Boids original-model transform comparison, 7 November 2026

The original registered dx_boids model, captured world/view/projection matrices and viewport reproduce all 4,680 submitted X/Y/Z/reciprocal-W components exactly across 39 naturally completed bird draws. Each of the three consecutive observation windows contains all 13 birds, wholly outside the homogeneous frustum. The personally inspected compositor capture shows cyan terrain without a visible bird. The captured guest transforms explain this terrain-only phase; no mismatch or emulator repair is demonstrated.

Evidence: scratch/runs/20261107-boids-model-transform/, especially oracle.json, oracle.py, recovered/original/upstream-transform.json, boids-disassembly.txt and review/startup.png.

The unchanged executable is SHA-256 b9327e6fb2758e232d0ca0fb410b11097fd557dc59b3d405167d976db208b8d9. Disassembly at 0x40da21 supplies primitive 4, vertex type 1, the original 16-vertex array at 0x420520, the 30-word index array at 0x41fa80, and calls Device2's indexed-draw method at vtable offset 0x78. The initialization and scaling writes to this model are also retained in that disassembly. At the qualified owning GPU's natural _draw boundary, the observer copies those original initialized model/index bytes using the runtime image base and translated guest pointers. All 39 model copies have identical SHA-256 8fc924ccf9febcd27bddad760ee88fbdce1215582a88efe35ef558b56d4b96c2. Their extents are X ±0.5, Y ±0.150000006 and Z −0.5 to 0.5.

Each raw draw also retains its immutable 4 KB command-state snapshot: world, view and projection at offsets 0/64/128, recorded composite at 192, viewport rectangle/scale/origin at 3088/3104/3120 and clip normalization at 3904/3928. Completion sources/spans correlate by exact command sequence and ordinal, including the first bird submitted separately from the following twelve. The first observation window begins partway through terrain; these are bird draw windows, not three independently captured complete compositor frames. The screenshot and geometry are sequential observations, not frame aligned.

The independent Python scalar oracle multiplies the three matrices from their raw values, projects the original indexed model and applies the captured viewport. It does not use the recorded composite as its input. Its composite has zero differing float32 components; all 1,170 submitted vertices then match exactly in all four position components. A separate 60-digit Decimal calculation applies world, view and projection successively without a rounded composite. Its largest differences from submission are 0.000488733 X pixels, 0.000512035 Y pixels, 2.113e−7 depth and 4.227e−8 reciprocal W, consistent with float32 rounding. Identity, translation, known perspective and deliberate wrong-world controls pass. Independent homogeneous classification from the Decimal clip coordinates places all 13 birds outside in each window.

This excludes a model-to-submitted-position transform discrepancy in these draws. It does not independently validate the original x86 simulation or camera-matrix construction. The historical SDK source describes time-seeded scattering and camera motion; this run neither fixes a seed nor reproduces an earlier phase. Native Windows parity, lighting, rasterization, full-flock convergence, FPS and historical27 completion remain unqualified.

Before the original launch, the actual observation branch passed on production guest/render objects: 25 raw draws and 25 completed submissions with state and translated model copies. Existing owner/module/freshness rejection controls, independent native-release controls, exact Arrow/quit ACKs, deferred-observer negative/positive tests and actual-driver 25-minute deadline controls passed. The focused draw-trace suite passed. No production code changed, so no repair regression or new full build was required.

One UUID-idempotent box, bx_2svdb5d8, was allocated for 40 minutes at 13:09:35 UTC, with immutable driver deadline 13:34:35.276 UTC. The inherited Globe staging omitted boids.exe; this was caught after allocation but before any original launch. manifest-http-preflight.json records the correction, all 38 registered executable/media inputs and 60 exact local HTTP bodies. Only the missing executable and completed manifest were supplemented on that single box. Its 526 pins and all 60 HTTP bodies passed before Chrome/original execution; 269 subsequently served pinned bodies match. There was one original launch, no lane retry, no old-box resume and no public deployment.

Runtime source is Git fe76b23838365466dab8e0f93513bffd5a0e3527 plus source-recovery.patch; source-recovery-verification.json reconstructs every pinned source input through a temporary Git index. Existing media remain path and SHA references. The executing guest/render module was recovered and SHA-verified as e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. This records tested inputs, not a fresh compilation provenance claim. No source-tree/media archive was made. Chrome reports version 151; its advertised Intel WebGL renderer is not hardware evidence: the provider wrapper forces ANGLE SwiftShader and /dev/dri is absent.

No original input, scene/camera/RNG/clock/configuration edit, guest execution probe, added fence, flush or readback was used. Only the already-qualified image-base and address-translation getters resolve passive copies. Natural submission hooks preserve batching and renderer results. The browser closed normally with exact quit ACK and zero native events. An immediate read-only terminal inventory safely rejected while closure was settling; a subsequent distinct inspection confirms both PIDs absent, both ports closed, profile and writer absent and no held input. Nothing was replayed.

All 35 remote artifacts (18,664,226 bytes), including the live module and raw transform observations, were retrieved and SHA/size verified before the explicitly delegated normal nonforced stop. The service confirms archived at 13:12:34.198 UTC, expiry cleared, stop completed and forced=false. Local fixture PIDs, ports, profile and writer are clear; home free space stayed above 2 GB. The global evidence audit retains the pre-existing Alice error; this investigation has no candidate-corpus entry and attaches to none.