Fluency is the failure mode
A confidently wrong quotation, a mis-attributed speaker, a hallucinated statistic — each would damage both the legal basis the project depends on and the credibility it is trying to earn. And each arrives looking exactly like correct output, which is what makes fluency the specific danger rather than a general one.
Step 9 of the pipeline is a person, not a process.
What the reviewer checks
Every generated script is read against the source transcript before it becomes an episode. Five checks, in order:
- Quotation fidelity — every quoted phrase appears in the transcript, verbatim, in context.
- Speaker attribution — each claim is attributed to the person who actually made it.
- No fabrication — no invented statistics, dates, votes or positions.
- Proportion — quotations are brief and illustrative, consistent with reporting rather than republication.
- Framing — the episode reports; it does not endorse, and implies no backing from the House or any member.
The run book enforces the sequence mechanically. The pipeline ticks Script Generated and stops; audio is generated only when a person presses approve, so the machine cannot take that step for itself. The gate is not a convention the team observes; it is a state the system requires.
One honest gap, since this section is about evidence. The approval is not itself written to the audit log — the log records machine stage events, and the reviewer’s decision is not one of them. What evidences the review having happened is that the audio artefact exists at all, because nothing else can produce it. That is weaker than a signed entry naming who approved what and when, and closing it is a change to the pipeline rather than to this page.
Fig. 05 — step nine is a person, and there is no way round it
Labelling, and provenance by construction
Each episode is labelled AI-assisted · human-edited. That is honest about provenance — a draft was machine-generated — without overclaiming: a person owned the final version. There is no UK statutory labelling obligation for this today; the label is voluntary, cheap, and consistent with where the 2026 copyright-and-AI report pointed on transparency.
The audit trail behind that label is not a document written afterwards. It is a by-product of running the pipeline: every episode page links to the source recording and states its basis, every artefact is hash-linked to its transcript, and every generation records the model and settings that produced it.
An audit trail you have to assemble later is a story. One the pipeline emits as it runs is evidence.
Corrections
Errors are inevitable in any published output, and a corrections policy is the difference between a mistake and a credibility problem. Corrections are made openly, the episode notes what changed, and the run book carries the change in its audit log alongside everything else. The same trail that proves what was produced also proves what was fixed.
Based on House of Lords proceedings (UK Parliament). Not affiliated with, or endorsed by, the UK Parliament.