A transcript can make close reading faster while making the evidence worse.
Searchable dialogue feels authoritative. Speaker labels add confidence. A clean paragraph is easier to quote than a difficult scene is to verify. Those conveniences can quietly move a transcript from preparation into proof, even when the source cannot support that promotion.
FrameRead treats transcript preparation as its own evidence tier. The material can help find a scene, narrow a viewing window, or prefill a worksheet. It cannot, by itself, verify who performed an action, how an event was staged, what an object did in the frame, or whether an inferred speaker label is correct.
The distinction is not editorial caution added after the workflow. It is part of the data model.
Preserve the source before improving it
Transcript work begins with a source record. The workflow retains where the material came from, when it was generated, what process touched it, and a cryptographic hash of the file. Full transcript bodies remain local; the committed manifest carries metadata and boundaries rather than reproducing the text.
That separation has two uses. It makes a derived file traceable back to the material that produced it, and it prevents an internal research convenience from becoming a public content dump.
When AI-assisted preparation adds speaker attribution, the manifest says that the basis is AI-inferred. It does not convert the label into a verified identity because the transcript looks orderly. A high-quality output can support more efficient search and worksheet preparation while remaining provisional.
This is a better record than a single confidence score. It tells the next person what the artifact may be used for.
Quality and authority are different axes
A transcript may contain clear dialogue, consistent formatting, and many usable speaker labels. Those are quality signals. They do not answer whether the transcript can verify a visual action or an endpoint in a story.
FrameRead’s preparation layer therefore defines allowed and disallowed uses by tier. A strong prepared transcript may support dialogue search, scene-window narrowing, and worksheet prefill with the AI-inferred attribution still marked. It may not lock speaker identity without corroboration, verify visual or causal action, or remove a pending direct-view check.
A weaker transcript receives a narrower role. It may orient a rough scene search without entering canonical analysis. A rejected output has no canonical use.
The product consequence is subtle but important: improving the artifact does not silently improve its authority. That promotion requires a different source or a recorded verification step.
Text cannot prove what happened in the frame
The strongest alternative is to treat a transcript as a compact substitute for viewing. For dialogue-led questions, it may appear sufficient. The problem arrives when a claim crosses layers.
A line can establish that a character said something. It cannot necessarily establish who heard it, what action followed, how the camera framed the exchange, or whether the spoken claim matched what the episode showed. Even official scripts may differ from the aired work.
This is why the workflow fact-locks a complete tuple before analysis depends on it: subject, episode, actor, action, object, status, and endpoint. A source that supports only part of the tuple cannot be made to support the rest through confident prose.
When direct viewing is unavailable, the record stays marked for viewing verification. The uncertainty remains in the operating layer until the appropriate evidence clears it. Reader-facing analysis does not inherit a guess merely because a draft needs to move forward.
AI belongs in the preparation lane when that is the job
AI can still be valuable here. Local transcription can turn lawful source media into searchable material. Speaker suggestions can reduce the time required to locate dialogue. Structured preparation can generate a worksheet starting point instead of an empty page.
Those capabilities are useful because the product has defined where they stop.
The alternative is not a fully manual workflow. It is a mixed one: automation prepares, the manifest records provenance and use limits, and canonical claims wait for evidence with the right authority.
That division also makes model changes easier. A new transcription or attribution method can improve search quality without forcing the editorial system to renegotiate what counts as proof.
A practical source-promotion test
Before a prepared artifact enters analysis, ask three questions.
First, what did this source directly observe? A transcript observes or reconstructs words and timing. It may not observe staging, identity, or causal sequence.
Second, what transformation occurred? Transcription, cleanup, speaker inference, summarization, and extraction create different risks. “AI-assisted” is too broad to describe them.
Third, what new action is the artifact being asked to support? Search is not quotation. Quotation is not attribution. Attribution is not visual verification. Each move needs an explicit reason to trust the source at the next level.
If the answers do not support promotion, keep the artifact where it is and let it remain useful there.
The record should survive the tool
Models, transcription engines, and local workflows will change. A durable research system cannot make its evidence policy depend on one tool’s interface or one output’s polish.
The source record, hash, transformation history, quality tier, provisional labels, and allowed uses should remain legible after the tool is replaced. That record lets a future editor reconstruct why an artifact was admitted, what it contributed, and what still required verification.
The goal is not to distrust transcripts. It is to use them for the work they can actually do.
These boundaries are visible in from close reading to editorial work and local transcription and source-linked research.