Give humans and agents the same experiment
Use when: animation, UI state, or sequence-dependent behavior makes “fix this bug” hard to verify. Lewis Metcalf’s Amp example starts with a ball teleporting through a rotating ring: randomly timed screenshots hide the causal sequence.
Action:
extract
the
existing
state-transition
logic
into
one
module
used
by
both
the
application
and
a
small
replay
harness;
do
not
write
a
second
simulation.
Encode
the
reproducer’s
inputs
in
a
URL
or
fixture,
including
initial
state,
seed,
step
count,
and
timestep
where
relevant.
Render
a
static
trajectory
for
inspection,
then
expose
the
same
inputs
through
a
headless
command
that
prints
bounded
frame
logs.
The
article’s
example
is
node
physics-cli.js
--vx=-7.71
--vy=2.14
--from-frame=17
--to-frame=25
--delta.
This
is
a
project-local
interface
example,
not
an
installed
utility.
Prompt:
“Reproduce
this
fixture
before
editing.
Use
frame
deltas
to
isolate
the
failure;
after
repair,
replay
the
same
case
and
neighboring
inputs,
then
show
the
visual
result.”
Acceptance check: freeze the bad fixture and expected behavior before repair. Repeated runs must reproduce the same trace on the same build. Require agreement between headless state and visual playground at selected frames. The original build must fail the agreed oracle; the repaired build must pass it without changing the fixture or weakening expectations. Nearby edge cases and the actual application must also pass. Preserve before/after logs, images, inputs, and source revision.
Evidence: Metcalf’s dated engineering post publishes exact velocity URLs, a shared-module CLI sketch, commands, and frame-level output. Its delta trace makes an edge-collision discontinuity inspectable rather than merely described. OpenAI’s separate harness report supports exposing UI, logs, and metrics directly to agents; it does not independently validate this physics repair.
Caveat: text diagnostics accelerate localization, not proof. A clean CLI trace can miss rendering, timing, integration, or nondeterministic failures. The published CLI sketch omits argument parsing and logging details; it is not a complete drop-in package. No demo, code, or runtime test was executed for this newsletter. Keep the harness small and retain human visual acceptance.