Workshop: Error-Catching Systematisation

Question

What is the space of error-catching techniques in an agent-operated KB — organized so that gaps become visible, each technique's economics are explicit, and the roles of gates, validators, critique, telemetry, and the human operator are placed in one scheme rather than existing as separate lore?

Provenance

Started 2026-07-18 from the distillation-migration retrospective (rename-lessons) and the conversation that followed it: an operator-caught category error (garbling misfiled as staleness) led to the activation-gap reading of why gates work, then to gate catch-rate statistics, then to the observation that every investigative loop in the session closed through the operator even though the triggering signals were already persisted. The systematisation is the attempt to put those pieces into one frame.

Threads

  • systematisation.md — the draft scheme: two axes (lens breadth × detection site), the technique catalogue with economics, the governing claims, and the gaps the grid exposes
  • gate-stats.md — first execution of the telemetry row (2026-07-18): the activation gap quantified (29.6% first-encounter catch, 67% peak, mechanical lenses catch 2–3× judgment lenses), the fix loop's 79% one-cycle resolution, within-partition absorption after the April sweep, the luna/sol 2.5× reviewer-variance finding, and three coverage-bet failures awaiting adjudication
  • calibration-run-01.md — seeded-violation calibration (2026-07-18): 8 gates × (4 seeds + 2 controls), blind reviewers; 30/30 valid seeds caught, zero clear false positives — recall is not the bottleneck, catch rate ≈ incidence, title-as-claim adjudicated absorbed-not-blind; two new second-order guards surfaced (seeder self-check; verdict-block content binding). Raw artifacts under run-01/

What closes this workshop

  1. The scheme survives contact with the existing catalogue: every current technique (validators, gates, conformance pairs, critique, friction gate, operator review) places cleanly, and every governing claim either cites existing notes or is marked as this workshop's conjecture.
  2. A decision per exposed gap: build (proposal in kb/reference/proposals/), defer (named in the systematisation), or reject.
  3. The durable claims promoted — likely one structure note (detection as a two-layer execution system with promotion-by-recurrence) and the technique grid either folded into it or kept as reference material.

Bookkeeping

Workshop register. Gate catch-rate numbers cited from the commonplace store as of 2026-07-18 (9,286 completed verdicts since 2026-03-13); re-run the queries to refresh.

Complete file listing (generated at build time)