Skip to main content

Tri-Witness - F3 Constitutional Consensus

Canonical source: 000_THEORY/003_WITNESS.md
Implementation: core/organs/_3_apex.py
Stage: 888 AUDIT . Motto: Tiga Saksi Lebih Baik Dari Satu (Three Witnesses Better Than One)

What Is the Witness System?​

The Tri-Witness system is the constitutional consensus mechanism. No verdict can be sealed unless three independent sources agree. A single dissenting witness prevents SEAL.

This is F3 in the constitutional floor system - a Mirror (feedback loop), not a Law. It does not block on its own (SABAR, not VOID), but no SEAL can be issued without tri-witness consensus >= 0.95.

The Formula​

W^3 = (S x S x S)^(1/3) >= 0.95

Where:

  • S = Human witness score (sovereign evidence, user input quality)
  • S = AI witness score (internal reasoning coherence, F8 Genius)
  • S = External witness score (reality grounding, web search, external data)

The geometric mean is used deliberately. Unlike arithmetic mean, the geometric mean cannot be "rescued" by one very high score if another is near zero. If any witness scores 0, W^3 0 regardless of the other two.

The Three Witnesses​

Witness 1 - Human (Sovereign Evidence)​

The human's input quality, context richness, and explicit confirmations contribute to S. A well-specified query with clear intent raises S. An ambiguous query lowers it.

When the human provides explicit confirmation (e.g. approving an 888_HOLD action), S approaches 1.0 for that specific decision.

Witness 2 - AI (Internal Reasoning Coherence)​

The AI engine's own consistency across stages contributes to S. This is connected to F8 Genius: G = A x P x X x E^2. A reasoning path that is internally contradictory, or that reaches a conclusion unsupported by its own stated premises, lowers S.

Witness 3 - External (Reality Grounding)​

External sources - web search (Brave API), citations, referenced documents - contribute to S. The reality_search tool exists specifically to raise S. A claim made without any external grounding has low S.

The Panopticon Principle​

There are no secrets between organs.

All witness activity is visible across the full pipeline. The AGI Delta, ASI Omega, and APEX Psi engines each contribute their scores to the tri-witness calculation at stage 888. No engine can pass a verdict privately - every judgment is visible to the other two.

This prevents:

  • Confirmation bias between engines (they are thermodynamically isolated until stage 444, then their scores are exposed to each other for the consensus calculation)
  • Any single engine becoming the sole authority on a verdict
  • Hidden reasoning that bypasses constitutional floors

What Happens When W^3 < 0.95?​

W^3 ScoreVerdictAction
>= 0.95Proceed to SEALAll three witnesses agree
0.85-0.95PARTIALProceed with documented warning
0.70-0.85SABARPause; gather more evidence; retry
< 0.70VOID (escalated)Rejected; return to earlier stage

Raising W^3 in Practice​

To raise S (external grounding):

# Enable Brave Search API for reality grounding
export BRAVE_API_KEY=your-key
# The reality_search tool will be called automatically when F3 is weak

To raise S (human witness):

  • Provide more specific queries with explicit intent
  • Approve 888_HOLD actions explicitly rather than leaving them pending
  • Supply context documents or citations with the query

To raise S (internal coherence):

  • Ensure the query does not contain contradictory premises
  • Use the full pipeline (stage 000999) rather than individual tool calls
  • Check that floor scores from earlier stages are not borderline

Stage 888 - The AUDIT​

Tri-witness consensus is computed at stage 888 (AUDIT) by the APEX Psi engine. This is the final constitutional checkpoint before stage 999 (SEAL). The verdict at 888 - SEAL, SABAR, VOID, or 888_HOLD - is what gets committed to VAULT999.

Source: core/organs/_3_apex.py . 000_THEORY/003_WITNESS.md