Agent Capability Substitution Resolver

Can a proposed capability B replace incumbent A for this specific declared task contract?

This tool evaluates a candidate already selected by your router or integration. It does not discover tools, recommend providers, optimize price or execute either service.

What it checks

• Input compatibility: the candidate must accept every input permitted by the task's supported schema.
• Output guarantees: the candidate's declared output schema must be contained in the task's accepted output schema.
• Goal and output semantic identifiers must match. Identical JSON types alone are insufficient; a positive-sentiment score and a spam score are different meanings.
• Candidate side effects must be explicitly declared and allowed by the task.
• Simple output-field invariants must be established by declared schemas. Passing a few examples does not prove a universal invariant.
• Supplied paired fixtures validate input, both providers' outputs, relevant output equality, invariants and observed effects.
• Optional explicit, lossless, top-level output renaming is assessed as an adapter plan.

All declarations and fixture execution results are supplied by you. The resolver cannot verify that a provider tells the truth about its guarantees or that a fixture actually ran.

Decisions

• DROP_IN: declared supported contracts align, semantic identifiers match, effects are allowed and the minimum supplied paired fixtures pass. safe_to_substitute is true only within this declared scope.
• ADAPTER_REQUIRED: an explicit lossless output rename makes the contract compatible and fixtures pass. The adapter is not executed or installed; integration testing remains required. safe_to_substitute stays false.
• SHADOW_TEST_REQUIRED: additional paired fixtures or a candidate runtime invariant guard is needed. Merely adding more examples cannot prove a universal invariant.
• INSUFFICIENT_EVIDENCE: semantic identifiers, effects, baseline evidence or supported schema guarantees are missing. Unsupported schema keywords are reported rather than ignored.
• NOT_SUBSTITUTABLE: the candidate fails the declared input/output guarantees, semantics, allowed effects or valid paired fixture checks.

When the supplied incumbent baseline is invalid, that fixture cannot establish candidate equivalence.

Input

On Apify paste the full contract in the request JSON field. HTTP takes the contract directly.

task_contract contains:

• goal_id: caller-defined versioned goal identifier.
• input_schema, output_schema: supported closed object schemas.
• output_semantics: field to caller-defined meaning/unit identifier.
• allowed_side_effects: identifiers of permitted effects.
• invariants: {id, field, constraint}; each constraint uses the supported schema subset.
• comparison_fields: exact-value comparisons on paired outputs; must include every required task output.
• minimum_fixtures: 1 to 100.

incumbent_capability and candidate_capability contain:

• id, goal_id, input_schema, output_schema, output_semantics, side_effects.
• side_effects: [] explicitly declares none; null means unknown and cannot pass.

fixtures contain unique id, input, incumbent_output, candidate_output, incumbent_side_effects, candidate_side_effects. They are evidence records supplied by the caller, not code to execute.

Optional output_mapping maps target task output field to candidate source field. Every candidate output field must be represented exactly once. No coercion, dropped fields, unit conversion or arbitrary transformation is performed.

Supported schema subset — v0.1

Single type: object, array, string, number, integer, boolean or null.

• Objects: explicit properties, required list, additionalProperties: false.
• Arrays: items, minItems, maxItems.
• Strings: minLength, maxLength (Unicode codepoints).
• Numbers/integers: minimum, maximum.
• Nonempty enum (up to 100 values).
• title/description are annotations only.

No $ref, unions, anyOf/oneOf/allOf, patterns, formats, conditionals, exclusive bounds, arbitrary extensions or open objects. Unsupported features yield INSUFFICIENT_EVIDENCE. Conservative inclusion may reject declarations whose compatibility would require reasoning outside this subset.

Root input/output schemas must be objects. At least one required task output is needed. Exact comparison is object-key order independent and array-order sensitive; missing and null differ. Every task output property, including optional ones, needs matching semantic identifiers.

Boundaries

• No LLM, network call, provider invocation, fixture execution or production traffic interception.
• No authentication of semantics, side effects, metadata or evidence.
• No general behavioral equivalence proof or production safety certification.
• Finite fixtures cannot establish correctness for all future inputs.
• An adapter result is a reviewable rename specification, not executable deployment.
• Up to 100 fixtures, 100 schema properties per object, 100 invariants, schema depth 8, JSON depth 32 and 64 KiB input.

The synthetic prefilled sentiment fixture uses supplied numbers. It does not call or contain a sentiment model.

Output

decision, safe_to_substitute, assessment_scope, contract_coverage, compatible_fields, behavioral_gaps, side_effect_differences, unknowns, unsupported_schema_features, fixture_results, adapter_spec, required_tests and limitations.

Pricing and contact

One result report per evaluation, including negative or insufficient-evidence decisions. Invalid contracts fail before producing a result. Consult the live Pricing tab for current result and startup charges.

bonometti.work.ai@gmail.com