The Catches That Only a Person Made

Attributed Catches Map the Blind Spots

The log records who caught what, and specifically marks the catches made by the human reviewer at the final viewing. That looks like credit and it functions as diagnosis.

A defect caught by an automated check tells you the check works. A defect caught by the author reviewing their own work tells you the self-review works. A defect caught only at the final viewing tells you something different and more useful: every gate before that one was blind to this category. Collect those and you have a map of where your process cannot see, which is precisely the information you need to decide where the next check goes.

Several of this pipeline's hard checks exist because of exactly such a catch. Someone watching noticed the piece began and ended abruptly; noticed that reused art was not new; noticed that a factual claim inherited from the source material was out of date. None of those are mechanically detectable, and each one named a blind spot precisely enough to act on.

The Mode Switch

Before the human gate there is a cheaper step: the author reads the work cold, as a first-time audience member, out loud, once per line. This sounds redundant — the author just wrote it, and presumably was paying attention.

It is not redundant, because authoring mode turns off the circuit that asks whether a sentence lands. While writing you are holding intent, structure, and the next sentence; the question "would someone without my context follow this?" cannot run at the same time. Review mode has that circuit on and the others off.

The important part is that this is a mode difference, not a capability difference. Both modes are available to the same person in the same session. The pilot demonstrated it directly: the same system that produced an inadequate draft correctly convicted its own output the moment it was reading rather than writing. Nothing about the capability changed — only which question was being asked.

Place the Gate Before the Spend

The human gate sits before any synthesis or rendering. That placement is what keeps rejection honest. A reviewer looking at a script has thrown away one authoring session by rejecting it. A reviewer looking at a finished render has to throw away paid synthesis, generated art, and hours of encoding — and at that price, "it's not quite right" quietly becomes "it's fine."

Sunk cost does not have to be believed to be effective. It is enough that the reviewer knows what rejection destroys, and the fix is structural: put the gate where rejection is cheap, so that the standard does not have to survive an argument with the budget.

Code

Reading a log for blind spots rather than for defects·text

caught by      what it tells you
-------------  ------------------------------------------------
automated      the check works. no new information.
check

self-review    the author's review pass works for this category.
(cold read)

HUMAN GATE     every earlier gate was BLIND to this category.
(final view)   <- this is the row that should drive new checks

the audience   the whole process was blind, and it shipped.
(after ship)   <- the row you are trying to keep empty


# from the pilot, catches made only at the final viewing:
#   "it starts and ends abruptly"        -> pacing, not measurable
#   "that's the art from last time"      -> became a cache fix
#   "this claim is out of date"          -> became a sourcing rule
#   "it talks as if others live here"    -> became a stance rule
#
# two of the four became code. the other two became doctrine,
# because they resisted encoding - and that is a real outcome too.

Exercise

Take your most recent piece of work that got reviewed and list every issue the reviewer raised. For each, ask which earlier gate should have caught it and did not. Then pick the single most common blind spot across the list and design one concrete check for it — automated if possible, a specific question on a checklist if not.

Hint

Look for issues the reviewer raised that you agreed with immediately. Those are the highest-value entries, because agreeing instantly means you could have caught it yourself and did not — which is a gate problem rather than a knowledge problem, and gate problems are the fixable kind.