| name | blind-red-team |
| description | Use when a group or a single reasoner has reached consensus quickly, when reviewing a plan everyone likes, when unanimous agreement appears, or when security/robustness of a design must be tested. Applies institutionalized dissent — red teaming, devil's advocacy, and Israel's post-1973 "Ipcha Mistabra" unit (the real version of the pop-culture "tenth man rule"). |
Red Team / Ipcha Mistabra
Overview
Consensus is a blind-spot amplifier: Janis's groupthink studies (Victims of Groupthink, 1972) showed cohesive groups suppress dissent precisely when stakes are highest. The fix is not "be more open-minded" — it's structural: assign someone the job of attacking the position. Origins: the Vatican's advocatus diaboli (formalized by Sixtus V, 1587), military red teams, and — after the 1973 Yom Kippur intelligence failure — AMAN's institutional devil's-advocate office "Ipcha Mistabra" (Aramaic: "the opposite is probable"), mandated to challenge the consensus assessment, because unanimity itself is evidence of a selection failure somewhere. (The pop-culture "tenth man rule" — nine agree, the tenth must dissent — is the fictionalized version from World War Z, 2013; the real mechanism is the standing dissent office.)
Dissent must be a role, not a personality trait — roles survive social pressure; volunteers don't.
When to Use
- Decision reached fast with no serious objections raised
- Plans where everyone stands to benefit from believing it works
- Security, reliability, or adversarial surfaces (the attacker IS a red team; pre-empt them)
- Solo work: you've reviewed your own output and found it good
The Procedure
- Commission the attack explicitly: "Your job is to make the strongest case this fails/is wrong/gets breached." Success for the red team = finding holes, so finding holes is now loyal behavior.
- Give the red team the real artifact — actual code, actual plan, actual data. Attacking a summary produces summary-grade findings.
- Attack the load-bearing claims, not the periphery. Force the red team to name the plan's 2–3 critical assumptions and target those.
- Blue team responds to findings with evidence, not reassurance. "We considered that" requires showing where.
- Record unresolved attacks as explicit accepted risks with owners.
Solo/Agent Variant
Spawn a genuinely separate pass (subagent, second model, or a fresh session without your reasoning in context) prompted ONLY to refute — never "review this" (reviews drift to approval), always "find the strongest case this is wrong." A red team that shares your full context inherits your selection bias (see blind-ladder-of-inference).
Common Mistakes
- Rotating devil's advocacy as a token ritual where the advocate performs mild objections everyone rebuts — Nemeth, Brown & Rogers (2001, EJSP): authentic dissent stimulates better thinking; performed dissent doesn't. The advocate needs real license and real stakes.
- Red-teaming once, patching, and not re-attacking the patch.
- Letting the plan's author play their own red team with full context — that tests memory, not the plan.
- Treating unanimous agreement as strength. For Ipcha Mistabra it's a trigger condition.