- Intent, quality, friction, cognitive load, and regret are constructs, not facts automatically present in telemetry. Poor operational definitions create precise-looking nonsense.
- The four scoring models are not on a universal common scale. Compare cohorts only when the archetype, schema, weights, and collection process match.
- Bootstrap intervals quantify sampling variability in the supplied ledger. They do not correct selection bias, interference, novelty effects, or confounding.
userHashsupports an approximate reach calculation. It must be privacy-reviewed and rotated; this project provides no consent or identity system.- SHA-256 binds a report to the sorted-key canonical JSONL form of its analyzed sessions. It does not preserve inconsequential source whitespace/comments or prove the events were observed, complete, or untampered before collection.
- Artifact publication rejects symlinks in every existing output-path component, rechecks directory identities, stages complete sets, reconciles ambiguous rename failures, and rolls back on a publish failure. Cooperative writers serialize through a bounded five-second
.artifact-write.lock; a crash or incomplete recovery leaves that lock in place so later writers fail closed. After verifying that no writer is active, an operator must inspect any.artifact-stage-*recovery directory before manually removing a stale lock. Non-cooperating processes are not serialized. Node does not expose portable directory-file-descriptor-relative rename APIs, so a process that can concurrently replace trusted ancestor directories may still race between identity checks and filesystem operations; choose an output tree not writable by an attacker. - Explicit JSONL input paths follow the operating system's normal symlink resolution; the CLI does not claim an input-directory confinement boundary.
- The demo is synthetic and designed to exercise code paths. It is not evidence that intent-first UI has a particular real-world effect.