---
license: mit
language:
  - en
pretty_name: Sentiance traces
tags:
  - cognitive-architecture
  - affective-computing
  - synthetic
  - sentiance
task_categories:
  - text-generation
---

# Sentiance traces

Self-labeled **deliberation traces** exported from the
[Sentiance](https://github.com/sanjaydoc/Sentiance) cognitive architecture — the
data used to train the *fused mind*
([`sentiance-film`](https://huggingface.co/CryptoGod97/sentiance-film)).

Each line is one moment of the cognitive cycle: what the mind was aware of, the
prompt its inner voice saw, the thought it produced, **and the numeric state
vector `m_t`** (valence, arousal, emotion, drives, and per-faculty signals) for
that moment.

## Format (JSONL)

One JSON object per line:

| field | meaning |
| --- | --- |
| `agent` | which named mind produced it (Iris, Milo, Rhea, Cass, Aria, …) |
| `system`, `prompt` | the (system, user) prompt the inner voice saw |
| `prompt_blind` | the same prompt with the felt state removed (state-blind training) |
| `thought` | the produced inner thought (the supervised target) |
| `state` | human-readable inner context (emotion, valence, drives, goals, signals) |
| `state_vec` | the 41-dim numeric `m_t` — the differentiable conditioning input |

## Use

- **Path A (voice):** fine-tune a small model on `(prompt → thought)`.
- **Path B (fused mind):** train a cognition-conditioned transformer on
  `(prompt_blind, state_vec) → thought`.

The full collect → prepare → train → eval pipeline is in the Sentiance repo.

## Provenance & honesty

Fully **synthetic** — generated by Sentiance's own (partly rule-based) cognitive
cycle, so it reflects *that system's* regularities, not the world's. No personal
data. These are **functional** state variables, not evidence of subjective
experience (see the repo's ADR-0002).

## Author

**Dr. Sanjay Anbu** — creator of Sentiance.
[github.com/sanjaydoc/Sentiance](https://github.com/sanjaydoc/Sentiance)

## License

MIT (the Sentiance repo).