Q2. The archive is eight screens shipped twice, English and Japanese --
not the "build 4 title, 5 main menu, 6/8/9 submenus" the handoff claimed.
Build 8 is the JAPANESE main menu; 6 and 9 are the EN and JP EXTRAS, and
EXTRAS is the only submenu GP_TITLE holds. The PRESS (A) BUTTON plate is
its own build (2/3), composited over the title art and faded in a beat
later, not a state of build 4.
Confirmed by booting to the main menu and walking it: title, PRESS (A),
main menu and EXTRAS each match their render element for element. Builds
0/1 and 10/11 -- a DELTASABER / SYLPHEED A.I. plate -- were looked for in
the whole boot filmstrip, every title-side screen and the attract loop,
and appear in none of them; the reach of that negative is written down
rather than filled in with a guess.
Two rig traps went into METHOD: the menus drop d-pad presses shorter than
~0.3 s, and a grab 2.5 s after a transition can catch a screen mid-fade
-- which nearly wrote "the returned title has no plate" into the corpus.
Q1 of the menu port, measured against the running game rather than reasoned
about. The developer-logo splash is the cheap target: it is the first thing the
guest draws and its bundle declares short, unambiguous ramps.
Two results, both frame-exact and both emulator-speed-independent (frame numbers
are VdSwap counts, the guest's own frames):
* the ramp is LINEAR. A declared 15-unit fade lands on round(255*k/15) for all
seven of its samples with zero error, k stepping 2,4,6,8,10,12,14. No ease
can reproduce a constant step of 34 at both ends.
* the animation clock advances 2.000 time units per submitted frame, over six
consecutive intervals with no residual, with 1 unit as the quantum
underneath (one frame in the fade-out advances by 1).
The conversion to seconds is one step further and is flagged as such: 300 frames
took 10.87 s = 27.6 present-frames/second, which reads as a 30 Hz title at 92 %
under the emulator and gives 1 unit = 1/60 s -- the title build 4.2 s, the main
menu build 1.1 s. That reading is not proven, because the rate was measured
while the guest was still streaming from the ISO; the page names the one test
that would settle it and says what changes if it goes the other way.
Committed beside it: the raw draw capture and the per-frame quad CSV, so the
numbers can be re-derived without a disc or an emulator.
The autonomous agent's loop prompt had grown to ~9 000 words of accumulated
findings, refutations and traps. That is a symptom, not a style: it was doing
the job the corpus should do, in the one place that dies with the container.
Three concrete failures followed from it, and each gets a structural fix rather
than a louder instruction.
1. IT REPEATED WORK IT HAD ALREADY DONE.
The "do not revive" list and the method traps existed ONLY in the prompt, so
nothing pointed a fresh iteration at them. Extracted verbatim into two
tracked files:
docs/re/REFUTED.md 105 claims tested and dead, grouped by subject so a
grep for your noun finds the neighbourhood
docs/re/METHOD.md the traps already paid for -- controls, inference,
searching, reading data, runtime
Both are linked from INDEX.md, and the loop prompt now opens by requiring
them to be read. This is the fix for "re-derived something already known":
the knowledge is now where the next iteration looks, not in a context window.
2. IT FORGOT TO ARM THE NEXT WAKEUP.
The prompt tried to solve this by shouting at itself in the first line. The
real fix is to stop asking: `loose` now defaults to a FIXED interval (45m),
so the harness owns the cadence and a forgotten ScheduleWakeup cannot end the
run. SYLPH_LOOP_INTERVAL= (empty) restores self-pacing.
3. IT COULD NOT PUBLISH, SO THE WORK ONLY EXISTED IN THE CONTAINER.
New `push-work`, plus a read-only credentials mount
(SYLPH_GIT_CREDENTIALS, default ~/.sylph-git-credentials). It pushes the
CURRENT branch only, refuses anything that is not auto/*, and never
force-pushes -- so the consolidated line stays a human's decision and a
confused iteration cannot rewrite history. The loop prompt now requires a
push on every iteration that commits, rather than at the end of some longer
arc, which is exactly when a container dies.
The prompt itself drops from ~9 000 words to 85 lines and, more importantly,
stops accumulating: findings go in the corpus, and the prompt points at it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
