seesaw

name	seesaw
description	Weighs competing findings or conclusions from a review/comparison output before it's finalized — checks whether the stated conclusion is actually balanced against the evidence on both sides. Triggers after `check` (comparison) or `ship` (ideation) produces a verdict.

seesaw — balance check

Process

1. Isolate the verdict. Find the conclusion/verdict the output actually asserts (e.g. "method A outperforms method B", "this is the most promising direction").
2. Weigh both sides explicitly. List the evidence supporting the verdict and, separately, the evidence or conditions that would weaken or reverse it (different eval setup, smaller scale test, cherry-picked benchmark, narrower assumptions).
3. Check the tilt. If the supporting side is thin (single benchmark, single paper, favorable-only setup) relative to the counter-evidence, the verdict is overweighted — flag it and soften the claim or add the missing caveat.
4. Re-balance the output. Rewrite the verdict to state the conditions under which it holds, not as an unconditional universal claim.

Anti-Rationalization

• "The comparison table speaks for itself" — a table without an explicit statement of what conditions the conclusion depends on invites overgeneralization by the reader.
• "It's obviously better, no need to caveat" — obviousness is not evidence; state the actual scope of the claim.
• "One counterexample doesn't matter" — one counterexample from a fair comparison does matter; note it rather than omitting it.

Evidence

Balance is achieved when: the final verdict states the conditions/scope it applies under, and any known counter-evidence or weaker-comparison caveat is explicitly surfaced rather than omitted.

Red Flags

• A verdict stated as universal ("X is better") when it only holds under one benchmark/setup
• Supporting evidence listed but no counter-evidence considered at all
• Silently dropping a comparison data point that contradicts the stated verdict