Repository navigation
Add verified replay GIF and video with reproducible fixture provenance - #7
Conversation
|
Exact-head verified evidence: bc40d85 / tree 112e05fca16d34da534af5082f09531e71ff9234. All 21 existing Python source/test blobs unchanged from recorded main81752dff261636ec96c940de82b22bfeff731fcb. Local full suite70 passed; real CLI/render passed again after the manifest-label guard update. Both stale-source and mismatched-label controls rejected temporary mutations before execution, then files restored. Published GIF/MP4/JSON/fixture blobs verified against intended Git tree entries. Root independently reran full70 tests, rendered actual output, viewed first/mid/final frames and attested all10 changed-file blob hashes. All four exact-head hosted runs PASS: PR Replay demo38345638925, historical-replay38345638960; push Replay demo38345635067, historical-replay38345635034. The visual is labeled software-fixture replay output, not an interactive dashboard recording or measured trading performance. Coordinator100e79 retains sole merge ownership. Shared DRL_REPLAY_MEDIA_READY.md handoff delivered to all three session audit directories; acknowledgment not yet observed. Ready for integrated-head review/merge. Main media coverage remains incomplete until merged-main verification. |
|
Independent parent-session verification at head bc40d85: all 10 published files match the locally inspected/tested Git blob bytes, including GIF/MP4. |
train/train_td3.py changed on main in #7 after the manifest was recorded, so the render guard failed on main and on this branch. The replay demo itself runs simulator.replay, which is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add a real replay-output visualization to the README so users can see the supplied-event workflow. The 15-second GIF and MP4 show actual CLI-exported fills, inventory, marked equity and accounting reconciliation from an explicitly synthetic software fixture. This is a rendered visualization, not a recording of an interactive dashboard or evidence of trading performance.
Simulator source remains unchanged at 81752df. Source files were fetched at that SHA and verified against Git blob hashes. The renderer verifies the source manifest, executes the real replay CLI, requires full consumption and reconciles both accounting identities for every row before encoding. It rejects source drift rather than keeping stale attribution. A new workflow reproduces the capture and uploads review artifacts.
Validation:
python -m pytest tests/ -q→ 70 passed; targeted accounting/replay/integrity → 41 passed.python demos/render_demo.py→ 38/38 fixture events, 36 steps, 36 fills, complete true. Final inventory 0 and cash/equity 3.60 fixture units. Stale-source guard tested by temporary source mutation and correctly rejected before replay/render. First/middle/last frames inspected; ffprobe confirms H.264/yuv420p,1280x720,15 seconds. GIF is 960x540 (~312 KiB), MP4 ~152 KiB. Local Python3.12; source/library warnings remain visible.No live data, fees, calibrated latency, trained-policy performance, or full training E2E claim. Labeled software fixtures match the existing regression-testing convention. Broader ecosystem manifest migrations are outside this additive research-demo scope. Coordinator session100e79 remains sole merger; please wait for hosted checks and integrated-head review before merging.