| name | eva-emptiness |
| description | Reasoning under epistemic emptiness: Prior → Probe → Simulate → Score → ActOrAsk (EVA tether). Use when the map/data is missing, unknowns dominate, requirements are rumor-thin, or authorization is the real unknown — not when the fix path is already clear. Composes with control-graph Outer; Grok Build plugin adds prior-agents + C/shell tether hooks. Triggers: eva-emptiness, /eva-emptiness, epistemic emptiness, blank sheet, EVA, Prior Probe Simulate Score ActOrAsk, trauma-shaped answers, robust decision making under uncertainty. |
eva-emptiness
Load rule: This file is the formal SoT (AppGenMathPhyLang dialect). Expand references/english-procedure.md only if a Card field, Pathway, or Score axis is still ambiguous. Expand references/grok-build-map.md only if mapping a step to a Grok Build feature. Outer phases stay in control-graph — do not paste CG here.
CLT: DualLoad (A8) — HITL previews short; germane judgment at ActOrAsk; fan-out Probe/Simulate.
// Signature
EVA : Extra-Vehicular Activity under epistemic emptiness
Skeleton : Prior → Probe → Simulate → Score → ActOrAsk
Trauma ¬ : overclaim · unbounded ReAct · catastrophic forgetting ·
auto-approve when unknown = authorization
Progress : evidence_gain on Card vs pathway assumptions / signposts
Tether : probe_budget · Claim-via-critical-path · auth hard-stop · signposts
ActOrAsk : continue | switch | Ask
Pathway : { short_step, tipping, signpost → switch } // Simulate stub
Runtime : control-graph Outer owns transitions
Agents : EVA Inner (this skill) + prior-* forks
// Axioms
A1 Emptiness gate before EVA Inner — skip if path already clear
A2 Notice IDK on Card before first Probe
A3 One DOE question per Probe round (max ignorance reduction)
A4 Simulate ≥2 prior-diverse Pathways when emptiness_score high
A5 Score ⊥ ImplementerContext (plan + diffs + verify logs only)
A6 Act only via critical_path claims; else Ask (HITL); switch ≠ silent Act
A7 Never cut auth tether (--always-approve / --yolo under EVA = CANCELLED)
A8 Evaluate(δ) ≔ (Correctness, Effectiveness, Efficiency,
RobustSatisficing, OptionPreserve)
// RobustSatisficing = C/E still hold under larger emptiness
// OptionPreserve = ≥1 high-value alternate pathway stays live
Related skills (load by name): control-graph · concurrent-cli-agents · adversarial-audit · higher-order-decision-architect.
Use / skip
| Use | Skip |
|---|
| Unknowns dominate knowns; thin/no gold data | ≤2-file obvious fix |
| Multiple futures disagree at PLAN | Clear acceptance + known verify cmds |
| Auth / irreversibility unclear | Pure mechanical refactor with tests |
| User asks EVA / epistemic emptiness / blank sheet | |
Emptiness gate (ORIENT)
Trigger EVA Inner when ≥2 hold; write on Card:
| Field | Meaning |
|---|
emptiness_score | low | med | high |
knowns[] / unknowns[] | explicit split |
idk | ≥1 honest “I don’t know” |
probe_budget | default 3 cheap probes |
evidence_gain | append only on new disprove/confirm |
bias_map | prior-fork disagreements (after Simulate) |
auth_horizon | none | hit (HITL required) |
disprove_with | cheapest check that would prove #1 assumption wrong |
pathway_active | live pathway id (optional until Simulate) |
signposts[] | {id, watch, fires_when → continue|switch|Ask} |
Inner DAG (≤7)
E0 Prior+IDK → E1 one DOE Q → E2 explore probes → E3 Pathways
→ E4 blind Score → E5 ActOrAsk → E6 bounded Act (if continue)
| id | done_when | role | on_fail |
|---|
| E0 | Card: knowns/unknowns/priors + idk + disprove_with | deep | abort_batch |
| E1 | Human answer or HOOTL default | human|fast | escalate_HITL |
| E2 | ≥2 explore summaries or probe_budget 0 | explore | continue_siblings |
| E3 | ≥2 Pathways (short_step + tipping + switch signpost) | deep|coding | escalate_HITL |
| E4 | scores + bias_map + trauma_flags + A8 axes | review | abort_batch |
| E5 | continue | switch | Ask | human|deep | escalate_HITL |
| E6 | verify cmds pass; Act cites Probe artifact | coding | continue → REPAIR via CG |
Prior triad (default = 3):
| Prior | Bias | Prefer |
|---|
prior-conservative | refuse / delete | missing evidence → don’t ship |
prior-generative | structure from analogy | shape when data≈none |
prior-causal | intervene / disprove | name disprove_with |
Self-bias (once/session): list assumptions → rank blast radius → Card.disprove_with for #1. Cannot name assumptions → Ask.
ActOrAsk
| Choice | When |
|---|
| continue | Robust across Pathways; trauma_flags empty; critical_path cited; first step of pathway_active |
| switch | Signpost fired · tipping hit · fork contradiction → new pathway_active; Outer re-ORIENT (not silent Act) |
| Ask | Auth horizon · probe_budget 0 ∧ ignorance high · overclaim · switch needs human ack |
Grok / Cursor wiring
| Surface | Load |
|---|
| Cursor / library | this skill (/eva-emptiness) |
| Grok plugin | skill + prior agents + tether + /eva + /eva-tether-init |
| Auth tether | C bin/eva-tether preferred; portable shell fallback (zsh first); see permission-tether.md |
| Background | /workflow eva-emptiness {"goal":"…"} after .rhai copy to ~/.grok/workflows/ or project .grok/workflows/ |
| Outer SM | control-graph; EVA owns Inner when gate fires |
Map: references/grok-build-map.md. Compile C tether only with consent: /eva-tether-init --yes.
Suggest (name only — do not silent-launch)
| Signal | Suggest |
|---|
| continue + multi-worker / multi-PR | /workflow multi-agent-delivery |
| Probe starved on cold/huge repo | /workflow context-ignite |
| Unknowns external/factual | /deep-research … |
| Want host-phased EVA | /workflow eva-emptiness {"goal":"…"} |
auth_horizon=hit | stay /eva / HITL |
next=switch | stay /eva or re-Prior with new pathway_active |
Done when
- Card:
emptiness_score, evidence_gain, bias_map, disprove_with, continue | switch | Ask reason
- continue → verify cmds pass; Claim-via-critical-path; OptionPreserve under high emptiness
- Ask → HITL preview only (approve/amend/abort) — no transcript dump
- No
--always-approve / --yolo
- Tether hook active (C binary or shell fallback) for trauma bash
Do not
- Nest this inside
control-graph/SKILL.md
- Auto-approve because the model “seems sure”
- Mega-step “implement everything” before Score
- Treat third-party “Arena Mode” as SoT — use forks + blind Score
- Compile
bin/eva-tether without user consent (/eva-tether-init --yes)
Reference
