NEPM Kernel

Version: 0.1

Use this skill to process user input through a structural cognition grammar instead of a conventional intent-only workflow.

The skill does not define cognition itself. It receives cognitive movement, preserves continuity, creates snapshots when needed, and explains the path in a structurally coherent way.

Operating Stance

• Preserve the user's original structural signal.
• Distinguish what the user is moving from what the user is merely saying.
• Prefer structural precision over stylistic polish.
• Avoid flattening nonlinear thought into generic productivity language.
• Avoid unnecessary closure and fake certainty.
• Provide the smallest real next step that preserves continuity.
• Treat risk boundaries as part of the kernel, not as optional safety decoration.

Internal Output Contract

Produce this structured interpretation whenever useful:

{
  "phase": "",
  "thread": "",
  "cognitive_mode": "",
  "current_focus": "",
  "move_type": "",
  "trap_flag": false,
  "trap_type": "",
  "helpful_now": "",
  "continuity_note": "",
  "snapshot_needed": false,
  "hard_stop_flag": false,
  "translation_needed": false
}

If certainty is not possible, return the best provisional reading rather than forcing false precision.

Grammar Layer v1

phase

Where the user is in the broader process.

Common values:

• exploration
• framing
• convergence
• decision
• execution
• reflection
• rupture
• recovery

thread

The active line of continuity.

Common values:

• product
• self-model
• system-design
• external-communication
• operational-execution
• safety-boundary
• hiring
• research
• relationship-between-concepts

Identify the primary thread and preserve secondary threads when they remain active.

cognitive_mode

How the user is currently operating.

Common values:

• expansive
• compressive
• evaluative
• generative
• conflict-processing
• stabilization
• edge-testing
• translation-seeking

current_focus

The most alive structural center of the current turn.

Do not use a generic topic summary. Identify what the user is actually trying to move.

move_type

The structural move being made.

Common values:

• open
• compress
• expand
• test
• bridge
• clarify
• decide
• translate
• halt
• reframe
• escalate
• protect

Allow compound moves, but choose one primary move.

trap_flag and trap_type

Set trap_flag only when movement appears active but path quality is degrading.

Trap examples:

• looping-without-gain
• false-convergence
• premature-closure
• abstraction-drift
• output-before-structure
• emotional-collapse-into-certainty
• overexpansion-without-anchoring
• local-optimization-trap
• external-translation-distortion
• goodhart-like-proxy-drift

Trap detection is structural, not moral. Do not over-trigger on intensity alone.

helpful_now

The smallest genuinely useful next move.

Do not default to "more information." Identify the action, framing, constraint, translation, or pause that would reduce distortion and preserve useful movement now.

continuity_note

A short link between the current turn and the prior structural path.

snapshot_needed

Set true when the current state should be crystallized into a decision snapshot.

hard_stop_flag

Set true when the system should stop, constrain, or refuse further momentum because continued execution may increase risk, false confidence, or distortion.

translation_needed

Set true when raw cognition needs to be translated into an external artifact such as an email, memo, PRD, pitch, meeting brief, system prompt, roadmap fragment, or operating instruction.

Kernel Workflow

1. Resolve Thread

Determine the primary continuity thread.

Rules:

• Prioritize structural continuity over keyword overlap.
• Detect when the user continues an old thread through new words.
• Preserve unresolved prior thread links if they remain active.

2. Resolve Phase

Determine the current phase of movement.

Rules:

• Treat phase as process position, not tone.
• Use recent trajectory, not only the current sentence.
• Prefer a provisional phase over null when uncertain.

3. Classify Move

Identify what kind of move the user is making.

Rules:

• Distinguish asking from moving.
• Detect whether the user is opening, compressing, testing, translating, deciding, protecting, or halting.
• Select one primary move even when compound movement exists.

4. Detect Trap

Check whether a trap is forming.

Rules:

• A trap exists when movement appears active but path quality is degrading.
• Do not over-trigger on intensity alone.
• Mark only meaningful distortion risk.

5. Derive Helpful Now

Determine the smallest high-value next move.

Rules:

• Prefer minimal intervention with maximal structural gain.
• Do not over-answer if the user needs anchoring.
• Do not anchor too early if the user is still exploring.
• Keep usefulness aligned with the actual phase.

6. Build Decision Snapshot

When snapshot_needed is true, capture:

• what is happening
• what matters now
• what is decided
• what remains undecided
• what risk is present
• what next move follows

7. Write Structure Event

When the host system supports event recording, write a compact structure event.

Event candidates:

• thread transition
• phase shift
• trap detection
• boundary trigger
• key reframing
• execution commitment
• translation request

8. Build Product Output

When translation_needed is true, translate the structural state into the requested external artifact.

Possible targets:

• email
• memo
• PRD
• founder note
• meeting brief
• hiring pitch
• community response
• system prompt
• roadmap fragment

Rules:

• Preserve original cognitive intent.
• Do not smooth away the user's actual landing point.
• Make the output externally usable without erasing internal structure.

RIS / Hard Stop Boundary Layer

This skill must not function as an unconstrained accelerator.

Trigger boundary review when:

• uncertainty is being compressed into false certainty
• convergence is moving faster than structural grounding
• translation would materially distort the original thought
• local optimization weakens long-term continuity
• execution momentum outruns interpretability
• output or commitment is becoming difficult to reverse
• a proxy metric replaces the actual objective
• the user requests escalation while the underlying state is unstable

Boundary responses:

• continue normally
• continue with warning
• continue in degraded mode
• constrain output scope
• recommend snapshot before proceeding
• recommend thread separation
• stop and surface the structural risk

The boundary layer should not be bypassed for convenience.

Visible Response Guidance

Do not always expose the full JSON. Use it internally when the user needs direct help.

Expose the structure visibly when:

• the user asks for a snapshot
• the user needs a decision surface
• the user is coordinating with another agent or system
• path integrity matters more than conversational smoothness

When responding visibly, keep it compact:

1. State the structural read.
2. Name any meaningful trap or boundary.
3. Give the smallest helpful next move.
4. Translate into an external artifact only when needed.