12 KiB
Read-only original process-state observer
2026-11-06. This investigation tool is separate from the modern runtime. It is intended to establish actual startup registry state and narrow the unresolved Carl skip-to-idle discrepancy. It does not supply a gameplay fix or certify exact callback order. No original execution is claimed by the initial compile/guard checks below.
Execution boundary
tools/recon/original_state_observer.c launches one explicitly attested generated
game copy with CreateProcessA, samples it with ReadProcessMemory, then requests
ordinary WM_CLOSE on that process's own visible window. It never writes target
memory, installs breakpoints, invokes target functions, forces actor state, or
terminates another process. If ordinary close does not succeed, the child remains
running; the caller must handle its own isolated environment. Existing Wine
capture infrastructure cleans up only its private wineserver afterward.
The first argument must be --copied-game=<path>. This is the caller's explicit
attestation that the target is a generated copy. Both the absolute path and the
handle-resolved final path reject a case-insensitive Original path component.
The tool accepts only the three exact hashes below and the original1019904-byte
file size. It holds an authenticated read handle without write/delete sharing
through process creation. The flag is not a cryptographic proof of provenance;
use the existing generated-copy wrapper, never a hard link to an original file.
Alternatively, --attach-copied-pid=<PID> explicitly attests that an already
running process uses a generated copy. This mode acquires only
PROCESS_VM_READ | PROCESS_QUERY_LIMITED_INFORMATION | SYNCHRONIZE, obtains its
actual image path with QueryFullProcessImageNameA, and applies the same final
path, size, file hash and live instruction guards. PID is a strict positive
32-bit decimal value; malformed/overflowed values are rejected. The process
handle binds the selected process throughout validation/sampling, rather than
reopening a potentially reused PID. Attach mode never calls CreateProcess,
focuses a window, sends input, requests WM_CLOSE, or terminates the target.
--menu-keys is rejected before opening the process. Sampling ends while the
original session remains under the user's control.
No network API is used. Run under tools/recon/reference_capture.py for its
existing no-network/no-original-mount boundary. Native Windows execution must
likewise use a generated directory containing copies of the game and its assets.
Game-created settings/saves then remain in that generated environment.
Supported subjects and evidence
| Profile | Exact executable SHA256 |
|---|---|
| canonical | f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8 |
| hd-resolution | b6675b7e7f3fc37512185ddb4d6ce1fbe9a36c8b83f3b4b3ed6979a56db3d2d4 |
| hd-nocd | 8cd5d21997835a40978b8b83e344c97c92217bf6cd6858f7458b76935d64dedb |
Canonical instruction Evidence is
ev_0d64cc58fa8f5573fc8e61e0f2c20e36288099212bd2257d3a3121361fd6c904,
retained in artifacts/jimmy-movement-audit/rea-original-disassembly.json.
The HDnoCD subject has separate complete REA instruction Evidence
ev_23555fac476985c913a80805a12b25c26c25347c6e8210ee188471dd815c276d,
retained in artifacts/jimmy-skip-original-sw/rea-hdnocd-disassembly.json.
The supplied HD resolution variant preserves the observed operands. The HDnoCD
profile is independently identified; its observation is not automatically a
canonical-original behavior claim. Other supplied variants are rejected.
Each global address is independently guarded by its actual live instruction operand before sampling. Runtime inspection is refused if those anchors do not match within5s. This avoids assuming that the other variant has identical data addresses. Preferred-base VAs and PE file offsets follow:
| Read role | Instruction VA / file offset | canonical and HD global | HDnoCD global |
|---|---|---|---|
| Raw control owner | 475A40..475A4B /75A40..75A4B setter | 5099E4 | 509A04 |
| Registry sentinel pointer | 473740 /73740 load | 50999C | 5099BC |
| Viewport seconds | 47374F /7374F load | 696030 | 696050 |
| Current game complete pointer | 46E360 /6E360 store | 509948 | 509968 |
| Shared skip | 430F3E /30F3E store | 4F83E4 | 4F8404 |
| Input block | 426062 /26062 load | 4F8430 | 4F8450 |
| Modal gate | 407060 /7060 load | 4F8198 | 4F81B8 |
| Skip suppression | 430F11 /30F11 compare | 4F8182 | 4F81A2 |
Registry nodes are {next,previous,payload} at offsets0/4/8, from the recovered
46FF4A..46FF7F insertion and473740 traversal. Traversal stops at the sentinel,
1024 nodes, an unreadable node or a repeated payload. complete:false means the
sample cannot establish the whole list; mutation during a read is possible.
Jimmy is recognized by game-interface vtable4A373C. The current game registry
payload is its complete pointer+8C. A successful full-list sample records their
actual indices without executing either actor or inferring constructor order.
Jimmy samples use the game interface: current Euler224..230, destination274..280, dirtyA, movement mode704, stored turn614, and action latch1D40 (=complete1E00). Support fields13F and4A0 are reported separately as cached and force-support bytes: actual458960 reads4A0 first, otherwise471A10 reads13F. They are not an executed support query. The source addresses are code-field contracts, not guessed layouts or corresponding fields from the modern simulation.
Output and ordinary input
Output is bounded JSONL on stdout: authenticated identity, registry changes,
nominal10ms state samples, optional normal input events, finish and completion
marker. Sleep(10) is a sampling request, not a guaranteed interval or a game
clock. Each record includes actual elapsed wall-clock milliseconds. Floating
point NaN/infinity is written as JSON null. Narrow Win32 path/title bytes outside
ASCII are escaped so the output remains valid JSON; use ASCII generated paths
for reliable cross-machine reproduction.
valid_mask uses bits0..3 for readable skip/input/modal/suppression bytes, bit4
for readable viewport delta, and bit5 for a readable full Jimmy field block.
Zero values without the corresponding valid bit are placeholders, not evidence
that the game state was zero. Jimmy fields are omitted without bit5.
By default no keys are sent. --menu-keys explicitly opts into ordinary Return
presses at approximately3/6/9s, each held100ms while sampling continues. It finds
only the child process's visible unowned window, requests normal window focus,
and uses SendInput only when that window is foreground. Records include title,
window/foreground handles, press/release and events accepted by SendInput.
Accepted events do not prove that the game's input system consumed them or that
a particular menu item was selected. This sequence is not a recovered New Game
macro. --copied-game-menu=<path> is an explicit single-argument equivalent for
the existing capture wrapper, which passes one --probe-model argument.
Samples consist of multiple independent reads without suspending the game. They can miss transitions shorter than the interval and can mix adjacent callbacks. They establish observed retained state and list membership; they cannot alone prove the first writer, same-frame control order, exact input/controller order, physics chronology, or the complete Carl skip lifecycle. A black screenshot is still not a usable visual reference even if process state can be read.
Build and use
From GamePort:
mkdir -p artifacts/original-state-observer
i686-w64-mingw32-gcc -std=c11 -O2 -Wall -Wextra -Werror \
tools/recon/original_state_observer.c -lbcrypt \
-o artifacts/original-state-observer/observer.exe
python3 tools/recon/reference_capture.py --asset-probe \
--probe-executable artifacts/original-state-observer/observer.exe \
--probe-model=--copied-game=NeutronHDnoCD.exe \
--probe-success ORIGINAL_STATE_OBSERVER_COMPLETE --seconds 24 \
--output artifacts/original-state-observer/capture
The observer itself defaults to15s and accepts --seconds=1..120 on native
Windows. For normal
input opt-in, substitute
--probe-model=--copied-game-menu=NeutronHDnoCD.exe above. Original variants with
a CD prompt may not proceed to gameplay in this environment; no prompt bypass or
executable modification is implemented by this tool.
On native Windows, first start an ordinary generated copy and reach the desired Carl/tutorial route. Obtain that game's process ID, then start the observer from a separate terminal, redirecting stdout to a local trace:
observer.exe --attach-copied-pid=1234 --seconds=60 > carl-original.jsonl
Replace1234 with the copied game's actual PID. The existing session is left
running and receives no observer-generated input or close request. --help
prints usage without opening a file or process. The same process path/hash
restrictions apply even if another executable is named Neutron.exe.
Initial verification: MinGW32 compilation with all warnings treated as errors passes. All24 independently PE-mapped source anchors (8 per supported subject) match; artifacts/original-state-observer/source-guards.json records original hashes, exact bytes, addresses and file offsets. These are compilation/static checks only. The primary investigator owns subsequent isolated runtime captures, integration, additional evidence and any commit.
Integrated Wine32 observation
The primary's generated-copy launch and focus/Return checks reach MAIN only: game-interface table4D0C0C, no Jimmy, black captures. An isolated xfwm4 desktop does not produce a usable gameplay reference. Viewport delta0 here does not prove the menu selected New Game or that menu/engine clocks share the gameplay clock.
The final attach revision is exercised by
tools/recon/original_observer_smoke.c. This Wine-only driver requires the
wrapper's generated Z:\work\game directory, launches the supplied HDnoCD
copy, then attaches the observer for10s. It verifies observer exit0 and that the
original remains active afterward, before sending its own child's ordinary close.
The observer itself sends no input or close in attach mode.928 JSON samples parse,
the exact live guards pass, and capture wineExit0 plus both completion markers
are present. MAIN only is observed; this is API/tool validation, not a Carl trace.
The tested observer source/build hashes, driver identity, logs/capture metadata
and parsed summary are in
artifacts/original-state-observer/attach-runtime/observations.json. The initial
launch capture lacked a pre-recorded observer executable hash and includes a
Wine diagnostic interleaved with JSON; it is retained separately as a limited
functional observation. The final attach capture has explicit build provenance
and928 intact samples. Native Windows execution remains unverified.
Reproduce after building the observer and driver, using the already prepared generated Wine copy/prefix (no source original is mounted):
i686-w64-mingw32-gcc -std=c11 -O2 -Wall -Wextra -Werror \
tools/recon/original_observer_smoke.c -o artifacts/original-state-observer/attach-smoke.exe
cp artifacts/original-state-observer/observer.exe artifacts/jimmy-skip-original-sw/observer-under-test.exe
python3 tools/recon/reference_capture.py --asset-probe \
--probe-executable artifacts/original-state-observer/attach-smoke.exe \
--probe-success ORIGINAL_OBSERVER_ATTACH_COMPLETE --seconds 18 \
--output artifacts/jimmy-skip-original-sw
Primary's launch observation, before attach support: separate HDnoCD profile passes all live guards; normal focused menu input and an isolated window manager still yield only MAIN (game index0, interface vtable4D0C0C), no Jimmy, and delta0. The final run recorded2042 samples across22.02s. This validates the observer's bounded launch/read path under Wine, not original gameplay or Carl's return. Attach support was added for a reference machine where the user can reach the route before starting a bounded trace. Its strict compilation/source checks are separate from native Windows attach execution, which remains unverified here.