Preflight

Preflight

Run this on a draft before plan-review, not instead of it. Preflight edits the plan; review only objects to it.

The premise is measured. One load-bearing process document took ten cold adversarial rounds and roughly sixty-two objections to reach a clean pass, and a post-mortem of those objections found four preventable classes:

• About half were hand-typed facts — counts, paths, URL params, state enumerations — written from memory and wrong.
• Fix rounds minted new contradictions, because the same rule was restated in several sections and a patch desynced them.
• Serial generalist panels each found one stratum; the single parallel specialized round found the most.
• The deepest bugs were only findable by executing the plan's first steps — a URL param the page silently ignored, a comparison instrument that could only render one state, a fixture whose absence degraded quietly instead of failing.

Three of the four are cheaper to prevent than to review. That's preflight: three moves, in order.

1. Generated-facts appendix

Every count, path, URL param, enumeration, version, or config value the plan states must be produced by a command. The appendix records command → output pairs; the plan's prose cites the appendix entry rather than repeating the figure from memory. A hand-typed fact is a defect, not a rounding error — the reviewer's job then becomes checking the command, which is cheap, instead of re-deriving the number, which is a round.

• An enumeration needs the command that enumerates it — grep -c per event type, an ls of the fixture directory, a --help dump, a schema query. "The six states" is a claim; the command that prints six is the fact.
• Re-run the appendix after any edit that touches a cited figure. A stale appendix is worse than none, because it reads as grounded.
• A fact about behavior — what a param does, whether an endpoint honors a flag — is not a grep. It belongs to move 2. Do not let a grep that finds the param's name stand in for evidence that anything reads it.

Verbatim command output

Every command → output evidence block is pasted verbatim from an actual command run. To abbreviate output, change the command so it produces the shorter output (for example, grep -m or head); never edit the captured output, insert ellipses, or drop matches. Manual edits are prohibited and require a BLOCKED verdict on their own.

2. First-hour spike

Before any reviewer sees the draft, execute the plan's opening steps against the real artifacts for the effort needed to establish its admitted behavior. Open the page and pass the param. List the network fetches. Run the instrument once, in each state it claims to compare. Delete the fixture and watch what the fallback actually does.

The rule that catches the expensive class: every claim of the form "X is addressable / served / discovered / honored by Y" gets one live probe. Those claims are the ones that read as obviously true and are silently false, and no amount of reading finds them.

Findings amend the plan before review starts. The scratch work is throwaway — a worktree, a scratchpad script, a curl log — and is never committed; the plan cites what the spike established, not the scratch.

3. Single-source rules

Each rule or constraint is stated normatively in exactly one place. Every other mention points at that place instead of restating it. This is what stops a fix round from minting a contradiction: a patch can only desync a rule that exists in two voices.

Before the doc leaves preflight — and again after every later fix round — grep each rule that appears in two or more places and reconcile it down to one normative statement plus references. Late rounds are when this decays fastest, because that's when patches are landing under time pressure.

Hand-back

Preflight returns the grounded plan to its caller, with the generated-facts appendix and the spike's findings attached. The caller continues its next authorized step, which may be plan-review; preflight itself does not claim the parent workflow ended.

Recommend that review's first round be three to five parallel specialized lenses rather than one generalist pass — machinery-interaction, cold-executor simulation, inventory verification, and domain judgment. Serial generalist rounds each surface one stratum and then repeat each other; parallel lenses surface the strata at once, which is what a preflighted plan needs, since the shallow findings are already gone.