---

name: write-contract description: Write production-quality GenLayer intelligent contracts. Always pins concrete GenVM runner version hashes and never uses local-only test/latest runner aliases. Covers equivalence principles, storage rules, LLM resilience, and cross-contract interaction. allowed-tools:

• Bash
• Read
• Write
• Edit
• Grep
• Glob

---

Write Intelligent Contract

Guidance for writing GenLayer intelligent contracts that pass consensus, handle errors correctly, and survive production.

Critical: Pin the Runner Version

All GenLayer networks reject py-genlayer:test, py-genlayer:latest, and unversioned runner aliases. Every generated contract MUST start with a pinned runner dependency header.

# { "Depends": "py-genlayer:1jb45aa8ynh2a9c9xn3b7qqh8sm5q93hwfp7jqmwsfhh8jpz09h6" }

test and latest are local-development aliases for GenLayer runtime developers. They may work only in a specially configured local Studio environment with a GenLayer developer environment variable, but they do not work on GenLayer networks and must not appear in generated user contracts.

Before returning any contract code, verify:

• The first line is a pinned Depends runner version hash.
• There is no py-genlayer:test.
• There is no py-genlayer:latest.
• There is no unversioned py-genlayer.

Always lint with genvm-lint check after writing or modifying a contract.

When to Use GenLayer

Before writing code, decide whether the feature actually needs GenLayer consensus. Recent builder feedback shows many projects start by treating GenLayer as a generic AI backend; push them toward a clear on-chain consensus role.

Use GenLayer when the contract must coordinate or settle around a subjective, external, or AI-mediated judgment that multiple validators should verify independently:

• Dispute resolution where evidence must be evaluated and the result affects escrow, payouts, reputation, or access.
• Prediction/oracle-style markets where the contract needs an independently validated outcome from external evidence.
• Compliance, moderation, or scoring workflows where the final decision must be reproducible enough for validator agreement but cannot be reduced to a simple deterministic API call.
• Autonomous agents that need transparent settlement, appeals, and auditable state transitions rather than a private off-chain decision.

Prefer a normal backend, frontend, or off-chain LLM workflow when:

• The frontend already computes the final answer and GenLayer would only rubber-stamp it.
• The contract only stores user-provided data with no validator-verifiable judgment.
• A deterministic smart contract, REST API, or database job can perform the work without AI consensus.
• The data-fetching/prompting step is not tied to an on-chain state transition, escrow, payout, or appealable decision.

For every contract, write down the boundary before implementation:

• Frontend/backend owns: UI, user auth, indexing, non-authoritative previews, cached market data, and convenience analytics.
• GenLayer contract owns: the minimum state transition that needs consensus, the evidence inputs, the validator comparison rule, the final settlement effect, and any appeal/rotation path.
• External sources own: raw facts or documents; do not treat them as trusted unless validators can re-fetch, normalize, and compare them.

If the boundary is unclear, create a one-page architecture note before coding: user action -> evidence source -> nondeterministic call -> equivalence principle -> state update -> user-visible settlement.

Contract Skeleton

# { "Depends": "py-genlayer:1jb45aa8ynh2a9c9xn3b7qqh8sm5q93hwfp7jqmwsfhh8jpz09h6" }

from genlayer import *

class MyContract(gl.Contract):
    # Storage fields — typed, persisted on-chain
    owner: Address
    items: TreeMap[str, Item]
    item_order: DynArray[str]

    def __init__(self, param: str):
        self.owner = gl.message.sender_account

    @gl.public.view
    def get_item(self, item_id: str) -> dict:
        return {"id": item_id, "value": self.items[item_id].value}

    @gl.public.write
    def set_item(self, item_id: str, value: str) -> None:
        if gl.message.sender_account != self.owner:
            raise gl.UserError("Only owner")
        self.items[item_id] = Item(value=value)
        self.item_order.append(item_id)

Runner Dependencies

The first line of a contract declares the GenVM Python runner. Always pin a specific runner version hash. All GenLayer networks reject test, latest, and unversioned runner aliases in generated contracts.

Single-file Python contracts

# { "Depends": "py-genlayer:1jb45aa8ynh2a9c9xn3b7qqh8sm5q93hwfp7jqmwsfhh8jpz09h6" }

Multi-file Python contract packages

Use py-genlayer-multi when the contract is packaged across multiple files.

# { "Depends": "py-genlayer-multi:06zyvrlivjga0d5jlpdbprksc0pa6jmllxvp8s20hq1l512vh5yk" }

Contracts using embeddings or semantic search

Add py-lib-genlayer-embeddings before the main Python runner with a Seq block.

# {
#   "Seq": [
#     { "Depends": "py-lib-genlayer-embeddings:0bmbm3cyfwxsyh454z53vxqjf47wz2q7smcqp1q4g4a6k2kidnyk" },
#     { "Depends": "py-genlayer:1jb45aa8ynh2a9c9xn3b7qqh8sm5q93hwfp7jqmwsfhh8jpz09h6" }
#   ]
# }

Equivalence Principle — Which One to Use

This is the most critical decision. Pick wrong and consensus will fail or be trivially exploitable.

Decision Tree

Can validators reproduce the exact same normalized output?
├── YES → strict_eq
│         Exact match. Use when outputs are deterministic or can be
│         canonicalized (e.g., JSON with sort_keys=True).
│         Examples: blockchain RPC, stable REST APIs.
│
└── NO  → Write a custom validator function (run_nondet_unsafe)
          Default: produce independent evidence. Usually rerun the same task
          and compare decision fields, derived status, scores, or other stable
          outputs with explicit tolerances. Only skip the second answer when the
          validator can judge the leader output against source data and criteria.

GenLayer also provides prompt_comparative and prompt_non_comparative as convenience wrappers, but most contracts outgrow them quickly. Start with a custom validator function for full flexibility.

Independent verification by default

For LLM and web operations, never trust the leader. The validator must verify the substance of the leader's answer using evidence other than the leader's answer alone. In practice that means one of:

• Rerun the same LLM/web task and compare the stable decision fields.
• Fetch the same source data and independently derive the status being stored.
• Run an explicit comparative LLM judgment over the leader output and validator output.
• For open-ended outputs, judge the leader output against the same input/source data and explicit criteria.

Do not write validators that only check leader_result.calldata for a valid JSON shape, allowed enum value, non-empty summary, or confidence in range. That is leader-output-only validation, not consensus. It trusts the leader's substantive answer 100% and only proves that the leader formatted the answer correctly.

Non-comparative validation does not mean "trust the leader." It means the validator does not produce a second candidate answer. It still must read the same input/source data and ask whether the leader output is valid under clear criteria. A summary validator, for example, should check whether the proposed summary is faithful to the article, covers the material points, avoids hallucinated facts, and satisfies length/style constraints.

Classification, scoring, extraction, authenticity decisions, safety decisions, ranking, and settlement logic almost always need comparative validation: rerun or independently derive the answer, then compare the decision field, extracted fields, score bucket, or derived status. If the validator only checks that the leader chose an allowed label such as authentic, suspicious, or inconclusive, the leader is deciding alone.

strict_eq — Deterministic calls only

def fetch_balance(self) -> int:
    def call_rpc():
        res = gl.nondet.web.post(rpc_url, body=payload, headers=headers)
        return json.loads(res.body.decode("utf-8"))["result"]
    return gl.eq_principle.strict_eq(call_rpc)

Never use for LLM calls or web pages that change between requests.

Custom Validator Function (most common)

The default choice for non-deterministic operations. You write the leader function and a validator function with your own comparison logic. The validator should independently perform or verify the same substantive task, then compare the result fields that matter.

def score_content(self, content: str) -> dict:
    def leader_fn():
        analysis = gl.nondet.exec_prompt(prompt, response_format="json")
        score = _parse_llm_score(analysis)
        return {"score": score, "analysis": str(analysis.get("analysis", ""))}

    def validator_fn(leaders_res: gl.vm.Result) -> bool:
        if not isinstance(leaders_res, gl.vm.Return):
            return _handle_leader_error(leaders_res, leader_fn)

        validator_result = leader_fn()
        leader_score = leaders_res.calldata["score"]
        validator_score = validator_result["score"]

        # Gate check: if either is zero (reject), both must agree
        if (leader_score == 0) != (validator_score == 0):
            return False

        # Tolerance: within 5x/0.5x bounds
        if leader_score > 0 and validator_score > 0:
            ratio = leader_score / validator_score
            if ratio > 5.0 or ratio < 0.2:
                return False

        return True

    return gl.vm.run_nondet_unsafe(leader_fn, validator_fn)

Convenience Wrappers

prompt_comparative reruns the task and sends both outputs to an LLM with your principle string. prompt_non_comparative does not rerun the task; it asks validators to judge the leader output against input data and criteria. Both are convenient for prototyping but limited - for most production contracts, prefer a custom validator function with explicit comparison logic.

Prefer prompt_comparative unless you can explain why independently doing the task again would be meaningless and how the validator will still verify the leader output against source data. If the only reason is "outputs may differ," compare the decision fields, normalize the output, derive a status, or use tolerance instead of dropping comparison entirely.

def resolve(self) -> str:
    def analyze():
        page = gl.get_webpage(url, mode="text")
        return gl.exec_prompt(f"Analyze: {page}\nReturn JSON with outcome and reasoning.")

    return gl.eq_principle.prompt_comparative(
        analyze,
        principle="`outcome` field must be exactly the same. All other fields must be similar.",
    )

Error Classification

Classify errors so validators know how to compare them. This is critical for consensus on failure paths.

ERROR_EXPECTED  = "[EXPECTED]"   # Business logic (deterministic) — exact match required
ERROR_EXTERNAL  = "[EXTERNAL]"   # External API 4xx (deterministic) — exact match required
ERROR_TRANSIENT = "[TRANSIENT]"  # Network/5xx (non-deterministic) — agree if both transient
ERROR_LLM       = "[LLM_ERROR]"  # LLM misbehavior — always disagree, force rotation

Canonical error handler for validators

def _handle_leader_error(leaders_res, leader_fn) -> bool:
    leader_msg = leaders_res.message if hasattr(leaders_res, 'message') else ''
    try:
        leader_fn()
        return False  # Leader errored, validator succeeded — disagree
    except gl.vm.UserError as e:
        validator_msg = e.message if hasattr(e, 'message') else str(e)
        # Deterministic errors: must match exactly
        if validator_msg.startswith(ERROR_EXPECTED) or