A record that is public but unread
The @HouseofLords channel carries question sessions, debates, committee-style content and the Lord Speaker's lectures: hours of dense, video-first, unrepeatable public record. It is published, searchable in principle, and almost entirely unwatched. Video is the wrong format for the audience it would serve — people who would happily give six minutes to a summary on a commute, and no minutes at all to a two-hour committee feed.
That gap is not an AI problem. Asking a model to "summarise a debate" is the easy part, and the part everyone demonstrates. The hard parts are the unglamorous ones: proving which recording you actually processed, keeping the source material inside your own building, and staying on the right side of the law that governs parliamentary broadcasts. This strand documents a working system built around those three constraints.
Any model can summarise. The engineering is in everything that happens before and after it does.
Four things it demonstrates
| Constraint | How it shows up in the system |
|---|---|
| Governed intake | Every source recording is registered in an Excel run book — with checkboxes and auditable columns — before anything is processed. The same workbook later accepts new URLs and guarantees nothing is processed twice. |
| Local, reproducible stack | LangChain orchestration, Ollama-served Llama models, FAISS vector search. No API keys, no third-party AI services, no source text leaving the machine. |
| Integrity and audit | Every transcript is SHA-256 hashed at ingestion; every artefact path, actor and timestamp is recorded. "What exactly did the pipeline see, and when?" is one lookup away. |
| Law in the pipeline | The legal basis for using parliamentary broadcasts is the reporting exception at s.45 of the Copyright, Designs and Patents Act 1988. The prompt that writes each script is engineered to stay inside it: report, summarise, quote briefly, attribute — never republish. |
What an episode is, and is not
An episode is an AI-assisted report of proceedings: a machine transcript of a recording, retrieved and synthesised by a local model into a spoken-word script, reviewed by a person, read by a synthetic voice, and published with attribution, an AI disclosure and a corrections policy.
It is a report about the proceedings — not a copy of them. That distinction is the whole legal design, and it is enforced in the prompt, not in a policy document.
The distinction has teeth. Republishing the broadcast would need a licence nobody is offering. Reporting on it — summarising, quoting briefly with attribution, analysing — is what the statutory exception exists for. Every downstream control in the system, from the prompt to the human gate, exists to keep the output on the reporting side of that line.
Fig. 01 — one URL in, one reviewed episode out
Fig. 02 — why the index is built per recording
Based on House of Lords proceedings (UK Parliament). Not affiliated with, or endorsed by, the UK Parliament.