STRATA.

Strata — an AI archaeologist for software whose source is lost. Software outlives its code. Games lose their servers, devices die with the company that made them, and legacy systems keep running long after anyone can read them. Often the only thing left is how the program behaves. Strata runs the old program as a black box, sifts through stale docs, partial code and runtime traces, rebuilds it in a new language, and proves the rebuild behaves the same. Every claim is labeled observed or inferred and backed by evidence. Why not just ask an LLM? LLMs guess, and the guesses look like working code. Decompiled code compiles 99% of the time but behaves correctly far less often, and one in five “solved” benchmark patches is wrong. On famous software a model simply recites what it remembers. Strata replaces recall with evidence. How it works Survey: reads the manual and any partial code into low-trust claims and ranks what to investigate. Check: tests those claims against recorded games and logs every contradiction. Probe: runs designed experiments on the live binary, predicts each move first, and records surprises. Hypothesize: proposes rules as executable code and scores them on held-out games it never trained on. Rebuild: writes a JavaScript engine from confirmed rules only. Verify: replays and fuzzes both versions side by side, then accepts or blocks the rebuild with the shortest counterexample. Models propose. Code decides. Every finding goes into an evidence ledger, built in layers like rock strata: raw traces, then what was observed, then what is inferred, then what is confirmed. A claim without trace IDs is rejected, and only deterministic code can promote a rule to Confirmed. Proving it learned instead of recalled. The black box is Pressure, an original card game we wrote, with hidden state, chained effects and scripted randomness. Every run plants secret rule changes, so reciting any known game fails visibly. In our demo the manual says Vent cools pressure by 3, twelve recorded moves say 4, and Strata catches the planted change from evidence alone. Opaque-token runs, held-out games and zero-probe ablations back this up. Built with Meta Muse Spark writes hypotheses, experiments and the rebuild. SubQ Snap makes 384K-token decisions over all the evidence: which experiment to run next, surprise triage and claim audits. Fetch.ai uAgents let you question Strata in ASI:One. Plain, deterministic code makes every final call. AI-written code runs sandboxed, with no keys. What you’ll see: a live table where cards fly and players get knocked out, plus a “Behind the scenes” view that shows, in plain words, what Strata saw, what it guesses and what it has proven. Strata. Software that outlived its creator — rebuilt, with proof.