	step, severity, layer, batchnorm, channels, count, method, optimizer_slices_reset, modification_class, bound, delta_norm, weight_norm_before, weight_norm_after, relative_change, rng	severity = "action", modification_class = "discrete", bound = null
loss_spike	step, severity, value, baseline	severity = "warning"
loss_recovered	step, value, baseline
nonfinite_loss	step, severity, value	severity = "critical"
nonfinite_gradient	step, severity, layers	severity = "critical"
skipped_step	step, severity, scale, layers	severity = "info"
sink_error	sink, error
run_end	steps, checks, counts, metrics_sha256

Field meanings

Field	Meaning
step	The optimizer step at which the check ran, counting from 0.
severity	One of info, warning, action (a repair), proof (a certificate), critical (non-finite values).
layer	The module name from model.named_modules().
kind	The module class name, e.g. Conv2d.
diagnosis	no_gradient, vanishing or exploding. See Alarms.
grad_norm, baseline	The layer's gradient norm at this check, and its baseline (10 to the running mean of log₁₀ gradient norm).
batchnorm	The BatchNorm paired with the layer, or null.
channels	In channels events: the layer's channel count. In certificate and repair: the channel indices (truncated to channel_list_limit).
level	How many of a layer's channels are stuck: none, some, most (half or more) or all.
n_certified, n_flat, n_observed, n_decayed	Counts of certified-dead, constant, silent and decayed channels.
certified, flat, observed	The channel indices of each (truncated to list_limit).
statement, assumes	The certificate's claim in words, and the assumptions it rests on.
reason, count	In certificate_refused: why channels that look dead from their weights were not certified, and how many. Written once per layer and reason.
message	In no_gradients: what to fix (written once per run; every such check is counted in run_end.counts.checks_without_gradients).
elements	In loss_reduced: how many elements the loss passed to step() had; its mean was recorded.
method	How the repair re-initialised the channels.
optimizer_slices_reset	How many optimizer-state tensors had the repaired channels zeroed.
modification_class, bound	Always "discrete" and null: repair is a jump with no enforced bound on its size. See Repair.
delta_norm, weight_norm_before, weight_norm_after, relative_change	The measured size of the repair: ‖ΔW‖, ‖W‖ before and after, and ‖ΔW‖ / ‖W‖ (null when ‖W‖ was 0).
rng	Where the new weights came from: the global RNG, or a seeded generator.
pairs	{producer: batchnorm}, learnt from the dataflow of the first sampled step: the BatchNorm's input is the producer's unmodified output.
relu_after	The layers whose output a ReLU module receives unmodified. Whether a layer may be certified (the ReLU being the only consumer, and the rest of the conditions) is in the manifest event's producers.
architecture_sha256	SHA-256 of name:shape for every parameter, joined by |. Two runs with the same value have the same parameter names and shapes.
metrics_sha256	SHA-256 of metrics.jsonl at the end of the run.

metrics.jsonl

One JSON object per check. A real row from the watch-only demo, one check after conv2 and conv3 were zeroed:

{"step":450,"loss":2.305945873260498,
 "grad":{"conv1.0":0.0,"conv1.1":0.0,"conv2.0":0.0,"conv2.1":0.0,"conv3.0":0.0,"conv3.1":0.0,
         "classifier.0":0.0558704249560833,"classifier.3":0.08159016817808151},
 "stuck":{"conv1.0":[0,0,0,0],"conv2.0":[64,0,0,0],"conv3.0":[128,0,0,0],"classifier.0":[0,0,100,0],
          "classifier.3":[0,0,0,0]}}

(Wrapped here for reading; in the file each row is one line.)

stuck lists, per producer layer, [certified, constant, silent, decayed] channel counts.

Seal and signature

record.json and record.sig, when present, are described in The training record.

Stability

The format is versioned by record_version. Readers should ignore fields they do not know. Before 1.0, event types may be added in any minor release; a removed field or changed meaning bumps record_version.