# TraceVox Research

> TraceVox is a visual, reproducible research environment for studying
> intelligent agents: reinforcement learning, trustworthy AI, multimodal
> agents, and adversarial robustness. It records complete per-timestep
> scientific evidence (ground truth, delivered observations, corruption,
> policy distributions, outcomes), replays it exactly, and exports it as
> provenance-carrying research bundles. Observable - Reproducible - Adversarial.

Operating model: the browser is the research control and visualization layer;
the TraceVox Research Runner performs computation on the researcher's own
hardware (laptop, workstation, lab GPU, cluster); the TraceVox backend
coordinates jobs, events, metadata, and publication. TraceVox is not a cloud
GPU service, and the public website never trains models on its own servers.

Public research is readable by humans and machines. Private research is
private by default; publication is explicit and granular.

## Canonical public URLs

- https://tracevox.ai/ : research homepage
- https://tracevox.ai/research : Public Research Library (published projects + experiments)
- https://tracevox.ai/research/projects/{slug} : public project page
- https://tracevox.ai/research/experiments/{slug} : public experiment page (results, recorded replays, manifest, citation)
- https://tracevox.ai/research/docs : documentation index
- https://tracevox.ai/public-research/index.json : machine-readable library index (tracevox.public.bundle.v1)
- https://tracevox.ai/public-research/experiments/{slug}/experiment.json : experiment record
- https://tracevox.ai/public-research/experiments/{slug}/manifest.json : provenance manifest
- https://tracevox.ai/public-research/experiments/{slug}/citation.json : citation metadata (URL/version; no DOI)
- https://tracevox.ai/public-research/experiments/{slug}/episodes/{episode_id}.json : full recorded step trace (research.v1)

## Documentation

- /llms-guide.txt : practical usage guide
- /llms-full.txt : comprehensive machine-facing reference (architecture, schemas, protocols, security)

## Data formats

- research.v1 : per-timestep scientific step record (environment ground truth,
  clean + delivered observation, corruption, policy distribution, value,
  outcome, trust placeholders that stay null until a trust model exists)
- tracevox.public.bundle.v1 : published-experiment bundle (plain JSON files;
  the format is the canonical contract, servable from any static host or a
  self-hosted TraceVox backend's /api/public routes)

## Current status (honest)

- Recorded pilot evidence is published at /research (dissertation gate
  experiment: PPO on MiniGrid-LavaGapS7 under observation corruption).
- Three usage modes, never confused: TraceVox Cloud (hosted, account
  required, experiments run on your own paired compute); Local/Offline —
  `pip install "tracevox-ai[rl]"` then `tracevox local start` runs the
  COMPLETE research environment (backend + persistent storage + UI) on your
  machine with no account, no cloud, fully offline; Explore Published
  Research (no account, no install). CLI: `tracevox doctor`,
  `tracevox local start`, `tracevox connect`, `tracevox runner start`.
- Trust calibration and protective actions (ACT/VERIFY/DEFER/ABSTAIN) are
  research milestones — NOT implemented; TraceVox never presents future
  research as existing capability.
