Prompt Serving

Cognition Prompt Service is the authority for how a prompt is composed for a model. It draws from kernels (core truth) and advisory artifacts (memory, knowledge, skill) and produces typed prompt context, with strict separation between the lanes.

Three Format Methods

Method	Serves
FormatCore	Kernel truth only — agent kernel + world kernel
FormatAdvisory	Validated advisory artifacts — memory, knowledge, skill
FormatAll	Composed core + advisory under admitted ordering

The three lanes are intentionally separated. FormatCore cannot include advisory content; FormatAdvisory cannot include kernel content; FormatAll is the admitted composition.

Why Lane Separation

Without lane separation, prompt assembly would silently mix identity context with recalled context. The model would not be able to tell "this is who I am" from "this is what I remember" from "this is what I'm currently working on."

With lane separation:

• FormatCore is identity context — slow-changing kernel truth.
• FormatAdvisory is recalled context — the agent's memory / knowledge / skill bundle relevant to this turn.
• Working state and routine evidence are excluded from prompt serving by design.

The model gets a prompt with explicitly-typed sections; nothing sneaks in.

What Working State Does NOT Reach Prompt

Working state is transient cognition scaffolding — the agent's intermediate notes, the routine's bookkeeping, etc. It isn't durable truth and is never served as prompt context.

If working state could leak into prompts, prompts would acquire arbitrary scaffolding. The hard boundary keeps prompts clean.

Prompt Lanes

Lanes are admitted under tables/prompt-serving-lanes.yaml. Each lane declares:

Field	Purpose
Serving order	Where the lane appears in FormatAll
Admitted families	Which artifact families this lane reads
Admitted inputs	What inputs the lane accepts
Derived view source	Which service-derived view backs this lane
Forbidden inputs	What is explicitly excluded

A reader who wants to know "what shape does this lane produce" goes to the table, not to docs prose.

Reader Scenario: An Agent's Turn Composes A Full Prompt

An agent is about to produce a turn.

1. FormatAll invoked. Cognition's prompt service composes the full prompt.
2. Core lane. Agent kernel + world kernel — slow-changing identity.
3. Advisory lanes. Memory recall, knowledge query, skill bundle — all validated advisory.
4. Admitted serving order. Lanes appear in admitted order.
5. Working state excluded. Transient scaffolding does not leak.
6. Result handed to Brain. Model sees typed prompt with typed sections.

The model cannot tell "this is identity" from "this is recall" unless the prompt distinguishes — which the lane separation ensures.

Reader Scenario: A Mod Tries To Inject Through Prompt

A mod that has access to the agent's surface attempts to inject extra context by smuggling it through prompt assembly.

1. Mod constructs intended content. Wants it in the prompt.
2. Submit to admitted lane. The lane validates the input against its admitted shape.
3. Forbidden input rejected. Working state, routine evidence, or other forbidden inputs fail validation.
4. Allowed input admitted under lane. Stays under that lane's typed shape.

A mod cannot smuggle working state through prompt assembly. The lane validation is the gate.

Reader Scenario: An Auditor Asks "What Did The Model See"

An auditor wants to reconstruct exactly what prompt context the agent had during a specific turn.

1. Locate turn. By trace id.
2. Read prompt composition. Cognition records what lanes produced what content for this turn.
3. Reconstruct. Auditor sees core lane content, advisory lane content, ordering.
4. No hidden content. Nothing is in the model's input that is not in the recorded composition.

Lane separation makes this reconstruction structural.

Boundary Summary

Concern	Owner
Prompt composition	Cognition Prompt Service
Lane semantics	tables/prompt-serving-lanes.yaml
Kernel truth source	Cognition kernels
Advisory inputs	Memory / knowledge / skill services (admitted)
Working state	Excluded by design

Source Basis

• .nimi/spec/cognition/kernel/prompt-serving-contract.md
• .nimi/spec/cognition/kernel/cognition-contract.md
• .nimi/spec/cognition/kernel/family-contract.md
• .nimi/spec/cognition/kernel/tables/prompt-serving-lanes.yaml
• .nimi/spec/cognition/kernel/memory-service-contract.md
• .nimi/spec/cognition/kernel/knowledge-service-contract.md
• .nimi/spec/cognition/kernel/skill-service-contract.md