Skip to content

Prompt Lanes

Status: Admitted, in build-out

The prompt-lane registry (tables/prompt-serving-lanes.yaml) and the C-COG-050..C-COG-052 prompt-serving contracts are admitted. The lane-routing implementation across families is in active build-out.

What Prompt Lanes Are

A prompt lane is an admitted route through which a particular class of cognition truth reaches the prompt. Lanes are not just formatter output layout; they govern what kind of truth may flow where, in what serving order, and what is forbidden.

The fixed authority surface:

ConcernAuthority
Lane registrytables/prompt-serving-lanes.yaml
Lane semantics + serving rulesC-COG-050
Lane separation + derived metadataC-COG-051
Failure modelC-COG-052

Required Per-Lane Declarations

Every admitted prompt lane must declare:

DeclarationPurpose
Serving orderWhere this lane sits in prompt assembly
Admitted familiesWhich cognition families may flow through this lane
Admitted inputsWhat input shapes the lane accepts
Derived-view sourceWhere service-owned derived views come from
Forbidden inputsWhat may NOT flow through this lane

A lane that does not declare these is not admitted. Apps cannot configure new lanes by convention.

Lane Separation Rule (C-COG-051)

The most important rule on this page:

Kernel truth remains in a dedicated core lane and must never be merged implicitly into advisory context.

A lane carries one kind of truth. Other kinds do not get to ride in.

Lane classCarriesForbidden
Core laneKernel truth (admitted authoritative artifacts)Advisory context
Advisory laneValidated artifacts or service-owned derived viewsWorking state, routine evidence (unless admitted later), caller-persisted metadata

Working state and routine evidence are excluded from prompt serving unless a later cognition rule explicitly admits them. Cleanup, support, or serving signals may appear in prompt output only when they come from explicit derivation logic — not from caller-persisted metadata.

Failure Model (C-COG-052)

Prompt serving must fail close on lane or derivation violations:

ViolationRequired behavior
Missing required kernel artifactsExplicit failure (no best-effort rendering)
Illegal lane mixingExplicit failure
Malformed derived viewsExplicit failure
Forbidden prompt inputsExplicit failure

Formatter convenience must not override family-truth ordering. The prompt output must not silently imply kernel truth from advisory-only inputs.

Why Lane Separation Matters

If kernel truth and advisory context were merged silently, two failure modes follow:

  1. Authority drift. The model can't tell which content carries admitted authority and which is suggestive — it treats them uniformly.
  2. Privacy / safety leak. Working state or caller-persisted metadata that was never admitted into prompts can now influence model output.

The lane system is a contract-level guard against both.

Reader Scenario: A Turn Pulls From Memory + Knowledge + Skill

A turn assembles a prompt that needs:

  • Memory (per-participant continuity)
  • Knowledge (retrieved facts)
  • A validated skill bundle's ordered steps
  1. Lane resolution. Each piece routes through its admitted lane: memory through admitted memory-serving lane, knowledge through admitted knowledge-serving lane, skill through admitted skill-serving lane.
  2. Family truth ordering preserved. Core kernel artifacts come first; advisory views compose afterward in their declared serving order.
  3. No implicit merging. The model sees clearly-segregated content per lane.
  4. Prompt assembled. Output respects every lane's declared inputs and derivation source.

Reader Scenario: A Forbidden Input Tries To Sneak In

A formatter convenience would let a caller pass arbitrary metadata into the advisory lane.

  1. Lane validates. Caller-persisted metadata is on the lane's forbidden inputs list (per C-COG-051).
  2. Fail close. Prompt serving raises explicit failure with typed reason.
  3. No best-effort rendering. The prompt is not generated with the bad metadata silently ignored — generation fails.
  4. Reviewer sees the truth. "Lane X rejected forbidden input Y."

The fail-close posture is what keeps the lane contract enforceable.

Reader Scenario: Working State Tries To Reach The Prompt

Working state (transient cognition processing data) is excluded from prompt serving by default.

  1. Working state attempted. A subsystem tries to include working state in a prompt.
  2. Reject. No admitted lane carries working state for this family.
  3. No silent demotion. The system does not quietly let the working state through as advisory.
  4. Explicit admission required. A later cognition rule must explicitly admit working state into a lane before it can flow.

What Prompt Lanes Do Not Do

  • They do not let core lanes silently merge into advisory.
  • They do not let formatter output choices override family-truth ordering.
  • They do not allow caller-persisted metadata as advisory derived views (only service-owned derivation).
  • They do not let working state into prompts without explicit admission.
  • They do not best-effort-render around lane violations.

Boundary Summary

ConcernAuthority
Lane registrytables/prompt-serving-lanes.yaml
Lane registry rulesC-COG-050
Separation + derived metadataC-COG-051
Failure modelC-COG-052
Service that owns prompt assemblyCognition PromptService

Source Basis

Nimi AI open world platform documentation.