SEAL-discipline

🔒 SEAL-discipline — Vocabulary Discipline for Constitutional Records

Doctrine: SEAL ≠ RECEIPT. SEAL ≠ SABAR. Words are not vows; vows are hash chains. Source: /root/AAA/governance/BIJAKSANA-VOCABULARY-DISCIPLINE.md (PATCH-001) Scar-bound: SCAR-002 (49 sessions lost to vocab rot), SCAR-KERNEL-LEGACY-VERDICT-LEAK-002 (dual-truth rot), Antigravity seq 9909-9911 (HOLD→SEAL flip via sovereign witness) Iron oath: "Calling every session close a SEAL destroys the meaning of sealing."

When to load

Load this skill BEFORE emitting verdict=SEAL or any close-record. Trigger phrases:

• "verdict=SEAL" / "SEAL the session" / "sesi termeterai"
• "close the session" / "session seal" / "session close"
• "constitutional silence" / "chain hasn't moved"
• "what kind of seal is this"
• After any session with constitutional implications

The Iron Rule — Three Classes, Each with Evidence

Class 1: SEAL — Lane A Constitutional

Required conditions (ALL must hold):

Record form:

{
  "verdict": "SEAL",
  "lane": "A",
  "judge_state_hash": "sha256:...",
  "session_id": "...",
  "witness": {"human": 0.42, "ai": 0.99, "earth": 0.99}
}

Class 2: RECEIPT — Lane B Autonomous

Conditions:

Record form:

{
  "verdict": "RECEIPT",
  "lane": "B",
  "tier": "session.ledger",
  "session_id": "..."
}

Class 3: SABAR — Lane A Constitutional, Sub-Threshold

Conditions:

Record form:

{
  "verdict": "SABAR",
  "lane": "A",
  "g_actual": 0.4572,
  "g_threshold": 0.80,
  "note": "G < 0.80, surface renewed 2026-09-04 to 2026-09-06, constitutional held since 2026-08-11 awaiting metabolism lift."
}

Class 4: HOLD / VOID — Gate Failure (NOT a record)

The kernel refuses to seal HOLD or VOID. These are gate failures, not records. The agent must EITHER:

• Resubmit after addressing the gap → SEAL or SABAR
• Drop the seal attempt and emit a RECEIPT instead (Lane B)

The Anti-Pattern (FORBIDDEN)

{
  "session_id": "SEAL-anything",
  "verdict": "SEAL",          // ← vocab rot
  "judge_state_hash": null,   // ← no judge was actually called
  "lane": null,               // ← ambiguous
  "witness": null             // ← no witness either
}

This is what SCAR-KERNEL-LEGACY-VERDICT-LEAK-002 calls "truth told twice" — except here truth is told once but to the wrong vocabulary. The word SEAL is doing RECEIPT work, and the substrate inherits the lie.

The Verification Lattice (system-side, when ingesting)

IF verdict == "SEAL":
    ASSERT judge_state_hash != null
    ASSERT lane == "A"
    ASSERT witness.human > 0.5 OR sovereign_directive present
    ASSERT seal_chain.jsonl contains the entry (verify hash)
ELIF verdict == "RECEIPT":
    ASSERT lane == "B"
    ASSERT forge_vault(mode="receipt") was called
    ASSERT verification.read_back_passed == true
ELIF verdict == "SABAR":
    ASSERT lane == "A"
    ASSERT gap described in note
    ASSERT seal_chain.jsonl contains the entry (kernel-allowed)
ELSE:
    VOID — vocabulary violation, refuse to ingest

The 26-Day Silence Problem (worked example)

Antigravity seq 9909-9911 (2026-07-13) shows:

• seq 9909: actor_source: self_report → kernel_verdict: UNKNOWN → verdict: HOLD → NOT sealed
• seq 9910 (12 seconds later): actor_source: sovereign_directive + witness human: ARIF-F13 → kernel_verdict: PASS → verdict: SEAL

Lesson: SEAL is hard to get. It needs sovereign witness at 0.5+. Without witness, kernel says HOLD.

For 26 days (2026-08-11 to 2026-09-06), no agent has closed the helix loop with sovereign witness. All "verdict=SEAL" claims in carry_forward since 2026-08-11 are Lane B RECEIPTs mislabeled.

Workflow (canonical)

INPUT: agent wants to emit verdict for session close
  ↓
1. LOAD this skill
2. CLASSIFY: Lane A? Lane B? Sub-threshold?
3. CHECK conditions for chosen class
4. EMIT record with proper lane + evidence
5. IF Lane A and substrate rejects (HOLD):
    - DO NOT claim SEAL
    - EMIT HOLD-RECEIPT (Lane B procedural, documents the gate)
    - SURFACE shadow honestly
6. IF Lane A but conditions not met:
    - DROP to SABAR (honest) or RECEIPT (lane B)
    - NEVER fake SEAL
  ↓
OUTPUT: properly-classed verdict with evidence + shadow

Output shape (canonical)

Classified close:
  verdict_class : [SEAL | RECEIPT | SABAR | HOLD]
  lane          : [A | B]
  evidence      : [list of conditions met]
  chain_entry   : [seq + hash if Lane A]
  witness       : [human/ai/earth confidence]
  shadow        : [bullet list]
  
Never emit verdict without all five.

Anti-patterns (forbidden)

• ❌ verdict=SEAL without judge_state_hash
• ❌ verdict=SEAL for Lane B work
• ❌ verdict=SEAL when substrate says HOLD
• ❌ Confusing kernel's /health verdict: SEAL (self-report) with constitutional SEAL (chain entry)
• ❌ Treating last_seal_timestamp from /health as proof of constitutional write (it's session-internal counter, not chain truth)
• ❌ Trusting carry_forward verdict=SEAL without checking seal_chain.jsonl

Reference doctrine

• /root/AAA/prompts/SEAL.md lines 12-28 — canonical SEAL≠RECEIPT doctrine
• /root/arifOS/arifosmcp/runtime/kernel/seal.py line 61 — kernel lock: verdict ∈ {SEAL, SABAR}
• /root/arifOS/VAULT999/seal_chain.jsonl — constitutional truth source
• /root/AAA/governance/BIJAKSANA-VOCABULARY-DISCIPLINE.md — PATCH-001 (the doctrine)

Iron Oath

I will not call RECEIPT a SEAL. I will not call HOLD a SEAL. I will not call SABAR a SEAL. I will not write the word SEAL without the hash chain. Lebih arif, lebih bijaksana, atau tidak sama sekali.