4. The framework

4. The framework

ANAD treats a correct agent answer as an effect of the whole network that produces it, and works on that network in two layers across four phases. This section presents each part and the reasons for it. Appendices B to H give every table and template.

4.1 Two layers, four phases

ANT supplies the network layer: who and what acts, where meaning changes, who holds authority, and what will disturb the network. NFE supplies the fidelity layer: what a value must carry to the decision, how it is lost, how each loss is controlled, and how fidelity is measured.

An analysis can end without a design. A finding that needs a person's decision is recorded and handed to its owner; it does not become a tool.

Eight principles govern the framework:

1. Analysis and design are different jobs. ANT is an analysis tool first; the design half is our adaptation.
2. Findings need not be tool-shaped. Unagreed rules, wrong source data and approvals that must stay with a role have their own record.
3. The unit is the network serving one decision. A value is judged at the decision, not at the component.
4. Assume mediators. Every hop changes meaning until shown otherwise.
5. Test outcomes, not shapes. The right answer is computed from the rules, never judged by a model.
6. Older systems are the normal case. Fields named for people, wrapped APIs, one word used in many systems, overnight snapshots and authority held in procedures are expected inputs.
7. Effort follows consequence. Low-stakes work needs three moves: the real question, one figure followed through the chain, and one test with a computed answer.
8. Claims are shown, not proven. What the framework does is shown in cases, not measured against other methods.

ANT's working rules (follow the actors, trace the translations, reopen what looks settled) become four questions for an agentic data flow. Who and what can change this outcome? Where does meaning change? Who holds authority? What will pull this apart later?

4.2 What the framework names

The ontology defines what is in the network and what the analysis finds (Appendix B). It has 22 types in five groups: network (decision, actor, role, authority, rule, interest, hop, path, value, qualifier, channel), fidelity (persuasive field, loss, finding), design (control, tool), test (oracle, fixture, assertion, run) and upkeep (disturbance, signal). Every document field and every chart mark draws from it. Six choices shape it:

• Rules are their own type. A rule is carried by an inscription, such as the wording of a policy, and the inscription is the actor. Tests and tools must point at a rule by id, and drift is usually a rule change.
• Authority is a grant held by a role, with a limit. People fill roles. An agent's remit is the same type, granted to the agent, and may be none when no automated actor acts on the value.
• A hidden mediator is a hop whose apparent role is intermediary and whose actual role is mediator. That turns "hidden" into something an analyst can check and record.
• Finding is the main output of analysis. A loss is one kind of finding. Each finding ends in a control, a referral to its owner, or an acceptance.
• A path is a graph, not a line. Side sources join the main value's path where they are combined.
• A tool is a design object, not an actor. Its name, descriptions and schema fields are the inscription actors, because they are what the model reads.

Meaning travels as qualifiers attached to a value. There are nine kinds: unit, basis, scope, entity, status, time, certainty, provenance and authority. A loss is a change, on a hop or inside an actor, after which a value no longer carries a required qualifier correctly; a required qualifier that never reaches the path is also a loss. Each loss has a dimension, which says what kind of fidelity breaks, and a mode, which says how. The six dimensions are source, semantic, transformation, contextual, temporal and authority; Appendix B defines each, with an example.

The first eight modes come from NFE. The ninth, repeat, was found by applying the framework to Harbourline (Section 5.1). Naming the mode tells the designer which control to apply: eleven standard control patterns, such as "carry", "compute at the passage point" and "gate authority", each close particular modes. A persuasive field is a value whose name or value suggests it answers the decision, but does not, such as a claims reserve read as money available to pay.

One glossary of 59 terms serves the whole framework, each marked as ANT's own, adapted from ANT, narrower or simpler than ANT's use, from NFE, or new (Appendix C). Templates use plain words, so practitioners can apply the framework without learning ANT; researchers can trace every adaptation through the glossary.

4.3 Eleven documents

The framework produces eleven documents. Each is a view of the ontology, not a separate source of truth (Appendix D).

Documents 1 to 6 are the analysis pack, the whole product when the framework stops at the analysis gate. At tier 1, three documents are enough: the decision record, the mediation map and one acceptance test.

Three documents carry the framework's main ideas. The decision record rewrites the question as asked into one that names every rule the answer depends on, sets the consequence tier and the tolerance, and records the remit and who may decide. The tier runs from 0 (exploration) through 1 (internal information) and 2 (an external commitment) to 3 (a financial, legal or regulated action). The tolerance is how much loss the decision can accept. The remit is what an agent may do: inform, recommend, act, or none. The findings register holds everything the analysis finds, including what no tool can fix; referrals and accepted losses are views of it. The acceptance test fixes every input and the as-of time (together, the fixture), computes the right answer from the rules (the oracle), and lists the assertions every run must meet. It is the framework's engineered trial of strength. The baseline record and score card mark each run as recorded or illustrative, so the framework's own records keep the caveat this paper applies to its cases.

4.4 Chart notation

The notation adapts data-flow diagrams, which practitioners already read (Appendix E). It keeps their four shapes (process, data store, external entity and flow) and adds one, a document with a folded corner, for inscriptions. A channel is the route by which text reaches its reader; for a model, the MCP client, the program that connects the model to servers, decides which channels arrive. Nine marks show what ANT and NFE need: the obligatory passage point (a double border), a hidden mediator (an accent border), operations, qualifiers on a value, a loss tag, authority lines, rival meanings (a dotted arrow), a channel the client does not forward, and an authority that was not consulted. Mediation is marked on the actor that changes the meaning; values travel on arrows. Shape carries meaning, and colour only repeats it.

The framework draws two charts. The mediation map is always drawn: one key value, followed from its authoritative source to the decision, top to bottom, with side sources joining from the right. The network chart is drawn only when the actor list is long or crosses organisations. Each chart is a freeze frame of one dated network version, an idea taken from Alexander and Silvis (2014).

Figure 1 follows Harbourline's invoice from the contractor's PDF to the handler's decision. The figure marks five of the eight losses: L6 at invoice capture, L2, L3 and L5 at the server's lookups, and L8 between the model and the handler. L1, L4 and L7 are in the loss table in Section 5.1 but are not drawn. The only actor that looks harmless is the one marked: accounts payable copies a total and, in doing so, drops what the total meant.

4.5 Lifecycle, gates and cycles

The lifecycle merges two earlier procedures into four phases and thirteen steps, and drops none of their steps (Appendix F). One is a ten-step procedure from the authors' earlier working paper, which turned ANT's concepts into stages and was first applied to Harbourline. The other is NFE's eight steps. Each step answers one question and leaves at least one record.

Four gates close the phases, each signed by a named role:

• G1, analysis, signed by the decision owner. Every hop is classified; every loss has a mode, dimension and effect; every finding has an owner and a proposed disposition, and referrals are sent. G1 is also the analysis-only exit. Live work needs confirmed owners; a case study may pass on proposed owners, and is recorded as such.
• G2, design, signed by the fidelity owner. Every loss has a control or a recorded acceptance. There is no write tool unless the remit is "act". Every write tool, and at tier 3 every tool not shown by test to be read-only, has been through safety analysis.
• G3, release, signed by the fidelity owner, and also by the decision owner at tier 3. The score card is within tolerance, every closed loss is guarded by an assertion, every run on the score card is recorded rather than illustrative, every tool has a capability record, and the disturbance and monitoring plan names its signals and owners.
• G4, re-certification, signed by the fidelity owner. The affected fixtures, including one dated under the changed rule, have been re-run, and the score card is no worse or the change is explained.

After the first pass, four cycles send work back. Fix returns a failed assertion for a known loss to design. Discovery returns a failure that matches no known loss to analysis, so a team does not patch a symptom when the analysis was wrong. Referral brings a decided finding back into the fidelity specification. Upkeep sends each reopening event, such as a rule change, a schema change, a new tool in the client or a model upgrade, to the step named for it. Fix is common practice. Discovery, referral and upkeep have counterparts in change management and re-certification; the framework ties each to named records and to the step where work re-enters.

The consequence tier, set at A1, decides how much of the lifecycle to run: a line of record at tier 0, the three moves at tier 1, all thirteen steps at tier 2, and at tier 3 all steps with an independent review of the oracle, safety analysis and re-certification on every change.

4.6 Product, roles and integration

The framework ends in one of two products (Appendix G). An analysis pack at G1 is the right ending when the main findings are an unagreed rule, wrong source data or an approval that must stay with a person. A certified network at G3 is the other. Done means G3, stated in one line: a network whose answers to one decision are tested against a computed right answer, with the records that say why. The unit is one decision, not one server; a server serving three decisions has three sets of records.

Six roles own the records (Appendix H). The decision owner owns the question, its rules and the meaning of its terms. The fidelity owner, the role NFE adds, owns one decision's specification, findings, tests and score card; without one, each team protects its own component and nobody protects the path. Source owners, the server owner, the client and platform owner, and approvers own the rest.

The framework does the discovery and keeps the records; other approaches design and build the tool surface. It joins a wider sequence at named steps: generating tools directly from existing application programming interfaces (APIs) supplies the naive baseline at A4; domain-driven design and data contracts fix one meaning per term before D2; agent-centred tool design (writing names, descriptions and responses for a model reader) and security-first design (trust boundaries and least privilege) shape the tool surface at D2; safety analysis such as STPA examines write tools at G2; evaluation-driven iteration runs the acceptance tests at T1 and T2. Each hand-off names what passes in each direction, so referred findings reach an owner and upgrades re-enter the lifecycle.

The framework does not build the model or the client, choose security controls, perform safety analysis of actions, correct source data or rules, or prove it beats other methods. It checks that the right meaning reaches the decision; it cannot make the decider act on it.