# jn-engine capture, tooling and oracle review 2026-11-03. Read-only source/artifact inspection; no friend code, scripts, legacy binaries, network services or Ghidra jobs executed. Paths below beginning `J/` mean `../../../jn-engine-master/` relative to this document. Confidence is high for source/artifact inventory; reported historical measurements remain provenance claims until independently replayed. **Useful now:** capture-derived Level 1 fixtures, bounded algorithm reference tests, and address/naming tools. Preserve GamePort's C#/.NET 10, SDL3 and Vulkan architecture; translate selected evidence rather than importing the C/SDL2/OpenGL runtime. **Binary scope.** Local SHA-256 recomputation confirms `Original/Neutron.exe` matches `J/docs/binaries.sha256:12` (`f5281076…1aea8`). However our `Original/OMT2.dll` is `3ec41ead…54943`; friend declares `a94f13e1…b0fea` (`:14`). Executable addresses can be cross-indexed; DLL addresses, helper semantics and captured renderer behavior need version reconciliation. No friend original DLL is included for byte comparison. This difference is not evidence that its captures are wrong. **Instrumentation exists, state tracing does not.** `J/instrument/proxy/ddraw_proxy.c:4` implements a replacement `ddraw.dll` forwarding through `ddraw_orig.dll`; COM wrappers hook scene boundaries, draws and state changes (`com_wrappers.c:202–234,389`). `protocol.h:12–83` specifies packed little-endian OMTC v4, accepting v1–4: transforms, viewport, textures, render/texture-stage states, light/material, draws; v3 adds pixels, v4 masks/color keys. Commands include capture controls, frame marks and camera changes (`:62–71`). Capture uses a 32 MiB producer/consumer buffer; overflow drops whole frames (`capture.c:43,596–616`). Frame time is `GetTickCount`, **not the simulation dt** (`:581–584`). No proposed state-tap addresses or gameplay-state record exist in this instrumentation. **Included evidence has limits.** `assets/capture/` contains 390 files, 13,512,573 bytes. `level1_hudfix/frame_meta.json:4–31` attributes anchor frame 8881 to a 2026-05-25 capture, 640×480, 3,523 draws and 342 textures. Local `draws.json` actually contains 3,523 entries; textures, HUD reconstruction, `scene.bin`, `scene_reproject.bin` and `scene_world.bin` are present. The world fixture combines seven accepted keyframes; three rejected registrations are recorded (`scene_world_summary.json:5–32`). These are useful rendering fixtures, not continuous movement/physics evidence. No `.omtc`, original screenshot named by the manifest, original executable tree, `.gpr` or `.gzf` project is supplied. `validate_replay_fixture.py:34–39` requires absent raw artifacts, so its accepted-frame replay cannot currently be reproduced directly. **Renderer claims deserve targeted calibration.** `PROJECT_HISTORY.md:1242–1259` records unusual projection `[3][3]=1`, zero vertex-alpha caveats, no fog in that capture, billboard tags, HUD FVF `0x152`/orthographic projection and `canvas_id=Canv+1`. Treat these as scoped capture/source findings. Do not infer universal no-fog or force every Vulkan material opaque. Verify matrix storage/transposition explicitly: wire schema says row-major (`protocol.h:92`) while architectural math uses column vectors. Historical histogram improvements are not image parity (`history/native_vs_capture_8881_phase5_report.md:94–117`). Particularly, `tools/check_oracle_diff.py:41–74` checks fixture counts/magic, player presence and FOV/clip values; it does **not** compare images or trajectories. **Oracles: 16 substantive Python scripts plus one template**, rather than the survey's 15. All substantive scripts build/run repository-specific C dumpers. Reuse expected-value functions, corpus enumeration and protocol formats with a C# adapter; no existing generic external-runtime CLI makes them drop-in tests. Strong cases include 3CAM clamp-order/branch precedence with real authored counterexamples (`C3DCutSceneCamera.py:8–35`) and animation mutation tests (`C3DAnimated_dispatch.py:460–498`). Scope remains narrow: 3MCA excludes six out-of-range entries (`C3DMultiCutSceneCamera.py:20–26`); walk-camera uses synthetic inputs/tolerances and documents omitted paths (`C3DPlayer_walkcam.py:29–43`); AI-trigger bypasses activation gates (`C3DAITrigger.py:21–25`). These prove translations against recovered models, not independent original runtime fidelity. **Adoption route.** First compare shared corpus/hashes, then adapt one camera/parser oracle and selected Level 1 render-state fixtures. Ghidra markup/dump scripts exist, but markup creates/replaces structures (`ApplyRttiClassMarkup.java:154–170`); apply only to a copied project with explicit paths. The analysed project remains an offer. State tapping is feasible future instrumentation: require pre/post-update phase, dt, input snapshot, stable actor identity, initial save/state/RNG, collision/camera state and dropped-frame reporting before replay comparisons. Begin with VR04 walking/jumping/jetpack traces. Keep execution local: receiver defaults to all interfaces (`receive.py:392`) with control enabled (`:501`); deployment uploads remotely (`deploy_xp.py:144–145`). Do not run those defaults or remote workflows. Use reviewed offline readers/local fixtures and separate generated directories. No new implementation or runtime validation is claimed here.