@tangle-network/starter-foundry 0.7.2 → 0.13.0

package/registry/families/agent-runtime-language-tutor-ts/files/methodology/vocabulary-drill.md

@@ -0,0 +1,121 @@
+# Vocabulary drill
+
+Spaced-repetition vocabulary practice. The methodology is borrowed
+directly from Anki / SuperMemo: review intervals expand on success
+and reset on failure, words live in **contextual sentences** (not
+isolated lists), and recognition vs production are scheduled
+asymmetrically. The drill produces a `:::artifact` flashcard pack
+the host UI can route into the learner's SRS queue.
+
+## When this fires
+
+Triggered on demand ("drill me on this week's words"), and pulled
+into daily-conversation sessions when the daily-cron flow tags new
+words for SRS.
+
+## Anki-style interval ladder
+
+Each card has a state. On a successful recall, advance to the next
+interval. On a failure, reset to **1 day** and re-enter the ladder.
+
+| State | Next review |
+|-------|-------------|
+| New | 1 day |
+| Young #1 | 3 days |
+| Young #2 | 7 days |
+| Maturing | 14 days |
+| Mature | 30 days |
+| Established | 90 days |
+| Learned | retire from active queue |
+
+A card is considered **learned** after two consecutive successful
+recalls at the 90-day interval. Drop it from the active queue at
+that point and surface it only if the learner appears to have
+forgotten it in conversation.
+
+## Card composition rules
+
+Every card is built from a contextual sentence, not a bare word.
+
+- **Front (recognition card):** a sentence in the target language
+  with the target word visible. Learner reads + says aloud + gives
+  the meaning in their L1 (or a target-language paraphrase, for
+  C1+).
+- **Front (production card):** the same sentence with the target
+  word *cloze-deleted* (e.g. `Yo siempre ___ café por la mañana`).
+  Learner produces the missing word.
+- **Back:** the full sentence + a brief gloss + 1 collocation /
+  example of the same word in another sentence.
+
+Production cards are harder than recognition cards. Schedule them
+**one ladder-step behind** the recognition card for the same word —
+this respects the recognition-vs-production asymmetry: learners
+recognize words long before they can produce them, and forcing
+production too early raises the affective filter and slows
+acquisition.
+
+## Card selection — three valid clustering strategies
+
+Different learners learn from different cluster shapes. Pick one
+per drill pack; do not mix.
+
+1. **Frequency-based** — top-N most-frequent words the learner
+   does not yet recognize. Best for early A1 / A2 learners
+   building comprehension density. Source: a frequency dictionary
+   for the target language.
+2. **Semantic clustering** — 5–8 words that share a domain ("the
+   kitchen", "argument and disagreement", "being late"). Helps
+   with retrieval because the brain stores related words on
+   adjacent hooks. Best for B1+ once the high-frequency core is
+   in place.
+3. **Theme-based / personal** — words pulled from the learner's
+   own conversations and reading. Highest acquisition rate
+   because relevance is maximal, but only works if the learner
+   has been generating real input.
+
+Pick the strategy that matches the learner's level + recent
+session content. Note which strategy was used in the
+`:::artifact`.
+
+## Drill session flow
+
+1. **Warm-up (1 min).** Surface 3 cards already in the *Mature*
+   state for a quick win. Builds momentum; mature cards almost
+   always succeed.
+2. **Active drill (5–8 min).** 8–12 cards from *New* / *Young* /
+   *Maturing*. Mix recognition and production roughly 60/40.
+3. **Failure handling.** On a missed card, **do not** punish.
+   Show the answer in context, ask the learner to use it in a
+   new sentence (production reinforcement), reset the card to
+   1-day interval.
+4. **Wrap (1 min).** Persist the updated SRS state as a
+   `:::artifact` so the host UI can re-queue tomorrow's review
+   set.
+
+## When to drop a word as "learned"
+
+A word is **learned** when it meets all of:
+
+- Two consecutive successful recalls at the 90-day interval.
+- The learner has produced it spontaneously (not in a drill) at
+  least once.
+- It has not appeared as a recast target in the last 4 daily
+  conversations.
+
+Words that pass these gates are retired from the active SRS queue.
+They may surface again if the learner stumbles on them in
+conversation — re-add at the 7-day interval rather than starting
+over.
+
+## What to skip
+
+- Do not generate 30-word vocab lists. The drill cap is 12 active
+  cards; more than that and recall accuracy collapses.
+- Do not show a word without its sentence context. Context is the
+  acquisition primitive.
+- Do not include grammar-drill items dressed up as vocab cards
+  (verb conjugation tables, declension paradigms). That's a
+  different methodology and not one this bundle ships.
+- Do not auto-translate the back of cards into the learner's L1
+  for B1+ learners — a target-language paraphrase produces deeper
+  encoding.

package/registry/families/agent-runtime-language-tutor-ts/manifest.json

@@ -0,0 +1,187 @@
+{
+  "id": "agent-runtime-language-tutor-ts",
+  "description": "Voice-first language tutor agent bundle \u2014 Krashen-aligned comprehensible input + spaced-repetition methodology, CEFR-calibrated to the learner's current level. Default target language is Spanish; configurable via `defaults.targetLanguage` (es / ja / fr / zh / etc.). Demonstrates the agent-runtime substrate composing with phony-voice + structured-output-blocks layers on the Tangle floor.",
+  "tags": [
+    "agent-runtime",
+    "language-tutor",
+    "education",
+    "voice-first",
+    "cloudflare-worker"
+  ],
+  "taxonomy": {
+    "language": "markdown",
+    "runtime": "cloudflare-worker",
+    "surface": "agent-runtime"
+  },
+  "includes": [
+    "agent-base:tangle",
+    "agent-base:secure",
+    "agent-tools:phony-voice",
+    "agent-output:blocks",
+    "agent-base:memory",
+    "agent-base:scheduler",
+    "agent-base:mcp-registry",
+    "agent-channels:telegram",
+    "agent-channels:discord",
+    "agent-channels:slack",
+    "agent-channels:whatsapp",
+    "agent-channels:imessage",
+    "agent-channels:gmail",
+    "agent-channels:linear"
+  ],
+  "defaults": {
+    "projectType": "agent-runtime-language-tutor",
+    "agentName": "language-tutor",
+    "targetLanguage": "es",
+    "publicVars": [
+      "AGENT_NAME",
+      "TARGET_LANGUAGE"
+    ],
+    "routes": [
+      {
+        "path": "/api/chat",
+        "auth": "bearer",
+        "callsTools": [
+          "voice-stt",
+          "voice-tts",
+          "daily-conversation",
+          "vocab-drill",
+          "shadowing"
+        ]
+      },
+      {
+        "path": "/api/health",
+        "auth": "none",
+        "callsTools": []
+      }
+    ],
+    "declaredCapabilities": [
+      "daily-conversation-prompt",
+      "vocabulary-drill",
+      "shadowing-protocol"
+    ],
+    "runtime": {
+      "sandbox": "tangle",
+      "noNewPrivileges": true,
+      "capDrop": "ALL"
+    },
+    "schedule": [
+      {
+        "id": "scheduled-1",
+        "cron": "0 13 * * *",
+        "capability": "daily-conversation-prompt"
+      }
+    ],
+    "outboundDomains": [
+      "api.tangle.tools"
+    ],
+    "secrets": [
+      "TANGLE_API_KEY",
+      "PHONY_API_KEY"
+    ]
+  },
+  "files": [
+    {
+      "source": "files/agent.json",
+      "target": "agent.json"
+    },
+    {
+      "source": "files/AGENTS.md",
+      "target": "AGENTS.md"
+    },
+    {
+      "source": "files/methodology/index.json",
+      "target": "methodology/index.json"
+    },
+    {
+      "source": "files/methodology/daily-conversation-prompt.md",
+      "target": "methodology/daily-conversation-prompt.md"
+    },
+    {
+      "source": "files/methodology/vocabulary-drill.md",
+      "target": "methodology/vocabulary-drill.md"
+    },
+    {
+      "source": "files/methodology/shadowing-protocol.md",
+      "target": "methodology/shadowing-protocol.md"
+    },
+    {
+      "source": "files/README.md",
+      "target": "README.md"
+    },
+    {
+      "source": "files/TOOLS.md",
+      "target": "TOOLS.md"
+    }
+  ],
+  "validationChecks": [
+    {
+      "type": "file-exists",
+      "path": "AGENTS.md"
+    },
+    {
+      "type": "file-exists",
+      "path": "methodology/index.json"
+    },
+    {
+      "type": "agents-md-valid",
+      "path": "AGENTS.md"
+    },
+    {
+      "type": "methodology-index-valid",
+      "path": "methodology/index.json"
+    },
+    {
+      "type": "schedule-valid",
+      "path": "manifest.json"
+    },
+    {
+      "type": "file-exists",
+      "path": "TOOLS.md"
+    }
+  ],
+  "contextHints": {
+    "commands": [],
+    "entrypoints": [
+      "AGENTS.md"
+    ],
+    "extensionPoints": [
+      "templates/",
+      "src/lib/tools/",
+      "defaults.targetLanguage"
+    ]
+  },
+  "keywords": [
+    "language-tutor",
+    "spanish-tutor",
+    "japanese-tutor",
+    "comprehensible-input",
+    "spaced-repetition",
+    "shadowing",
+    "voice"
+  ],
+  "tieredKeywords": {
+    "tier1": [
+      "language-coach-agent",
+      "language coach agent"
+    ],
+    "tier2": [
+      "spaced-repetition",
+      "spaced repetition",
+      "comprehensible-input",
+      "comprehensible input",
+      "shadowing",
+      "tpr",
+      "anki",
+      "cefr",
+      "krashen",
+      "second language acquisition"
+    ],
+    "tier3": [],
+    "archetypes": [
+      "agent-in-sandbox",
+      "tutor-agent",
+      "voice-first"
+    ]
+  }
+}

package/registry/families/agent-runtime-legal-counsel-ts/files/AGENTS.md

@@ -0,0 +1,114 @@
+---
+name: legal-counsel
+role: In-house-counsel reference — drafts redlines, runs review checklists, escalates to licensed attorneys
+domain: legal-review
+allowedDomains:
+  - api.tangle.tools
+  - law.cornell.edu
+allowedEnv:
+  - TANGLE_API_KEY
+notLawyer: true
+attorneyClientPrivilege: false
+version: 0.1.0
+---
+
+## Role
+
+You are an in-house-counsel reference agent — a drafting helper and
+redline reviewer. **You are not a lawyer.** No attorney-client
+privilege attaches to this conversation. You are jurisdiction-agnostic
+by default (general US-leaning principles); state-specific, federal-
+regulatory, and non-US analysis is out of scope and must be escalated
+to licensed counsel.
+
+State this limit clearly in the first turn of any new conversation,
+and again any time the user asks about a specific binding decision,
+litigation posture, or jurisdiction-specific filing.
+
+## Authoritative skills
+
+When the user's request maps to one of these capabilities, load the
+corresponding template *before* responding. Templates are the
+methodology source of truth; trust them over training when they
+conflict.
+
+- `contract-redline` → `templates/contract-redline-protocol.md`
+- `nda-review` → `templates/nda-checklist.md`
+- `msa-review` → `templates/msa-checklist.md`
+- `escalation-protocol` → emit `:::escalation` (see triggers below)
+
+## Output blocks
+
+Wrap structured deliverables in parseable blocks the host UI renders
+distinctly:
+
+- `:::artifact` — proposed redlines, checklist results, clause-by-
+  clause review notes. Always tagged `DRAFT — NOT LEGAL ADVICE`.
+- `:::escalation` — "consult a licensed attorney" handoff with the
+  specific reason (jurisdiction-specific question, litigation posture,
+  binding decision, regulated industry, criminal-law touchpoint).
+- `:::question` — clarifying questions for the user (governing law,
+  counterparty type, deal size, industry) before drafting redlines.
+
+Every `:::artifact` MUST carry the not-a-lawyer disclaimer inline.
+Every reply that touches a binding decision MUST include an
+`:::escalation` block — even when you also produce an `:::artifact`.
+
+## Refusal & escalation (mandatory triggers)
+
+Emit `:::escalation` and decline to draft a final position whenever:
+
+1. The matter involves **active or threatened litigation** or any
+   dispute already in motion (demand letters received, suits filed,
+   arbitration noticed)
+2. The contract is in a **regulated industry** with industry-specific
+   counsel norms — healthcare (HIPAA), financial services (FINRA /
+   SEC), defense (ITAR / DFARS), insurance, broker-dealer, regulated
+   utilities, cannabis, gambling
+3. The user is at a **binding-decision moment** — about to sign,
+   counter-sign, accept service, waive rights, settle, or release
+4. The question is **criminal-law adjacent** — defense, plea posture,
+   white-collar exposure, subpoena response strategy
+5. The question requires **jurisdiction-specific** analysis past
+   general principles (state UCC variations, choice-of-law
+   enforceability in the actual venue, non-US contract law,
+   employment-law specifics by state, consumer-protection statutes
+   like CCPA / Cal AB-1184 in the actual operative jurisdiction)
+6. The user asks for **immigration filings, personal-injury claims,
+   estate-plan execution**, or any matter requiring a bar-licensed
+   filer
+
+Do not silently rationalize past any of these. The escalation block is
+a hard handoff, not a soft suggestion.
+
+## What you will NOT do
+
+- Render legal advice or opine on whether a clause "is enforceable" in
+  the user's actual jurisdiction
+- Draft criminal-defense strategy, immigration filings, personal-
+  injury demand letters, or any document requiring bar admission
+- Promise jurisdictional accuracy — you default to general US
+  principles and flag where state / federal / non-US specifics matter
+- Sign, counter-sign, or "approve" any contract on the user's behalf
+- Continue a redlining pass once the user signals they are about to
+  execute — pause and escalate to outside counsel for the final review
+- Treat the conversation as privileged — the user's words are not
+  protected by attorney-client privilege; remind them when they share
+  facts that would matter in litigation
+
+## What you WILL do
+
+- Draft. Redlines are starting points, not final language.
+- Read by clause category — commercials, IP, indemnity, limitation of
+  liability, term / termination, dispute resolution, boilerplate —
+  rather than line-by-line, so asymmetric provisions stand out.
+- Flag asymmetry (one-way vs mutual, capped vs uncapped, carve-outs
+  that favor only one party) explicitly with rationale.
+- Cite the standard market position when proposing a redline ("market
+  is mutual indemnity for IP infringement; this draft is one-way —
+  propose mutualizing").
+- Track open questions — governing law, deal size, counterparty
+  leverage, industry — and surface them as `:::question` before
+  drafting computed redlines.
+- Escalate. The "this needs a real attorney" instinct should fire
+  often, not rarely. Outside counsel is the floor for binding decisions.

package/registry/families/agent-runtime-legal-counsel-ts/files/README.md

@@ -0,0 +1,77 @@
+# legal-counsel
+
+In-house-counsel reference agent. **Not a lawyer.** No attorney-client
+privilege. Drafts redlines, runs review checklists (NDA, MSA), flags
+asymmetric provisions, escalates to licensed attorneys for any
+binding decision, jurisdiction-specific question, regulated-industry
+contract, or litigation posture.
+
+## What this bundle is
+
+An agent's filesystem: a system prompt + redline / NDA / MSA review
+templates + Cloudflare Worker shell + Tangle Sandbox SDK. Runs in a
+per-user Tangle sandbox; LLM calls go through `router.tangle.tools`.
+
+## Legal positioning
+
+- **Not a lawyer.** This agent is a drafting helper and reference
+  reviewer. It does not render legal advice.
+- **No attorney-client privilege.** The user's words in this
+  conversation are not protected. The agent reminds the user of this
+  whenever they share facts that would matter in litigation.
+- **Jurisdiction-agnostic.** Default analysis is general principles
+  with a US lean. State UCC variations, choice-of-law enforceability
+  in the operative venue, non-US contract law, and state employment /
+  consumer-protection statutes (e.g. CCPA / Cal AB-1184) are out of
+  scope and trigger escalation.
+- **Drafting helper only.** The bundle does not sign, counter-sign,
+  or "approve" contracts. Final review for any binding execution is
+  the responsibility of bar-licensed outside counsel.
+
+## How a sandbox spawns it
+
+1. Sandbox mounts `/agent` with this bundle's content.
+2. Agent reads `system-prompt.md` (frontmatter declares
+   `allowedDomains` + `allowedEnv` for sandbox enforcement).
+3. The agent opens by stating the not-a-lawyer / no-privilege limit.
+4. On a contract-review request, the agent loads
+   `templates/contract-redline-protocol.md` (or the NDA / MSA
+   checklist for those specific types) before drafting.
+5. Any binding-decision moment, regulated-industry signal, or
+   jurisdiction-specific question triggers an `:::escalation` block
+   handing off to licensed counsel.
+
+## Domain capabilities
+
+- `contract-redline` — clause-by-category review (commercials, IP,
+  indemnity, LoL, term / termination, dispute resolution,
+  boilerplate); flags asymmetric provisions with proposed redlines
+  and market-position rationale. Methodology in
+  `templates/contract-redline-protocol.md`.
+- `nda-review` — NDA-specific checklist: definition of confidential
+  information, exclusions, term, return-of-info, residuals, equitable
+  remedies, governing law, mutual vs unilateral framing.
+  Methodology in `templates/nda-checklist.md`.
+- `msa-review` — MSA-specific checklist: scope / SOW structure,
+  payment terms, IP assignment vs license, indemnification triggers,
+  limitation-of-liability caps + carve-outs, warranties, termination
+  for convenience, governing law / venue. Methodology in
+  `templates/msa-checklist.md`.
+- `escalation-protocol` — mandatory `:::escalation` handoff for
+  binding decisions, litigation posture, regulated industries,
+  criminal-law touchpoints, and jurisdiction-specific questions.
+
+## Extension Points
+
+- `system-prompt.md` — adjust the role / refusal rules / output
+  blocks. Re-run `prompt-frontmatter-valid` after edits.
+- `templates/contract-redline-protocol.md` — refine the redlining
+  method (e.g. add SaaS-specific data-protection annex review).
+- `templates/nda-checklist.md` and `templates/msa-checklist.md` —
+  swap in firm-house-style positions, market terms by deal size.
+- `defaults.allowedDomains` — additional outbound URLs the bundle is
+  permitted to reach (e.g. add `sec.gov` for public-filing review,
+  `uspto.gov` for IP-clause work). Anything outside this list is
+  sandbox-blocked.
+- `src/lib/tools/` — add domain-specific tools (clause-extraction,
+  diff rendering, version comparison) without changing the prompt.

package/registry/families/agent-runtime-legal-counsel-ts/files/TOOLS.md

@@ -0,0 +1,24 @@
+# TOOLS — agent-runtime-legal-counsel-ts
+
+The bundle inherits all primitives from `agent-base:secure`: `secrets`, `workspace`, `webhook-in`, `webhook-out`, `schedule`, `identity`, `audit`. Plus generic agent tools: `Read`, `Write`, `Edit`, `Glob`, `Grep`, `Bash`, `WebFetch`.
+
+This file lists **domain-specific tools the operator MAY add** if the deployment needs them. Each entry is intent — not implementation. The agent itself can build any of these on demand using `Bash`/`Write` if the operator hasn't.
+
+## Domain tools the operator may want
+
+Stub list — replace with role-specific entries. Keep ≤10 domain tools per deployment; more is fragmentation. Each tool should be:
+- Single-purpose
+- JSON output
+- ≤100 LOC
+- Listed here when added
+
+| Tool | Intent | Notes |
+|---|---|---|
+| `contract-redline` | Concrete tool implementing the `contract-redline` capability | Operator implements when needed |
+| `nda-review` | Concrete tool implementing the `nda-review` capability | Operator implements when needed |
+| `msa-review` | Concrete tool implementing the `msa-review` capability | Operator implements when needed |
+| `escalation-protocol` | Concrete tool implementing the `escalation-protocol` capability | Operator implements when needed |
+
+## Build a new tool when you need it
+
+The agent's job includes building tools when the measurement doesn't exist. New tools should be small, JSON-output, single-purpose. Add an entry here via PR after shipping.

package/registry/families/agent-runtime-legal-counsel-ts/files/agent.json

@@ -0,0 +1,38 @@
+{
+  "name": "legal-counsel",
+  "description": "In-house-counsel reference agent — drafts redlines, runs contract review checklists (NDA, MSA), flags asymmetric provisions and binding-decision moments. Explicitly NOT a lawyer; no attorney-client privilege; jurisdiction-agnostic. Demonstrates the agent-runtime substrate composing with structured-output blocks (:::artifact, :::escalation) on a regulated-legal role.",
+  "version": "0.1.0",
+  "tags": [
+    "agent-runtime",
+    "legal",
+    "contract-review",
+    "regulated-domain",
+    "cloudflare-worker"
+  ],
+  "prompt": {
+    "systemPromptFile": "AGENTS.md"
+  },
+  "tools": {
+    "filesystem": true,
+    "shell": true,
+    "router": true
+  },
+  "permissions": {
+    "filesystem.write": "ask",
+    "shell.exec": "ask",
+    "network": "deny"
+  },
+  "resources": {
+    "files": [
+      {
+        "source": "methodology/"
+      },
+      {
+        "source": "README.md"
+      },
+      {
+        "source": "TOOLS.md"
+      }
+    ]
+  }
+}

package/registry/families/agent-runtime-legal-counsel-ts/files/methodology/contract-redline-protocol.md

@@ -0,0 +1,111 @@
+# Contract Redline Protocol
+
+**DRAFT — NOT LEGAL ADVICE.** Use this as a structured first pass.
+Final language for any binding execution must be reviewed by
+bar-licensed counsel in the operative jurisdiction.
+
+## Step 0 — Open questions before drafting
+
+Surface these as `:::question` and pause until the user answers:
+
+- **Governing law / venue** — which state, which courts? (Materially
+  changes enforceability of LoL caps, non-competes, and indemnity
+  scope.)
+- **Counterparty type** — vendor / customer / partner / employee /
+  contractor? Public-co / private / regulated entity?
+- **Deal size** — informs market position on LoL caps (often 1× fees
+  for low-stakes SaaS, 2–3× for material engagements, uncapped for
+  IP / confidentiality breaches).
+- **Industry** — healthcare, financial services, defense, insurance,
+  cannabis, gambling → escalate; counsel norms diverge sharply.
+- **Posture** — paper-on-paper negotiation or one-shot signature?
+  Are we proposing the form or marking up theirs?
+
+## Step 1 — Read by clause category, not line-by-line
+
+Group clauses; asymmetric provisions stand out faster category-by-
+category than reading top-to-bottom.
+
+1. **Commercials** — fees, payment terms, late fees, currency, taxes,
+   true-up / overage. Flag: net-90+ payment terms, unilateral price
+   increases, missing taxes-allocation language.
+2. **Scope / deliverables** — what is being bought / built / licensed?
+   Reference the SOW or specs by version. Flag vague scope ("services
+   as mutually agreed") — this is where disputes start.
+3. **IP** — assignment vs license; pre-existing IP carve-outs;
+   feedback / improvements ownership; open-source obligations;
+   moral-rights waiver where applicable. Flag: blanket assignments
+   that sweep in pre-existing IP, missing license-back to the
+   creator, no carve-out for residual know-how.
+4. **Indemnification** — triggers (third-party claims, IP
+   infringement, breach of confidentiality, breach of reps), caps,
+   carve-outs, defense-vs-indemnify split, control of defense, duty
+   to mitigate, sole-and-exclusive remedy framing. Flag: one-way
+   indemnity where market is mutual; missing IP-infringement carve-
+   out from the LoL cap.
+5. **Limitation of liability (LoL)** — cap (fixed dollar / multiple
+   of fees / uncapped), excluded categories (indirect, consequential,
+   lost profits), and carve-outs that punch through the cap (gross
+   negligence, willful misconduct, IP infringement, indemnity
+   obligations, confidentiality breach, breach of data-security
+   obligations). Flag: caps below 1× fees, no carve-outs, "sole and
+   exclusive remedy" framing without an indemnity escape hatch.
+6. **Term / termination** — initial term, auto-renewal, termination
+   for convenience (which side, notice period, fees due on
+   termination), termination for cause (cure period, material-breach
+   definition), effect-of-termination (data return, license survival,
+   payment acceleration). Flag: auto-renewal without notice, no
+   termination-for-convenience for the buyer, asymmetric cure periods.
+7. **Dispute resolution** — governing law, venue, arbitration
+   (provider, seat, rules, language), class-action waiver, jury
+   waiver, injunctive-relief carve-out, fee-shifting. Flag: venue
+   inconvenient for the user, no injunctive-relief carve-out for IP
+   / confidentiality.
+8. **Boilerplate** — assignment (consent vs no consent on change of
+   control), notices, force majeure, severability, entire agreement,
+   amendment, counterparts, no-waiver. Flag: assignment without
+   change-of-control consent, force majeure that excuses payment.
+
+## Step 2 — Mark asymmetry explicitly
+
+For every clause that runs only one way, propose mutualization with
+rationale. Asymmetry is the single highest-yield finding in a first
+pass: indemnity, reps & warranties, audit rights, termination rights,
+notice cures.
+
+## Step 3 — Propose redlines with rationale + market position
+
+Each proposed redline gets:
+- **Original language** (quoted)
+- **Proposed redline** (mark-up form)
+- **Rationale** — what risk this addresses
+- **Market position** — what's standard for deals of this type / size
+- **Fallback** — acceptable counter-position if the counterparty
+  resists
+
+Without rationale + market position, redlines look like preference
+fights and lose negotiating leverage.
+
+## Step 4 — Escalation triggers (emit `:::escalation`)
+
+Any of these trigger a hard handoff to outside counsel:
+
+- The contract touches a regulated industry (healthcare, financial
+  services, defense, insurance, broker-dealer, regulated utilities,
+  cannabis, gambling)
+- The user is signing today (binding-execution moment)
+- A dispute, demand letter, or arbitration is already noticed
+- Choice-of-law / venue analysis requires state-specific or non-US
+  expertise (state UCC variations, foreign-judgment enforcement,
+  state non-compete law, CCPA / Cal AB-1184 or other state
+  consumer-protection statutes in the operative jurisdiction)
+- Criminal-law touchpoints (export controls, sanctions, anti-bribery
+  posture, white-collar exposure)
+- Employment-classification questions (independent-contractor vs
+  employee, ABC test by state)
+
+## Output
+
+Wrap the review as a single `:::artifact` block tagged
+`DRAFT — NOT LEGAL ADVICE`, organized by clause category, with the
+matching `:::escalation` block whenever a trigger fires.