# Pulse Court runtime evidence ## Prospective runtime identity, September 21, 2026 `runtime-identity.json` records source `8e4c1ef` and its local documentation checkpoint `f4c01db`. Both evaluator entrypoints retain the same checked, loaded DLL through environment teardown. A completed identity record is required before the producer writes a report. Uncertain native destruction retains ownership until process exit and produces no completed record. The retained logs cover 62 software test methods: eight wrapper/evaluator, 17 producer/export, 15 receipt/protection and 22 offline preflight methods. The integration tests exercise actual entrypoints with synthetic native and ML dependencies. Receipt tests also exercise Windows file protection. Separately, a native child process called exactly five build-identity/ABI getters. It did not create, reset or step an environment, deserialize a model or run inference. The publication review rechecked the 12 source files, Git identities, receipt, DLL and retained logs. It did not rerun those executions. The public JSON is a curated summary with hashes, not a runnable raw receipt. Its local caller pin is not signed attestation. Build pre/post checks are not an immutable source snapshot, and dynamic runtime dependencies remain outside scope. Historical checkpoint binding remains `SOURCE_TO_DLL_UNBOUND`. Export remains rejected and strict deployment disabled. This prospective evaluated-DLL contract does not backfill historical training identity, certify a policy or establish playing strength. The owner checkpoint is local, unmerged and unpushed according to its documentation; the publication review checked local state without a remote fetch. The earlier evidence below remains unchanged. ## Local v4 integration, September 20, 2026 `v4-integration.json` records the later local migration at source `e22769854684166f524548f5c0248bc42b67358d`. The simulator and observation semantics now match the identified v4 reference on 34 synthetic fixtures and 78 snapshots, including the earlier 32/40 subset. Native adapter tests use generated untrained policies. Independent checks ran the recorded owner-built binaries and reconciled their source to the commit; they did not independently rebuild them. Strict deployment remains disabled pending accepted certification and artifact identity. The branch is local, unmerged and unpushed. No trained-model or graphical acceptance is implied, and the earlier preview results below remain unchanged. The unchanged historical v2 replay uses a separate frozen-core verifier on this new branch, after building it from the identified local source: ```powershell & 'build/release/Release/pulse_replay_v2_verify.exe' --replay 'synthetic-preview.pulse' ``` The v4 loader intentionally rejects that v2 replay. Do not rewrite its header. ## Ruleset comparison, September 20, 2026 `ruleset-comparison.json` records a synthetic contract audit of native v2 inputs at `3a3fe53030fb71e60f26c0e969462e825985bffd` and reference v4 inputs at `9a5dfef3cecf82ca985425852036c3fdc4bbefe5`. Audit tooling is `0e28c2217590c05776e059fd6af9ef0ff0c7481a`; the final evidence checkpoint is `057528df4a749154abceb1eaa617567430262814`. The compact public record includes source, tooling, binary and raw-probe hashes plus selected actual records. Across 32 fixtures and 40 snapshots, all 648 decoded action mappings and 80 mask comparisons matched. There were 78 differing observation-slot entries and 16 differing state/touch records. These counts describe diagnostics, not a success rate or model strength. The comparison table shows identical velocity encoded differently, different ordinary-strike motion, and matching body-contact motion with different ownership. Velocity values use raw fixed-point units. Transitions cover one tick; normalization is measured before stepping. The owner built four Release/Debug probe binaries and checked repeated output and configuration agreement. Separately, the coordinator verified 17 source hashes, three tooling hashes and four binary hashes, reran all four binaries with exact matching output, and performed 12 read-only replay checks. It did not independently recompile the binaries. Own-ruleset replays verify and cross-ruleset replays reject. The historical v2 replay remains byte-identical. The probe directly links the existing core and observation encoder. It does not execute the policy adapter or C ABI wrapper. v2 touch-kind values come from a compatibility shim. Raw state hashes include ruleset identity and are not used as semantic equality criteria. This is measured non-parity on synthetic fixtures, not a core migration, full parity proof, policy evaluation, training, fresh holdout, certification or graphical acceptance. Historical September 18 preview evidence and September 20 replay-admission evidence retain their scopes. ## Historical preview, September 18, 2026 Synthetic, untrained development-policy preview. Not certified and not a learned-strength evaluation. Observed 2026-09-18T02:27:14.426091+00:00; exported September 18, 2026. Executable source commit 661f35dba583d51bb33a32e398b3471b29b48e12. Six preview games exercised both seats and recorded 9,483 native network decisions. All six recorded replay-verification commands succeeded. Six explicitly scripted control games recorded zero network decisions. A missing-input strict load returned exit 1 without creating session evidence. Release and Debug each passed four suites. These are execution and regression-test results, not performance, graphical acceptance or certification. results.json preserves sanitized commands, exit codes, per-game counts and hashes. release-tests.txt and debug-tests.txt are sanitized source logs. provenance.json distinguishes checks performed while exporting from the owning project's native run evidence. SHA256SUMS.txt covers every exported file. No new native runs occurred during export. ## Native replay synthetic-preview.pulse is game index 4, selected because it is the smallest artifact. It contains 5,720 ticks and 478 network decisions. The synthetic policy lost. It is not trained-policy play, a browser demo, a video or a screenshot. Use the project's native replay viewer or headless verifier after building the matching source: ```powershell & 'build/release/Release/pulse_headless.exe' --replay 'synthetic-preview.pulse' .\scripts\run_viewer.ps1 -Replay 'synthetic-preview.pulse' ``` Run from the project root with the downloaded replay there, or supply its actual path. No policy bundle is needed for replay verification. Only this replay is packaged; hashes and outcomes for all six remain in results.json. That historical runtime uses ruleset 2; its commands above apply to matching historical source. The later local v4 integration is documented at the top of this guide. Strict trained-policy play still needs deployment-artifact acceptance, latency, certification and human testing. The self-play replay browser and full D2D gameplay migration remain unfinished. The preview policy bundle hash identifies the source fixture but the bundle itself is not included. ## Local replay admission repair, September 20, 2026 replay-admission.json records a separate repair at source commit 596d1f5e8d139bfb39c981e2296f0e8430a7e889, with final evidence documentation at 3a3fe53030fb71e60f26c0e969462e825985bffd. The owning source and documentation are locally committed and unpushed. Their source-file hashes were checked against the working files and Git blobs before this export. The loader requires exactly `34 + 18 * tick_count` bytes before allocating tick storage, rejects malformed files including trailing bytes, and preserves the caller's destination on failure. The old, preexisting binary accepted an appended byte; it was identified by hash but was not independently rebuilt from the baseline commit. No huge-count probe ran against that old binary. The owning project passed 4/4 Release and 4/4 Debug suites plus 14 native checks across both builds. Separately, the portfolio coordinator reran 4/4 Release suites and seven native cases: the unchanged historical replay, a zero-tick replay, trailing bytes, truncated header/frame/hash, and a forged maximum count. Caller-state preservation is covered by core regression tests rather than observable through the CLI. These scopes are recorded separately, not added together. This is structural admission evidence. It adds no semantic input validator, file-size quota, allocation-exception policy or concurrent-file snapshot guarantee. It establishes no physics validation, ruleset-v4 certification, graphical acceptance or learned strength. No training, fresh holdout or new gameplay evaluation occurred. The September 18 replay, results, provenance and test logs remain unchanged; the new JSON contains a compact whitelist of identities and checks rather than raw build logs. SHA256SUMS.txt covers all packaged files except itself.