IDent dynamics | c. 2026

v2.1.1

Score analysis

Compare this folder run's decisions against its labels. A label is a hit when a decision falls within its time span ± the margin below — needed because some labels are only milliseconds wide.

Target

Margin (s)

Metric	Value

Time: — Freq: — Energy: —

Drop a wav file, add a folder, or connect to a stream.

or click the file icon on the left

Loading…

Name	Size	Modified	Models

Import labels

File contents

Separator First row is a header Pad zero-length to 1 ms

Ignore tags

Defaults (used when a field isn't mapped to a column)

Low frequency (Hz) High frequency (Hz) Tag Description

IDent dynamics learn detect notify

Guest preview · Sign in · Privacy

Data

FFT

FFT size

Hop overlap

Window

Channel

Display

Magnitude scale

dB floor: -150 dB

Saturation %: 99.5

Frequency axis

Colour map

Height scale: 0.00

Light intensity: 0.65

Labels

Show overlay

No labels loaded.

Src	Tag	fmin (Hz)	fmax (Hz)	tmin (s)	tmax (s)	Desc	Author

Decisions

Show overlay

Custom colour

No decisions loaded.

Src	Signature	t (s)	Decision	Reason	Frame	Active URL	All URL

Model inspection and builder

Loading…

Name

Rules

thr notify

Show overlay

Emit detections as decisions

Use templates when running this model

Suppress periodic (sonar) detections cov ≤

My models

Active runs

Target Source Sort by Run on

Loading…

	Colour	Name	Target	Source	Fitness	Macro F1	Updated	Actions

Benchmark models

Evaluate selected models on selected recordings, rank them by macro F1, and export the report (JSON / LaTeX / PDF). Runs synchronously on the training server.

Name (optional)

Target tag

The label tag the per-frame target mask is built for. Filters the model list below.

Additional target tags (optional)

Also score every model against these tags (per-tag metrics) plus the combined "any target". Ctrl/⌘-click to pick several.

Type any label tag — useful for tags that no model was trained on (so they're not in the list above).

Models

Input files

Saved files — files you've uploaded

Stream files — server-side, no upload needed (Enterprise only)

Loading streams…

Files without a sibling labels file are greyed out — they'll still contribute to the benchmark as evaluation-only inputs (just no target-positive frames to score against).

Advanced (fallback STFT params)

Each model is normally evaluated with the FFT/hop/window it was trained at (resolved from its training run's config_json). These settings only apply to models that were pushed manually and have no training run on record. The effective STFT used for each file is shown in the results table.

FFT

Hop

Window

Stream decisions: —

Delete all by model:

No stream selected.

File	Model	t_min (s)	t_max (s)	Score

Active stream: —

Press Connect on a stream in the Saved files panel to start.

Show stream decisions on the spectrogram

Subscribed models — the Brahma service runs each of these on every new file:

Notifications sent

When (UTC)	Type	What	How	To

No notifications sent yet for this stream.

Files & decisions

Decisions written by the Brahma service appear here. Click a file to load it with its decisions overlaid.

Acoustic templates

Target

Loading…

AIS vessels

Connect to a stream to see vessels.

	Vessel	MMSI	Points	Type	Flag	Speed	Course	Destination	Position	Distance	Last seen (UTC)	Since report

Search AIS

Search vessels logged near the sensor.

Vessel	Type	Closest	Closest approach (UTC)	Fixes

Days with AIS data

Click a day to see its hours.

Days with recordings

Click a day to see its hours.

ML Runs

your ML runs

Loading…

Saved setups

Name

Target label tag *

Recordings to train on

Saved files — files you've uploaded

Tick none to train on every saved recording.

Stream files — server-side, no upload needed (Enterprise only)

Loading streams…

Files without a sibling labels file are greyed out — pick them as evaluation-only inputs (they won't contribute training-positive frames).

Generations

Population

Fitness

advanced

Frame spacing Δt (s)

Band min (Hz)

Band max (Hz)

Min rules

Max rules

Seed

Mutation rate

Crossover rate

Elitism

Tournament size

Survivor fraction

Size penalty

Fine-tune detection threshold per bot

candidates

Skip the rendered run images (faster)

Seed model (optional) copies

Allowed rule types — all (no restriction)

Use images — let the GA gate detections on similarity to saved templates of this target prob

Build ensemble — keep a diverse team of bots (not just the single best) and save it as a brahma_ensemble_… model whose score is the fraction of members firing

Templates are the target (no label file) — each picked template clip becomes a fully-positive target; the GA (incl. ensemble pools) evolves bots that score highly on it. No input files needed; tick recordings above to add negatives.

Runs (max 5; each gets its own model)

Or run it from your terminal — copy this brahma train command:

Log

Most recent at the bottom.

Time	Cat.	Level	Message

Agents

your agent tasks

Each agent will plan + train + evaluate models on your behalf, reporting back here. You set the target tag and minimum F1 — the agent decides everything else within the budget you give it.