The Augmentation Layer

The runtime that enforces substrate-bundle integrity automatically — describe the workflow's shape in one sentence and the canonical primitive bundle is applied.

The problem it solves

The substrate criterion (proven empirically via the 5-phase ladder) says certain primitives must travel together to preserve calibration, auditability, and adaptability. Calling primitives one-at-a-time risks degrading the bundle — skipping a required primitive for cost reasons, or forgetting to extract sources before scoring source credibility.

The augmentation layer enforces this contract at runtime. Required primitives are CONTRACTS, not suggestions. Optional primitives are caller opt-in. The augmentation report makes the bundle visible to downstream consumers so they know what to surface.

The 4-layer architecture

Layer 1 — Universal Augment Endpoint  (POST /api/v1/augment)
   Drops into any HTTP-capable platform. Foundation for all others.

Layer 2 — Workflow-Shape Recipe Library  (7 canonical shapes + bundles)
   summarization / enrichment / monitoring / decision / generation /
   routing / evaluation + generic fallback.

Layer 3 — Judgment Co-Pilot  (Claude Code template + Anthropic Skill)
   Reads workflow.json, identifies augmentation points, rewrites with
   augmentation injected. Distributed via existing Skills + Claude
   Code template infrastructure.

Layer 4 — Per-Platform Native Blocks
   Importable templates (n8n / Zapier / Agent Designer / Lindy) that
   wrap Layer 1 in platform-native auth + UX.

The 7 canonical workflow shapes

Each shape encodes a substrate-criterion contract — required primitives cannot be skipped regardless of cost optimization.

Shape	When	Required primitives	Cost
`summarization`	Text → text summary	extract_claims, extract_sources, score_probability_language	11 credits
`enrichment`	Record → record + context	extract_claims, score_source_credibility_nato	6 credits
`monitoring`	Current+prior → alert	compute_analytic_delta, extract_information_gaps, check_regression_robust	14 credits
`decision`	Question + options → recommendation	compose_decision_frame, generate_alternatives, extract_information_gaps, gate_with_confidence	18 credits
`generation`	LLM-generated content	extract_claims, score_probability_language, extract_assumptions	13 credits
`routing`	Classification	compose_decision_frame, generate_alternatives, gate_with_confidence	13 credits
`evaluation`	Predictions vs outcomes	score_brier, check_regression_robust, compute_beta_binomial_ci	3 credits
`generic`	Fallback when no shape clearly fits	extract_claims, extract_sources, extract_alternatives, extract_information_gaps, score_probability_language	22 credits

Minimal example

curl -X POST https://api.leadershipunderuncertainty.org/api/v1/augment \
  -H "Authorization: Bearer nb_..." \
  -H "Content-Type: application/json" \
  -d '{
    "value": "Reuters reports Tencent reduced its OpenAI API spend by 60%...",
    "intent": "Summarize the press item for the executive team morning brief.",
    "workflowShape": "summarization",
    "options": { "confidenceGate": 0.6, "attestation": "hash" }
  }'

# Returns:
# {
#   "value": "<original input, unchanged>",
#   "augmented": {
#     "claims": [{"text": "Tencent reduced its OpenAI API spend...", "confidence": 0.7}],
#     "sources": [{"identifier": "reuters_tencent_openai", "type": "news_article"}],
#     "probabilityLanguage": {"score": 70, "issues": []}
#   },
#   "report": {
#     "workflowShape": "summarization",
#     "primitivesExecuted": [...],
#     "substrateCoverage": {"populated": ["Claim","Source","UncertaintyEstimate"], "score": 3},
#     "totalCreditCost": 11,
#     "confidence": {"mean": 0.7, "lowConfidenceFields": []},
#     "downstreamGuidance": "...",
#     "issues": []
#   },
#   "attestation": {"mode": "hash", "hash": "8b9c5c7a..."}
# }

Substrate criterion enforcement

If report.issues is non-empty, the substrate criterion is broken — typically because a required primitive failed. Surface the issue to the consumer before treating the augmented output as authoritative.

The augmentation report's substrateCoverage.score (0-9) tells you which of the 9 substrate primitives the run populated. A score of 3-4 is typical for narrow shapes (summarization, enrichment), 6-9 for broad shapes (decision, generic).

Distribution surfaces

REST — POST /api/v1/augment (this page).
MCP — luu_augment tool (Claude Desktop, Cursor, ChatGPT, custom MCP clients).
SDK — Augment class in the TypeScript SDK at packages/sdk (in-repo; not yet published to npm) with summarize(), enrich(), decide() wrappers.
Skills — packages/luu-skills/skills/judgment-augment/SKILL.md for Claude Code / Claude.ai / Claude Desktop / Cursor / ChatGPT (when MCP-enabled).
n8n — importable workflow at packages/n8n-templates/judgment-augment/.
Zapier — recipe at packages/zapier-templates/judgment-augment/.
Google Agent Designer — flow template at docs/distribution/flow-templates/judgment-augment.json.
Lindy — install guide at docs/distribution/install-judgment-augment-lindy.md.
Claude Code template — Judgment Co-pilot at packages/claude-code-templates/judgment-copilot/ reads any workflow.json and rewrites it with augmentation injected.