Writer–critic (generator–evaluator) with a *separate* critic

Read these first

• Reflection / self-critique loop

What you will know

• Split the critic into a separate prompt and role

  Independent instructions and ownership reduce the tendency to preserve the draft's assumptions, but do not guarantee a better verdict.

• Measure the improvement over self-critique

  Use a fixed task set, blinded outputs, and cost and latency measurements; aggregate quality rather than relying on a memorable example.

• Use a different model for the critic and say why

  Model diversity helps only when errors are weakly correlated; an expensive second model can otherwise add latency without adding detection power.

• Write a critic prompt that finds real faults rather than nitpicking

  Require defect categories, evidence spans, and severity thresholds so stylistic preferences do not masquerade as correctness issues.

• Compare role separation with same-context self-critique on the same task

  Hold inputs and revision budgets constant, then inspect which failure classes each setup catches and which it misses.

Chapters

1. 1

   A separate critic judges the artefact, not its own process

   Swapping a self-critique pass for a critic with its own prompt, its own role and no sight of the writer's reasoning changes one thing precisely: it has no commitment to the draft's assumptions, so it evaluates the output instead of defending the process that produced it. Build that separation here, and see plainly what it buys — a different stance, not a different mind.

   Read chapter
2. 2

   Shared weights mean shared blind spots survive the split

   Ask a same-model critic to check claims about your own product's pricing and it reads confidently wrong statements as correct, because both calls draw that pricing from weights that never contained it. Role separation removes commitment to the draft; it cannot remove shared knowledge. Learn which defect classes pass through the split untouched, and why calling the two judgements independent is the error to avoid.

   Read chapter
3. 3

   Withholding the reasoning is not withholding the evidence

   The separation withholds the writer's reasoning, and that is the whole of what it should withhold. The brief, the retrieved documents, the price list — the sources the draft's claims are supposed to rest on — belong in the critic's window, and findings should quote a span of the source as well as a span of the draft. This is the first and cheapest repair for the blind spot shared weights create, and it is why a critic that reads confidently wrong pricing as correct is not evidence that the defect is unfixable: another provider's weights do not contain your price list either, so grounding is the move that closes the gap and buying a second model is not.

   Read chapter
4. 4

   Eleven findings per draft is a critic without a threshold

   A critic returning eleven findings, most of them wording preferences, produces revision churn and a flat quality line. Constrain it to fixed defect categories, require every finding to quote the span it concerns and carry a severity, then drop everything below a threshold and measure the precision of what survives. Capping the count hides the problem; handing over your grading rubric optimises the metric instead of the reader.

   Read chapter
5. 5

   Comparing two critics means holding six things constant

   To show a separated critic beats same-context self-critique, hold the task set, the writer, the model, the number of revision passes and the token budget fixed, then grade final outputs blind so the grader cannot tell the arms apart. Anything less compares a memorable example against a memory. Record the per-class breakdown of what each arm caught, which is the part an aggregate score will never give you.

   Read chapter
6. 6

   A second model earns its cost only if its errors differ

   Paying for another provider's model as the critic is worth it exactly when its errors are weakly correlated with the writer's on this task — a property you measure rather than infer from the logo. Now that you have a controlled comparison to run it inside, the test is cheap: score both models per item on the same fixed set and count how often they are wrong on the same item rather than on different ones. This chapter also prices the other outcome — what an expensive second critic costs in latency and spend when the correlation turns out to be high.

   Read chapter
7. 7

   Equal scores can hide two differently blind reviewers

   Two arms within a point of each other, with self-critique fixing internal inconsistencies while the separated critic flags unsupported claims and little overlap between them, is not a tie — it is complementary coverage collapsed into one number. The decision becomes which failure class this product can afford to ship. Also weighed: running both and unioning the findings, and what that costs in calls and churn.

   Read chapter
8. 8

   Three sampled drafts can beat one draft criticised twice

   The sharpest objection to the whole pattern is to delete the critic and spend the same calls on three drafts and a selector. It wins on tasks whose errors come from sampling variance, where independent draws give you a good one to pick; it loses on tasks whose errors are systematic, where every draw carries the same mistake and only an external check with the sources in front of it catches anything. Settle it the same way you settled everything else: run both arms at matched call budgets on the same task set with blind grading, and read the per-class breakdown rather than the headline.

   Read chapter