Tokenchaining

Each call runs N parallel shadows; a pluggable comparator (exact match, embedding cosine, AST diff, JSON shape) scores divergence, and a judge — deterministic medoid by default — resolves disagreements. Every result appends to a hash-chained SQLite audit log. It is a per-call stability detector, not an ensemble and not a guardrail.

The problem: multi-stage agent pipelines fail silently — a subtly unstable output propagates downstream before anything surfaces. Measured against a production triage pipeline: 60% surface divergence at stage one but only 15% structural; separately, 20% of runs disagreed on output cardinality. That last class is the target.

Status: v0.4 alpha, private repository; playground demo behind an access code. Results published as they stabilize.