Agentic discoverability framework

Framework status

The Agentic discoverability framework is a versioned proposal. Version 0.1 formalizes a vocabulary, a transition chain, axioms and an evidence regime.

It is not a standard, certification or adopted industry protocol. It does not promise that an object will be cited, recommended or selected by a third-party system.

The proposed canonical definition owns the term. This framework organizes its application and observation.

Problem addressed

Visibility measurements often compress several states: a resource exists, it can be retrieved, it appears in an answer, it is cited, compared, recommended or used.

These states are not equivalent.

A system may:

• find a resource without resolving the entity correctly;
• resolve the entity without understanding its offer;
• understand the offer without considering it for the current intent;
• consider it without selecting it;
• select it while representing it incorrectly;
• recommend it without being able to invoke it or act with it.

The framework prevents these states from being silently merged.

Unit of analysis

Every observation must declare at least:

D = object × intent × agent × context × time

Where:

• object means an entity, content item, service, offer, resource or digital capability;
• intent means the problem, need, comparison, search, decision or action;
• agent means the system, model, version, configuration and available tools;
• context means language, region, constraints, available data, user profile and session state;
• time means an observation instant or window.

A conclusion that omits one of these dimensions must be qualified as partial.

Transition chain

The public model can be summarized as:

Presence → Exposure → Access → Resolution → Interpretation → Admissibility → Consideration → Selection → Restitution

These nine gates do not claim to reproduce the internal architecture of every system. They are a verification grid for transitions an observer may attempt to establish.

1. Presence

Question: does the object exist in an addressable and sufficiently stable environment?

Possible evidence: route, identifier, page, catalogue entry, resource, version and publication date.

Typical failures: absent object, unpublished surface, missing version or unrepresented entity.

2. Exposure

Question: can a system learn that the object exists?

Possible evidence: internal links, sitemap, catalogue, index, structured data, discovery surface or qualified external relation.

Typical failures: orphan surface, poor routing, incomplete catalogue or contradictory signal.

3. Accessibility

Question: can the system reach and read the resource under applicable conditions?

Possible evidence: HTTP response, usable rendering, access policy, authentication, format negotiation and route stability.

Typical failures: blocking, empty rendering, critical dependency on a late state, ambiguous permission or inaccessible resource.

4. Resolution

Question: is the object correctly identified and distinguished from adjacent objects?

Possible evidence: identifiers, canonical name, relations, version, owning organization and declared scope.

Typical failures: homonymy, brand merge, wrong version or attribution to the wrong entity.

5. Interpretation

Question: are role, meaning, conditions and boundaries reconstructed faithfully?

Possible evidence: canon-output comparison, claim matrix, scope tests and omission detection.

Typical failures: expanded role, omitted condition, invented capability or erased conceptual distinction.

6. Admissibility

Question: can the information or capability legitimately support the current intent given its authority, provenance, state and limits?

Possible evidence: canonical source, source hierarchy, dated proof, applicability condition and declared authority.

Typical failures: weak source, stale state, insufficient evidence, inapplicable scope or missing permission.

7. Consideration

Question: does the object enter the field of relevant options for the declared intent?

Possible evidence: instrumented-agent trace, explicit candidate list, visible comparison or routing explanation.

Typical failures: accessible object never considered, implicit filter, incompatible context or ignored candidate.

In a black-box system, this gate is often NOT_EVALUATED.

8. Selection

Question: is the object chosen, recommended, cited as a primary source or retained for an action?

Possible evidence: final output, explicit recommendation, source choice, invocation or traced decision.

Typical failures: competitor selected, object demoted, unsupported choice or out-of-context recommendation.

9. Restitution

Question: is the selected object represented faithfully with the required limits, attribution and conditions?

Possible evidence: canon-output comparison, attribution, claim accuracy and preservation of exclusions.

Typical failures: incorrect description, misleading citation, removed conditions or expanded scope.

Axioms

1. Relationality axiom

Discoverability is never absolute. It depends on the object, intent, agent, context and time.

2. Non-equivalence axiom

Presence, retrieval, understanding, citation, consideration, recommendation, selection and action are not interchangeable.

3. Eligibility axiom

A missing recommendation is not a failure when the object was not appropriate for the intent. Recommendation and selection rates must use scenarios in which the object was genuinely eligible.

4. Fidelity axiom

An appearance or selection based on an incorrect representation is not a complete success.

5. Observability axiom

An unobservable internal step must not be asserted. The correct status is NOT_EVALUATED, not an invented explanation.

6. Temporality axiom

Every observation is tied to a system version, region, language, configuration and time.

7. Permission separation axiom

Discovery does not grant permission to train, extract, invoke, transact or act.

Evidence statuses

Each gate receives an explicit status:

SUPPORTED_INFERENCE must never be presented as OBSERVED.

Three observation modes

Mode 1: surface audit

This mode verifies what the publisher controls directly:

• routes and canonicals;
• HTML and structured data;
• discovery surfaces;
• entity graphs;
• access policies and conditions;
• service or capability catalogues;
• provenance, versions and source hierarchy.

It measures controllable preparation. It does not prove consideration or selection by a third party.

Mode 2: instrumented agent

A controlled agent runs with traces that expose:

• consulted resources;
• identified candidates;
• criteria;
• rejections;
• selected source or capability;
• possible invocation;
• produced restitution.

This is the preferred mode for directly evaluating consideration and selection.

Mode 3: external black-box panel

Versioned scenarios are submitted to several external systems. Controlled variation may include language, region, intent, constraints, wording and time.

Only visible outputs are declared observed. Internal mechanisms remain NOT_EVALUATED.

Three output profiles

Controllable readiness profile

This profile describes what the publisher can correct:

• resource exposure;
• entity resolution;
• clarity of offers and capabilities;
• provenance and authority;
• boundaries, permissions and conditions;
• route and interface stability.

Observed discoverability profile

This profile describes what systems actually do in tested scenarios:

• appearance;
• citation;
• inclusion in comparison;
• recommendation;
• selection;
• variation across agents, languages, regions and time.

Interpretive integrity profile

This profile describes restitution quality:

• claim fidelity;
• attribution;
• scope compliance;
• preservation of limits;
• conformity with the canonical source;
• absence of unjustified expansion.

These profiles must not be silently merged.

Initial metrics

Version 0.1 forbids a universal score out of 100. It allows separate metrics with declared denominators and limits.

Every metric must publish its corpus, sample, observation mode and NOT_EVALUATED states.

Minimal observation contract

A record should retain at least:

• object identifier;
• object type;
• intent;
• system, model and version;
• language and region;
• context and constraints;
• observation time;
• evaluated gate;
• evidence status;
• result;
• provenance;
• raw artifact or reference;
• observed-content digest;
• limits and unevaluated factors.

The following separation must be preserved:

raw → structured → interpreted → derived

Boundary with actionability

Discoverability ends at selection and restitution.

Actionability begins when an agent must invoke a capability, transmit data, book, purchase, submit, modify or execute.

An object may be discoverable without being actionable. A capability may be actionable without being appropriate for the intent. Authorization and action governance remain separate objects.

Non-promises

The framework does not prove:

• that a third-party system will consult a surface;
• that a brand will be cited or recommended;
• that an internal rank exists or can be reconstructed;
• that an observed result will remain stable;
• that a technical protocol causes selection;
• that a recommendation is commercially favorable;
• that an invocation is authorized or will succeed.

Authorship and citation

This publication does not claim invention of the generic expression “agentic discoverability.” It establishes the dated publication of this particular formalization by Gautier Dorval: relational unit, nine gates, seven axioms, evidence statuses, output profiles and separation from actionability.

Suggested citation:

Dorval, Gautier. “Agentic discoverability framework,” proposed version 0.1, August 27, 2026, https://gautierdorval.com/en/frameworks/agentic-discoverability-framework/.

History