Arnav.S*k-3

Kuttaka 3 is an attempt at a systematic cognitive substrate around a fixed model: the weights stay unchanged while a persistent workspace accumulates methods, checks, and the evidence behind them. Later tasks can retest and reuse that work. The runtime checks proposed tests, generates or composes candidate methods, uses disagreements to guide repair, and can withhold an answer when verification remains unresolved. Here, self-awareness means recognizing weak or conflicting evidence. The submission includes a Python CLI, HTTP/MCP server, method library, Docker sandbox, judge quickstart, and three engine editions. The packaged default suite previously passed 1,248 tests with Docker enabled. In the recorded W4-o development run, Meta Muse Glimmer alone solved 8/12 tasks in 12 model calls; Glimmer with Kuttaka solved 12/12 in 53 calls, with 79/79 private cases passing on rescore and zero wrong answers delivered. A separate growth run solved 12/12 with memory versus 11/12 with a fresh workspace, using 69 and 52 calls respectively.