Stampede observes trade order without claiming money flow
Stampede looks for a specific sequence: one address sells coin A, then buys coin B within a selected window. It labels direct sequences that happen in one transaction, clean sequences where the wallet sold only A, and ambiguous sequences where other sales also occurred. The interface keeps ambiguous rows visible but excludes them from edge weight. That distinction prevents a useful observation from turning into a story the chain data cannot support.
Its scope is Robinhood Chain ID 4663, specifically PONS v2 bonding curves and the Uniswap v4 pools that supported coins graduate into. Plain transfers and other exchanges are outside the dataset. The v0.1.0 sample records 60 minutes on September 10, 2026. It is suitable for learning the tool and inspecting its calculations, not for making general claims about wallet behavior across chains.
Four views expose the same underlying transactions
The terminal shows a live feed and a radar screen. Browser views add a ranked RADAR table, a FLOW diagram around one coin, and a 3D MAP of routes. An evidence card opens the wallets, sell and buy transactions, timestamps, gaps, and grades behind a route. The ranking shows its score parts instead of presenting one unexplained number.
The recorded sample contains 146,329 trades and 33,743 wallets according to the README. Its default 30-minute view has 12,164 observed sequences after the documented direct and clean filtering used by the demo. Stampede also warns that timestamps may be interpolated when exact block times were unavailable, and it marks those values as approximate in the interface.
What happened when we ran it
Our sandbox installed commit 2f47870 in 37 seconds, pulling 57 Python packages and using 100 MB on disk. The build passed in another 6 seconds. Pip-audit reported 0 known vulnerabilities. We used an unprivileged Debian container with 3 CPUs, 8 GB of RAM, Python 3.12, and no secrets.
The test command exited 1 after 72 seconds. Pytest reported 97 passed, 0 failed, and 8 collection/setup errors among 105 collected tests. The log tail for test_radar.py, test_shared_session.py, and test_traders.py says Starlette's test client requires the httpx2 package and gives pip install httpx2 as the remedy. The log does not explain why that package was absent, so we will not assign a cause.
The recorded demo avoids credentials, while live mode does not
The demo fetches a compressed release asset once, unpacks its database, and serves a replay on localhost. It does not read API credentials or connect a wallet. The web interface needs Node 22 or newer, while the terminal and server use Python. A user can inspect a known route, pause the clock, and open its transaction evidence without making a live data request.
Live ingestion changes the setup. The sample environment asks for an Alchemy key, with a HyperSync token available as an optional source and a separate optional token for X mentions. External market context and holder data are disabled in demo mode. The 3D scene needs WebGL 2. The guide's verified platform was Apple silicon macOS 26.5.1, so Linux and Windows users are following an unverified path even though the commands are expected to be similar.
A visible score is still only a ranking of observations
RADAR combines recent distinct-wallet inflow, acceleration, source breadth, wallet quality, age, and curve stage. Stampede shows each part and records what happened 30 and 60 minutes after an alert on indexed trades. That makes the score auditable. It does not turn the score into evidence that a coin will rise, that addresses belong to one person, or that a preceding sale financed a later purchase.
The README says thresholds were calibrated on 5 hours from one launchpad, while the public sample covers 1 hour. It also says most spikes described in the research note ended below the alert price by the end of the measured horizon. Anyone evaluating the ranking should preserve those limits beside screenshots and exports. Removing them would convert descriptive research into a claim the project rejects.
The missing license blocks ordinary reuse
The documentation is frank that no license has been chosen and that the code is available for reading and running the demo, not redistribution. Issue 1 asks for a LICENSE file. Issue 3 requests explicit permission to port the project to TypeScript. Neither issue had a response when checked, so a public repository should not be mistaken for permission to create a derivative product.
GitHub showed 143 stars, 3 open issues, and a last code push on September 18, 2026. The v0.1.0 release arrived on September 11 with the sample and demo video. This is an early, focused research tool with unusually careful claims and a currently incomplete test environment. Its best use is inspection: follow an observed sequence back to its transactions, then stop where the evidence stops.