Deterministic pipeline. Agent judgment. Auditable end to end.
The hard part of an AI system isn't the model — it's keeping it trustworthy. This one splits the two cleanly: a deterministic CLI does everything mechanical, and an agent is invited only where genuine judgment is needed — always inside a versioned contract, always verified afterward.
Six stages, one direction
Discover new episodes from each channel (YouTube RSS / Rumble). Deterministic; runs on a schedule.
Fill the transcript — a paid caption API, or local GPU speech-to-text for Rumble. Deterministic.
An agent analyzes one transcript into a structured report, under a per-channel contract. Judgment, on rails.
Coins, themes, and frameworks accumulate into a wiki of durable knowledge. Deterministic.
The weekly pass distills it all into the live per-channel and global dashboards — marking where the channels agree and surfacing where they disagree. Judgment, on rails.
A deterministic gate renders only the safe synthesis pages out to this website. No human copy-paste.
Guarantees, not vibes
Every mechanical action is one deterministic command — reproducible, scriptable, testable.
The agent's inputs and required outputs are pinned in versioned contract files, not buried in a prompt.
State only advances when a run checks out Expected vs Claimed vs Actual. Mismatch = refuse.
Every claim carries a source tag and date, traceable back to a specific episode.
Each run leaves a record — what was expected, what changed, what was verified.
The publish gate is a closed allowlist with a fail-closed check — paid material and secrets can't leak.
Regenerated on every publish from the gate's own output — not a screenshot, and not written by hand. The withheld citations are sources from a private channel: the conviction publishes on its public receipts, the private citation is dropped rather than republished. Checked 2026-08-25.
Compression without losing provenance
A week of both channels is ~14 hours of watching. The pipeline compresses that into an ~22-minute read without losing the trail: you can always trace a line in the synthesis back to the episode and date it came from. The summarization is the visible part; the discipline around it — deterministic steps, contracts, verification — is what makes it something you can actually rely on.