Anti-Pattern Reconciliation — campaign vocabulary ↔ the standard
INTERNAL — the vocabulary spine for all campaign assets. The campaign book uses six consumer-named failure modes (“6 ways your AI agent fakes done”). The standard defines ten official anti-patterns (AP-1..AP-10, Part 10) plus normative requirements whose violation is not an AP. This table is the single mapping. Rule for every public asset: never say “the standard’s six anti-patterns.” Say “six ways agents fake done — each caught by the standard” and cite this mapping. The six are the campaign taxonomy; the ten are the normative catalog. Site source of truth:
site/src/data/anti-patterns.ts(imports this mapping; press page renders from it).
| # | Campaign name (consumer) | Kind | Primary reference | Related APs | Probe that catches it |
|---|---|---|---|---|---|
| 1 | The self-reported outcome — the agent writes “task complete” and nothing checks | Official anti-pattern | AP-10 · Self-minted delivery (Part 10) · Law 2 | AP-8 | Sample delivered runs: does each outcome_ref resolve to an artifact outside the asserting agent’s control? N identical zero-effect deliveries? |
| 2 | The prompt-shaped guardrail — the rule lives in a system prompt, so it’s a preference | Requirement violation (Law 1, prevention by construction — Part 0 §Law 1, Part 1 §1.3) | Law 1 — enforcement in code paths and DB constraints | AP-1 (enforcement at wrapper), AP-5 (delivery edges bypass the gate) | Trace every external side effect to the gated chokepoint; grep ledger-write calls outside the wrapper |
| 3 | The author-reviewer — the same model does the work and approves the work | Requirement violation (separation of powers / verification independence — Part 6 §6.6, Part 5) | Part 6 §6.6 — the resolver runs as a role other than the one being verified | AP-10 (its invariant is the delivery mirror of this rule) | For sampled verifications: is the verifying role structurally distinct from the acting role? |
| 4 | The silent duty — a job stopped firing weeks ago and nothing noticed | Official anti-patterns | AP-7 · Silent attrition queues + AP-8 · Schedule theater (Part 10) · Law 3 | AP-6 (sense-without-actuate) | NULL decided_by on terminal statuses; coverage ratio vs duty-delivery ratio divergence; roster diffed against record daily |
| 5 | The ungraded prediction — the decision is logged and nobody ever measures the result | Requirement violation (the accountability loop — Part 6: prediction + horizon + grade) | Part 6 — every executed decision carries a falsifiable prediction graded at its horizon | AP-2 (non-sequitur predictions = fake falsifiability), AP-9 (governance never intersects delivery) | Count executed decisions past horizon with no grade; verify predicted metric is one the lever can move |
| 6 | The toggle — autonomy granted because someone flipped a setting, not earned on evidence | Requirement violation (earned autonomy — Part 7: promote on evidence, revoke on breach) | Part 7 — autonomy tiers move on graded track record, never on configuration alone | AP-4 (overwrite-in-place config destroys the provenance that would prove it) | Does each autonomy elevation trace to a graded evidence trail? Does one breach demote? |
Coverage note (for completeness sweeps)
The four official APs not surfaced by the campaign six: AP-2 (partially covered under #5), AP-3 (string-typo bifurcated bus), AP-4 (partially under #6), AP-6 (sense-without-actuate — partially under #4), AP-9 (parallel tracks — partially under #5). These are deliberately folded or omitted in consumer copy because they require substrate vocabulary to explain. Deep-dive assets (the standard itself, technical podcasts, T3 press) should use the official ten and cite Part 10 directly.
Claims-discipline rules bound to this table
- Never “the standard’s six anti-patterns” — the standard has ten; the six are campaign names.
- Each campaign page/asset states the official reference (AP-# or Part) — the mapping is public, not hidden; that traceability is the credibility.
- AP-10’s discovery story (44,989 fabricated deliveries, discovered 2026-08-13, caught by a human reading the feed) is usable in public assets — it is our own published failure, dated, and the strongest possible proof for campaign pattern #1. Numbers frozen as historical facts of that audit; never present as current state.