Research
Rubric-first research informs every foundation choice in SIS. Candidates evaluated against a locked rubric. Board-gated before ratification. Every artifact carries SIP attestation.
01Decision question + axioms + scoring dimensions written before candidates are evaluated.
02One agent per candidate. Sources cited inline. Findings dense, not sprawling.
03Decision matrix → recommendation → /starlight-board pre-pass → /bless ratification.
2 published · 1 in flight.
Seven memory architectures scored against a 5-axiom 10-dimension rubric. Letta MemFS and LangGraph + LangMem emerge as the head-to-head Phase 0 dog-food candidates. Anthropic Memory API rejected on model lock-in. The incumbent currently fails its own filesystem-native axiom.
Benchmark survey for SIS 3D substrate scene — Apple Liquid Glass, Linear, Bruno Simon, Active Theory, NASA Eyes, Heptabase. Twelve references, seven principles, three code-level techniques.
Living head-to-head LLM eval surface, run natively in Claude Code. R1: correctness parity; Fable 5 edge = instruction compliance. R2 stress card (governance traps, injection, lying docs): Fable 5 = precision instrument, Opus 4.8 = judgment instrument. Receipts in-repo.
Why a research surface
The wrong memory architecture, the wrong attestation surface, the wrong cross-model bridge — these errors don't stay contained. They propagate into every Intelligence System, every agent, every session. So substrate decisions get evidence before they get committed.
Research lives here, in the open, because forks of SIP inherit the decisions. The reasoning has to travel.
Methodology in the open. Decisions earned.