TRACE-FORMAT.md

Importing a capture

Use Export trace to download a complete working example. The same JSON imports into this viewer. Import reads the file locally and never sends its contents to a server. Closing the tab discards imported state.

Supported schemas are flint.trace.v1 for Flint and model.trace.v1 for Flint or Qwen. Legacy Flint exports without model metadata remain supported. Both registered checkpoints use Q8_0. A different architecture or quantization requires a separate registry entry and compatible measurements.

A trace contains:

Field	Contents
schema	Supported schema identifier
model	Model identity and quantization; id is flint or qwen
architecture	Text-decoder configuration matching models.json
prompt	Original prompt, up to 4000 characters
settings	Recorded output limit, temperature, seed and reasoning flag
start.tokens	Decoded input-token pieces, including chat-template tokens
frames	1 to 256 output-token frames, in generation order

Each frame records text, token_id, optional raw UTF-8 token_bytes, input_position, duration, elapsed, and one entry per decoder block in layers. input_position begins at start.tokens.length - 1 and increases by one. It identifies the input that predicted this frame's output token. Capture context is limited to 2048 positions.

Every layer includes finite, nonnegative heads, ffn and residual RMS measurements. Head count follows the layer type. Flint MoE layers add six unique experts, six normalized weights, and shared and routed output RMS. Its first two layers are dense.

Attention layers include one attention entry per head, with top position/probability pairs, entropy, and Flint's sink probability. Only the strongest eight positions are retained. Recurrent layers instead include write_gate, retention, and state_rms arrays per value head.

The viewer checks all frames and relevant architecture dimensions before changing the loaded trace. It accepts stopped captures containing valid complete frames. It rejects empty captures, malformed JSON, inconsistent model identity and invalid measurement dimensions. Files larger than 160 MB are rejected before parsing.

A compatible exporter must record actual tensors from an instrumented model execution. An ordinary text completion does not supply attention, routing or recurrent-state measurements. The bundled captures provide concrete, downloadable examples of the complete format.