Search

Overview

Search is one of the core execution capability families provided by the Forge runtime.

This section contains canonical verification examples for Search capabilities implemented by Forge. Every example is derived from a maintained execution profile and validated against the production execution pipeline.

These documents are not tutorials.

They are executable capability specifications that demonstrate how a Search capability behaves under independent verification.

Search in Forge does not mean document retrieval or semantic lookup.

Search means deterministic exploration of a scenario space through mutation, scoring, ranking, and evidence generation.

---

Primitive Profile

Search capabilities are implemented through the search@1 primitive.

Each documented capability corresponds to a versioned Primitive Profile registered in the Forge execution surface.

Primitive
search

Version
1

Profile
<profile-id>

Representative profiles include:

• finance.system.systemic_fragility_search.v1
• finance.credit.scenario_set.discovery.v1

Each profile defines a specific scenario-space discovery contract.

---

Capability

Search capabilities discover important candidate scenarios inside a bounded execution space.

A Search profile typically receives:

• a base scenario or base state
• mutation rules
• a ranking objective
• candidate limits
• optional evidence from prior executions
• optional scaling or scoring configuration

The runtime then explores the defined mutation space and returns ranked candidates, worst-case candidates, score distributions, mutation traces, and replayable execution metadata where available.

Search is used when the evaluator needs to ask:

Where does this system begin to fail?

rather than:

What is the average outcome?

---

Canonical Contract

Execution is performed through the canonical Forge execution contract.

Primitive : search
Version   : 1
Profile   : <profile-id>

Requests are validated by the canonical execution validator before entering the runtime.

The canonical payload shape is:

op.name
op.version
op.profile
args

Optional execution-level fields may include:

ctx
seed
policy

The exact argument contract is profile-specific.

---

Required Inputs

Search profiles generally require:

• iterations
• base_scenario or base_state
• mutation_rules
• ranking
• ranking.objective
• ranking.top_n

Some profiles use domain-specific required fields inside the base scenario.

Execution cannot begin unless the submitted payload satisfies the profile-specific canonical contract.

---

Optional Inputs

Supported optional inputs vary by profile.

Common optional surfaces include:

• evidence
• scoring weights
• scale configuration
• additional mutation axes
• candidate ownership metadata
• upstream Monte Carlo summaries
• upstream Graph summaries

Optional inputs refine discovery and scoring behaviour without changing the primitive identity.

Agents and evaluators must not invent optional fields.

Only fields declared by the profile contract should be used.

---

Canonical Smoke

A canonical Search smoke should verify:

• profile discovery
• contract inspection
• minimum valid payload construction
• successful validation
• successful execution
• ranked candidate generation
• artifact generation
• replay metadata availability

The maintained smoke suite remains the canonical source for executable smoke payloads.

Documentation should describe verification expectations rather than duplicating every maintained payload.

---

Verification Expectations

A successful Search execution should demonstrate that:

• the primitive resolves to search@1
• the profile resolves to a registered versioned contract
• canonical validation succeeds
• mutation rules are accepted
• ranking objective is supported
• execution completes successfully
• ranked candidates are returned
• runtime evidence is inspectable
• replay metadata is available
• artifacts are generated when requested or supported

Successful validation alone is not sufficient.

A valid Search payload may still produce low-information results if the mutation space is too narrow.

---

Runtime Evidence

Successful Search execution exposes observable runtime evidence including:

• execution identifier
• primitive profile
• execution metadata
• selected ranking objective
• candidate count
• ranked candidate count
• best or worst candidate
• score distribution
• mutation trace
• replay metadata
• artifact references where available

The exact evidence surface depends on the profile and execution policy.

---

Replay

Search profiles support deterministic replay when executed using the same canonical contract, deterministic seed, and compatible runtime version.

Replay should confirm that the ranked candidate set and execution evidence remain consistent under the documented replay guarantees.

Replay verification should be performed after successful execution.

---

Artifacts

Typical Search artifacts may include:

• ranked candidates
• best candidate
• worst-case candidate
• score distribution
• mutation trace
• candidate scoring surface
• execution summary
• replay metadata

Artifacts should be inspected before interpreting the result.

A Search result is not complete if the evaluator only reads the top candidate without inspecting how that candidate was produced.

---

Applied Intelligence Modules

Search capabilities are reused across multiple Forge Intelligence Modules, including:

• Capital Intelligence
• Credit Intelligence
• Banking Intelligence
• Insurance Intelligence
• Reinsurance Intelligence
• Climate Intelligence
• Infrastructure Intelligence
• Systemic Risk Intelligence

The same search@1 primitive can support multiple domains because the primitive verifies discovery behaviour, while the profile defines the domain-specific scenario space.

---

Related Documentation

• /examples/
• /verification/
• /verification/capability-verification
• /verification/smoke-suite
• /verification/replay-determinism
• /verification/artifact-inspection
• /api/kernel/execution-model

---

Verification Checklist

Verification Surface	Status
Primitive resolved	✓
Profile resolved	✓
Contract inspected	✓
Required inputs identified	✓
Optional inputs bounded	✓
Canonical smoke available	✓
Runtime evidence inspected	✓
Artifacts inspected	✓
Replay verified	✓

---

Final Principle

Search capabilities are verified as versioned discovery contracts.

They do not prove value by returning a single candidate.

They prove value when an evaluator can inspect the mutation space, ranking objective, generated candidates, score distribution, artifacts, and replay metadata that produced the discovered result.