How Identity Anonymization Works in Round 2 of the Council of High Intelligence

Round 2 anonymizes panel members by mapping real identities to abstract labels like "Member A" to suppress conformity bias and ensure arguments are evaluated purely on quality rather than source prestige.

The Council of High Intelligence employs a sophisticated multi-round deliberation protocol defined in SKILL.md where identity anonymization serves as a critical safeguard against social conformity. In Round 2, the coordinator transforms identifiable Round 1 outputs into anonymized submissions, enabling panel members to cross-examine arguments without knowing which specific agent produced them.

How Identity Anonymization Works in Round 2

According to the coordinator protocol in 0xNyk/council-of-high-intelligence, the anonymization process follows four distinct steps implemented in SKILL.md:

Stable Label Mapping

The coordinator creates a deterministic mapping from real panel members to abstract labels (Member A, Member B, etc.) based on the order members appear on the panel. This mapping is defined in lines 354-362 of SKILL.md and ensures consistent label assignment throughout the deliberation.

Header Rewriting and Self-Reference Removal

Each Round 1 output undergoes surgical editing to replace identifying information. The original header (e.g., "Socrates") is replaced with the assigned abstract label, and in-body self-references like "As Socrates, I..." are rewritten as "As Member B, I..." or stripped entirely (lines 361-363).

Private Mapping Retention

The real identity-to-label mapping remains confined to the coordinator's internal state. It is explicitly not exposed to members during Round 2, though it is preserved for restoration in Round 3, tie-breaking scenarios, and the final transcript generation (lines 363-364).

Anti-Conformity Directive

The Round 2 prompt explicitly instructs each member that identities are masked, directing them to evaluate arguments by quality rather than source and prohibiting attempts to infer who said what (lines 374-378).

Benefits of Identity Anonymization in Round 2

Research cited in the protocol—specifically Choi et al. (arXiv:2510.07517), Cui et al. (Free-MAD, arXiv:2509.11035), and controlled-study (arXiv:2511.07784)—demonstrates that visible identities create social conformity pressures. The anonymization strategy delivers four key benefits:

• Mitigates conformity bias: Members assess argument strength independently rather than gravitating toward majority positions or high-status voices.
• Encourages dissent: The combination of anonymization and weight-balanced voting makes minority positions easier to surface and defend.
• Preserves diverse reasoning: Building upon polarity-pair separation and provider spread from earlier steps, anonymization safeguards epistemic diversity during cross-examination.
• Improves auditability: Because the coordinator retains the original mapping, final verdicts can be audited with real names attached, ensuring transparency without compromising the deliberation process.

Implementing the Anonymization Logic

The following Python implementation reproduces the label mapping and text rewriting procedure described in Step 3 of SKILL.md:

# Example: Build the label map and anonymize Round‑1 outputs

def anonymize_round1(round1_outputs, panel):
    """
    round1_outputs: dict{name: str}   # raw texts from each member

    panel: list[str]                  # ordered list of member names

    Returns: (dict[label: str], mapping dict)
    """
    # 1️⃣ stable label mapping

    label_map = {name: f"Member {chr(65 + i)}" for i, name in enumerate(panel)}
    # 2️⃣ rewrite headers & strip self‑references

    anonymized = {}
    for name, text in round1_outputs.items():
        label = label_map[name]
        # replace header line (assumes first non‑empty line is the header)

        lines = text.splitlines()
        if lines:
            lines[0] = f"# {label}"

        # strip identity mentions like "As Socrates,"

        body = "\n".join(lines)
        for real in panel:
            body = body.replace(f"As {real}", f"As {label_map[real]}")
        anonymized[label] = body
    return anonymized, label_map

For shell-based workflows, this bash script mirrors the prompt template transformation:

# CLI‑style illustration (mirrors the prompt template in SKILL.md)

# Assume $ROUND1_DIR contains one file per member (e.g., socrates.txt)

LABELS=("Member A" "Member B" "Member C")
i=0
for f in $ROUND1_DIR/*.txt; do
    label=${LABELS[$i]}
    # Replace header and remove real names

    sed -E "1s/.*/# ${label}/; s/As (Socrates|Feynman|...)/As ${label}/g" "$f" > "anonymized/${label}.txt"

    ((i++))
done

# The resulting files are fed to Round 2 prompts.

Summary

• Identity anonymization in Round 2 replaces real agent names with abstract labels (Member A, Member B) to prevent conformity bias.
• The process involves stable label mapping, header rewriting, private mapping retention, and an anti-conformity directive as defined in SKILL.md lines 354-378.
• Research-backed benefits include reduced social conformity, encouraged dissent, preserved epistemic diversity, and maintained auditability.
• The coordinator retains the real-to-label mapping internally, restoring identities only for Round 3 and final transcript generation.

Frequently Asked Questions

What specific line ranges in SKILL.md define the anonymization protocol?

The core anonymization workflow is defined in lines 354-362 for label mapping, lines 361-363 for header rewriting, and lines 374-378 for the anti-conformity directive.

Why is anonymization limited to Round 2 rather than the entire deliberation?

Round 2 specifically involves cross-examination where conformity bias poses the greatest risk; restoring identities in Round 3 allows for accountability and attribution in the final synthesis while preserving the unbiased evaluation phase.

How does the system prevent members from inferring identities from writing style?

The protocol relies on the anti-conformity directive (lines 374-378) instructing members not to attempt inference, combined with the structural separation of polarity pairs and diverse provider selection implemented in earlier steps.

Can the anonymization mapping be audited after the deliberation completes?

Yes, the coordinator maintains the private mapping throughout the process, enabling full auditability of the final verdict with real names attached without compromising the anonymized deliberation phase.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}