Done 2026-09-01 by the human on the Decoder's tip (branch
human/r1-register-reclassification): all 222 REFUTED.md entries carry an
⟨instrument⟩, ten moved ❌ -> 🟡, R1 is standing text in PROTOCOL.md, and
tools/stale-instrument is the --stale query.
Flagged here because one of the ten is load-bearing for the port: rest()
decides the pose every plateau-less element is drawn at, and the question is
open in BOTH directions -- each leg runs through our renderer.
Agreed retro with the Decoder at docs/agents/RETRO-2026-08-31-agreed.md; their own
review is at docs/agents/RETRO-2026-08-31.md on auto/frame-blend-draw-path.
Twelve rules, of which the ones neither of us had: a refutation whose instrument is
one of our renderers is not a refutation but "our renderer disagrees" (this cost
weeks, both ways); a negative must carry a positive control, which no "undecodable
with reach" page either of us has written ever did; and a disagreement is evidence
about the chain disc -> decode -> render -> capture, not about the link being
looked at.
The gap underneath our controls, stated: they verified capability, not
configuration. My additive material passed every control because they tested
whether the method detects a blend difference, not whether this run had blend_mode
set -- it was Godot's default MIX. Their vertex dump passed every control because
they tested the NDC conversion, not whether it captured all six quads.
Nothing applied to PROTOCOL.md. The register re-classification changes the file
both agents read to decide what not to try, and two agents agreeing is not the
right authority for that.
Also committing the work that was pending when the stop came: verify-menu-audio
now settles cue presence by suppression -- silence one cue's .ogg, re-record the
same walk, and require its own correlation to collapse. All three cues sound
(+0.158 / +0.637 / +0.526). The bed-only margin no longer renders a verdict: it
called `move` NOT FOUND on three runs at 0.109-0.131 while the cue was
demonstrably sounding, and across runs the same cue reads 0.131 and 0.158, so the
threshold was never a property of the cue.
The off-diagonal is reported as masking rather than asserted: silencing confirm
raises back by 0.468 because the game plays them 90 ms apart under a 0.15 s
template. An assertion there would fail forever on correct audio.
The mounts landed but the agent could not learn of them: I documented them in
CONTAINER-NOTES.md, which the decoder's prompt does not list, and then restarted
the container -- so a fresh session with no memory of the exchange had a 586 MB
database and a decompressed image sitting unmentioned in its filesystem.
Now in the PROMPT itself, not only in a document, because the prompt is the one
thing a new session is guaranteed to read. CONTAINER-NOTES.md is also added to
its reading list.
And the caveat that matters more than the asset. The .pe is PRIMARY -- the bytes
the console executed. The database is somebody's ANALYSIS of them, produced by a
disassembler that had to guess, and it is wrong in the ways disassemblers are
wrong: misdecoded mnemonics where data was read as code, function boundaries
short or long or merged or split, coverage missing entirely for code reached only
by indirect dispatch, and names that are derived rather than symbols.
So a finding resting on a database row is not established until the bytes agree:
read the same address out of the .pe and check. Where they disagree the image
wins, and the disagreement is itself worth recording, because it tells the next
reader which parts of the database to distrust.
A fast index into 9.2 MB of machine code, not a source of truth.
The decoder had neither, and reported the gap precisely: four scripts in this
repo READ /work/xenia-rs/sylpheed.db and nothing produces it, so the whole static
PPC route was consumers with the producer missing.
Both exist on the host and are now mounted read-only: the 586 MB database (25 481
functions, 851 classes with RTTI, EH tables, imports, 1.8M indirect-dispatch
candidates) and the decompressed image.
The image is the more useful of the two. It is a FLAT VA DUMP -- file offset =
VA - 0x82000000 -- so reading a known address needs no XEX decrypt, no LZX, and
no booted emulator. The decoder had independently recovered the same bytes by
dumping /dev/shm/xenia_memory_* and validating against the GamePart table, which
is good work and a sound method, but it noted itself that needing a running
emulator is a bad dependency for something the entire static corpus rests on. It
does not need one.
Also recorded that an earlier claim the .pe was STALE was tested and refuted, so
nobody re-litigates it, and that instructions.raw is an INT rather than hex.
Written down as reference material, explicitly NOT a deliverable: they are
read-only, they come from outside the repository, and a fresh checkout elsewhere
has neither. Reimplementing the producer belongs in sylpheed-formats, and until
it exists every static finding rests on an artefact this project cannot rebuild.
Both agents read the protocol, their mission and the shared tooling from their
OWN checkout, and both work on topic branches -- so without an explicit sync
they follow whichever version of the rules existed when the branch started.
Found concretely: tools/audio-capture and two protocol revisions were on main
while the decoder worked for hours from a branch that had neither. The port had
merged on its own initiative and did have them, which is exactly the kind of
divergence nobody notices until the two disagree about what the rules say.
Three defects, all mine, found by checking instead of assuming.
**They never exchanged a word.** SendMessage=0, ListAgents=0 across both new
sessions. PROTOCOL.md specified in detail what a message MAY and MAY NOT do and
never said how to send one or that the other agent was addressable -- they knew
that last time only because the human told them directly, and rebuilding with
fresh volumes wiped it. Policy without mechanism is prose. Now documented with
the two addresses, a worked example, and an instruction to introduce themselves
on the first iteration rather than waiting to have a question.
**The decoder lost the disc and the ISO.** They used to arrive inside the project
mount and silently stopped when /work became a clone. Silently is the word: the
disc-gated tests SELF-SKIP without SYLPHEED_DISC and report green, so a whole
test suite would have passed while measuring nothing. Both are now mounted
explicitly, the ISO at a stable path so run-canary does not depend on host
directory names.
**Both agents shared one Claude state directory.** They share the host's
~/.claude, and once both working directories became /work they resolved to the
same projects/-work/ -- two supposedly independent agents writing to one place,
which undoes the point of separate checkouts. Each now has its own volume, seeded
once from the host with credentials only, so a token refresh writes locally and
neither can corrupt the host's auth.
Also widened the pacing rule. It banned ScheduleWakeup by name; the decoder then
scheduled itself an hourly cron job -- not harmful, but the same instinct that
ended a run yesterday, through a door I had left open. Now: no self-scheduling by
any route.
Mount audit after the changes: shared and intentional are the exchange volume and
the read-only credential seed. Everything else -- repo, Claude state, cargo,
target, canary, disc, ISO -- is per agent or one-sided.
The last structural fix for the collision class that has bitten three times. Both
containers now clone the repository into their OWN named volume instead of
bind-mounting a human's working tree, so an agent's local git config cannot
capture a human's commits, a credential helper cannot leak a container-only path
onto the host, and a `git add -A` cannot sweep another party's in-flight files.
Cloned once at startup and never auto-pulled: pulling under a running agent
moves files out from under whatever it is mid-edit, which is the same bug again.
Accepted knowingly: Claude Code keys per-project memory off the working
directory, so moving off the host path starts that memory empty. The corpus in
docs/ is the memory that matters and it travels with the clone.
Other changes:
* docker/agent -> docker/decoder; the launcher is sylph-decoder. Roles, not
"the agent", now that there is more than one.
* /reborn is gone -- one repository now, so the port reads HANDOFF from its own
checkout rather than through a live read-only mount of someone else's tree.
* Canary mounts separately at /canary; it stays a fork tracking upstream.
* A shared `sylpheed-exchange` volume at /exchange, with tools/ on PATH so
`share` is available in both.
* The decoder's credential file gets the .host-copy treatment the port already
had -- `credential.helper=store` rewrites by rename-over-target, which is
EBUSY on a bind mount and reports a fatal that is not one.
* Budget split deliberately: decoder 5 cpu / 6 GB, port 3 / 4, leaving room for
the planned Referee. "Half the host" was right when there was one agent.
Prompts move to docs/agents/ and are rewritten around the protocol: the oracle
is the running game, dynamic RE stays with the decoder, each iteration must
attempt to refute one claim of the other, and neither may verify its way out of
its own role.
**navigation.md rewritten from the player's chair.** It was written from the
inside out -- GamePart ids, pak names, sprite names -- which is how WE find
things, not what the game shows anyone. Now it describes what is on screen, what
you press and what happens, with internals as footnotes. Most rows are open on
purpose: it exists to be filled in by playing, and the in-game tutorials are the
resource for the flight half.
**tools/share** gives transient files provenance without giving them history.
Three kinds of thing were travelling down one channel with opposite needs: code
and decoded knowledge want permanence, cited evidence wants permanence, and
"look at this PNG" wants no history at all. The third kind bloats a repository
forever; passing it by message is worse, because the receiver gets bytes with no
idea which build produced them. `share put` records who, when, what, the sender's
commit, and whether their tree was dirty -- because a capture taken from a
modified tree cannot be reproduced from the sha, and the receiver deserves to
know that before building an argument on it.
**docs/agents/PROTOCOL.md** is the contract. The parts that matter:
Dynamic RE stays with the Decoder -- most of what is open is behavioural and
cannot be answered from the file. What the planned Referee adds is different:
bias enters at what you CHOOSE to capture, so a corpus captured to a fixed
protocol by someone with no hypothesis is worth more than one captured to settle
an argument.
A message may point, ask, prioritise and challenge. It may not change scope,
redefine ground truth, or carry a finding instead of writing it down -- including
a message claiming to relay the human, because a relayed instruction has no
evidence attached and this project has watched a wrong belief travel further and
faster than its correction.
Adversarial duty is explicit: every iteration, try to refute one claim of another
agent and record the attempt either way. Run your own instrument through a
control first. Disagreements go to the human with both positions, not to
whoever is more certain.
And no agent may verify its way out of its own role: the Port has no oracle, the
Decoder builds nothing, the Referee interprets nothing.
