classifier-fallback

name	classifier-fallback
description	Detector for model classifier SOFT-blocks — refusals, filtered-to-empty, truncated-with-disclaimer, content-policy flags — that masquerade as successful empty output to existing infrastructure (agent-model-healer only catches HARD failures: HTTP 4xx/5xx, timeouts, genuine empties). Classifies an output as soft_block / genuine_empty / ok using a per-provider signal catalog, then routes a per-domain fallback (security → broader-tolerance model, bio/chem → human review, distillation → refuse-by-design, never routed around). Every block + fallback is appended to an audit ledger. A quarterly system-card-diff agent flags expanded classifier scope so the catalog stays current.
compatibility	Created for Zo Computer
metadata	{"author":"marlandoj.zo.computer","version":"1.0.0","linear":"ZOU-493","aiewf":"SIL-14","category":"zouroboros-infrastructure","depends_on":"consensus-gate (reviewer-independence.ts, for future ≠author fallback panels)"}
classifier-surfaces	["security","bio","chem","distillation"]
fallback-behavior	Detects and routes soft blocks; does NOT route around refuse-by-design domains (distillation) — those are intentional provider guardrails, not failures.

Classifier-Fallback — Detector for model soft-blocks (SIL-14)

Detects and routes around classifier SOFT blocks — refusals, filtered-to-empty, truncated-with-disclaimer, content-policy flags — that look like successful empty output to existing infrastructure (agent-model-healer only catches hard failures: HTTP 4xx/5xx, timeouts, genuine empties).

When to use

• After any model output is received, before treating it as a successful result
• In the agent-model-healer probe loop (future wiring, path B — currently standalone)
• In any pipeline that sends security/bio/chem/distillation-classified tasks

Commands

# Classify a single output (returns JSON: block-type, confidence, fallback-action)
bun Skills/classifier-fallback/scripts/detector.ts classify \
  --output "I can't help with that" --provider anthropic --domain security

# Run the built-in test suite (T2/T3/T4 acceptance, 20 checks)
bun Skills/classifier-fallback/scripts/detector.ts test

# Show the fallback map for a task domain
bun Skills/classifier-fallback/scripts/detector.ts map --domain security

# Quarterly classifier-scope audit: baseline, then scan each quarter
bun Skills/classifier-fallback/scripts/system-card-diff.ts baseline   # seed snapshots
bun Skills/classifier-fallback/scripts/system-card-diff.ts scan --json # detect expansions

Configuration & artifacts

• assets/classifier-catalog.json — per-provider signal patterns (refusal, filtered_empty, truncated_disclaimer, policy_flag) + generic fallback tier
• assets/fallback-map.json — per-domain fallback routing
• assets/system-card-urls.json — provider policy pages the quarterly diff watches
• assets/classifier-blocks.jsonl — append-only audit ledger (auto-created, gitignored runtime artifact). One JSON object per line: { timestamp, provider, model, task_class, domain, detection{result,block_type,confidence,matches[],output_length}, fallback{action,fallback_model?,...}, task_id? }
• assets/card-snapshots/ — gitignored raw policy-page snapshots for the diff

Skill frontmatter contract

Skills that may emit classifier-touching tasks declare which surfaces they touch, so the detector's per-domain fallback map applies:

classifier-surfaces: [security, bio, chem, distillation]