The Evidence Freeze Before Final Polish
TL;DR
Final polish is where many good drafts become less honest.
The fix is simple: before cleanup, freeze the approved claim set.
An evidence freeze means:
- supported claims stay in scope,
- uncertain claims stay labeled,
- unsupported claims do not get promoted by prettier wording.
Once the evidence is frozen, the final pass can improve readability without renegotiating what is true.
Context
Earlier workflow patterns in this repo focused on sourcing, uncertainty, and verification. The missing piece is the transition between "verified enough to draft" and "ready to publish."
That transition is dangerous because the incentives change. At first, the writer is thinking about correctness. At the end, the writer is thinking about flow, rhythm, and elegance. Those are valuable concerns, but they often create pressure to remove caveats, compress context, and make conclusions land harder.
This is exactly the kind of local optimization that produces global trust problems. NIST’s AI RMF frames trustworthy outcomes as lifecycle properties rather than isolated checks. Editorially, that means the verification stage and the polish stage cannot be allowed to quietly overwrite each other.
Key Points
1) Freeze the claims, not the prose
An evidence freeze does not mean the wording is locked.
It means the underlying approved claims are locked:
- what is supported,
- what is mixed,
- what remains uncertain,
- what was excluded.
The prose can still change. The truth budget cannot expand without another review pass.
2) Most late-stage drift is small but consequential
It rarely looks dramatic.
More often it looks like:
- shortening a caveat out of existence,
- merging two ideas into one stronger idea,
- turning an example into a recommendation,
- converting a bounded claim into a general lesson.
These are easy to miss because they happen sentence by sentence.
3) Freezing evidence makes style work safer
Once the claim set is fixed, the editor can move faster on:
- repetition,
- transitions,
- paragraph order,
- headline sharpness,
- sentence rhythm.
That is useful because it stops every style change from reopening every factual question.
4) The freeze should be visible in the workflow
If it only lives in someone’s head, it will not hold.
The easiest form is a compact claim sheet:
- approved claims,
- claim status,
- source links,
- uncertainty notes.
When the final draft deviates from that sheet, it triggers a return to review.
5) Freezes are most useful when you publish frequently
Daily or near-daily publishing makes it tempting to treat the final pass as magic.
But frequent cadence is exactly where workflow boundaries matter most. You do not have the luxury of rediscovering editorial discipline from scratch every day.
Steps / Code
Minimal evidence freeze note
Approved claims:
- AI systems often over-answer when workflows reward completion over uncertainty.
- Final editing can strengthen unsupported claims if caveats are removed.
- Publish review should re-check all high-impact claims.
Mixed / uncertain:
- Exact effect size of these controls depends on team process maturity.
Excluded:
- Any universal claim about all AI-assisted writing workflows.
Freeze rule
Before final polish:
1) lock approved claims,
2) mark uncertain claims,
3) reject new strong claims,
4) reopen verification if certainty increases.
Trade-offs
Costs
- Adds one explicit checkpoint before publishing.
- Can feel rigid when you want to keep "improving" the draft.
- Requires a small claim artifact outside the prose.
Benefits
- Prevents late-stage credibility erosion.
- Lets editors focus on readability with clearer guardrails.
- Makes review disagreements easier to resolve.
- Scales better for high-frequency publishing.
References
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0): https://doi.org/10.6028/NIST.AI.100-1
- NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile: https://doi.org/10.6028/NIST.AI.600-1
- OpenAI, Why language models hallucinate: https://openai.com/index/why-language-models-hallucinate
Final Take
The smartest time to get stricter is right before the prose starts looking finished.
That is when evidence needs to freeze.
Otherwise polish becomes a loophole.
Changelog
- 2026-03-17: Initial publish on the evidence-freeze pattern for final editorial polish.