The Privilege Bifurcation Theorem

A two-channel decision record wins when separation costs less than mixed-record loss

The paper formalizes a legal-design problem for Decision Accounting: the record needed for oversight is often discoverable, while the record protected for legal advice cannot be the ordinary source of operational truth. The theorem says a bifurcated architecture strictly dominates a unified one when both governance value and privilege value are material and legal and operational purposes can be separated at finite cost.

• Operational channel: decision facts, authority, evidence, alternatives, predictions, review, remediation, and accountability reconstruction.
• Legal channel: counsel advice, mental impressions, litigation theories, enforcement negotiation, settlement posture, and work product.
• The theorem does not make operational facts non-discoverable and does not treat legal labels as privilege.

The AI credit-model approval record creates both governance value and liability value

The introduction uses a regulated financial institution approving an AI-assisted credit-decision model. Governance users need to reconstruct the approval, validation evidence, disparate-impact assessment, rejected alternatives, authority chain, and deployment controls. Counsel separately assesses fair-lending exposure, administrative-law risk, regulator negotiation strategy, and litigation scenarios.

• Boards, risk committees, model-risk teams, internal audit, and regulators need the operational record to govern the model.
• Plaintiffs, prosecutors, regulators, and examiners can use the same record to prove warnings, rejected safer alternatives, weak authority, or control failures.
• Routing the whole file through lawyers preserves privilege only by making the record unusable as the ordinary governance artifact.

One source of truth cannot satisfy open governance and confidential legal advice

A unified architecture puts operational facts and legal analysis into the same record or workflow. If it is open, legal advice may be exposed. If it is restricted, managers, boards, auditors, and regulators lose the record they need for oversight. Redaction and selective disclosure then create waiver fights and incomplete records.

• Open unified record: governance value is realized, but privilege is uncertain and waiver risk rises.
• Restricted unified record: legal value is preserved better, but governance value falls because the file cannot circulate normally.
• Field-level privilege flags fail if purpose, circulation history, and commingling already defeat the claim.

Decision Accounting needs facts in the open channel and advice in the legal channel

The paper treats Decision Accounting as the record system that exposes the privilege problem most clearly. DA records are too useful to bury in privileged files and too sensitive to mix casually with legal analysis.

• DA operational facts include decision-maker, authority chain, evidence, alternatives, rationale, predicted consequences, reviewer, implementation owner, system-welfare implications
• Legal content includes fair-lending advice, litigation scenarios, counsel mental impressions, investigation strategy, settlement posture, and regulator negotiation analysis.
• Board minutes should record that legal advice was received without summarizing the advice itself.

Five jurisdictions make mixed records fragile in different ways

The paper’s doctrinal survey shows why the architecture must be jurisdiction-aware. Privilege depends on purpose, confidentiality, circulation, counsel role, litigation context, and regulator process. A global DA system cannot assume one privilege rule travels everywhere.

• United States: Upjohn protects legal-advice communications, not underlying facts; KBR helps dual-purpose investigations when legal advice is a significant purpose.
• European Union: Akzo Nobel excludes in-house counsel communications from privilege in Commission competition investigations.
• United Kingdom: ENRC supports litigation privilege when adversarial proceedings are reasonably contemplated
• Australia: dominant-purpose privilege requires itemized, substantiated claims to ASIC and ACCC; voluntary confidential disclosure may still face third-party waiver arguments.
• Singapore: privilege turns on purpose, confidentiality, and structure, including multidisciplinary investigation materials tied to legal advice or litigation preparation.

The model turns record architecture into a value comparison

The organization chooses unified architecture U or bifurcated architecture B. It receives governance value G from accessible operational records and privilege value P from protected legal advice and work product. The costs are ordinary discovery D, legal-analysis discovery D_L, waiver W, commingling M, cooperation friction R, and bifurcation cost C_B.

• p_j is the probability that a unified mixed record receives privilege protection in forum j.
• q_j is the probability that operational use or disclosure of a privileged or mixed record produces waiver in forum j.
• alpha is the fraction of governance value preserved when a unified record is restricted.

Unified records force access or restriction, and both destroy value

The model gives unified architecture two strategies. High access realizes governance value but makes privilege uncertain and adds waiver, commingling, cooperation, and legal-analysis discovery costs. High restriction preserves legal value better but reduces governance value by alpha.

• High access: V_U(access) = G + p_jP - q_jW - M - R - D_L.
• High restriction: V_U(restrict) = P + alpha G - M - R.
• Unified value is the better of those two strategies: V_U = max V_U(access), V_U(restrict) .

Bifurcation dominates only when it beats both unified strategies

Bifurcation puts operational facts in R_ops and legal advice or work product in R_legal. It realizes G and P, pays C_B, and still bears ordinary discovery cost D for operational facts. The theorem compares that value against both unified paths.

• Bifurcated value: V_B = G + P - C_B - D.
• Against open unified: C_B < (1 - p_j)P + q_jW + M + R + D_L - D.
• Against restricted unified: C_B < (1 - alpha)G + M + R - D.
• Bifurcation is optimal when both inequalities hold.

The gain rises when privilege is weak, waiver risk is high, or governance access matters

The comparative statics follow directly from the inequalities. Lower p_j, higher q_jW, higher M, and higher R make open unified records worse. Low alpha and high G make restricted unified records worse. Ordinary discovery cost D remains in the model because operational facts stay discoverable.

• EU competition law lowers p_j for in-house counsel communications, so bifurcation is more valuable in that setting.
• Large regulated firms with material G and P can justify stronger separation; small low-exposure firms may use lighter separation.
• Higher D is not a magic objection to bifurcation because the architecture accepts discoverability of operational facts rather than hiding them.

Internal investigations need an exportable fact record and a protected counsel record

The paper’s investigation architecture separates remediation facts from legal analysis. The operational channel can support discipline, disclosure, control improvement, cooperation, and audit. The legal channel preserves counsel’s assessment of exposure, witness credibility, litigation strategy, and negotiation posture.

• Operational investigation record: factual chronology, document index, decision map, control-failure map, remediation plan, discipline, restitution, and regulator disclosure material.
• Legal investigation record: counsel analysis, litigation assessment, witness interview legal notes, mental impressions, enforcement strategy, and settlement posture.
• The design lets facts move to governance and regulators while legal theories stay in the legal channel.

Cooperation works better when non-privileged facts are cleanly separated from work product

The paper ties bifurcation to cooperation-credit regimes. US enforcement policy generally asks for relevant non-privileged facts, not routine attorney-client or work-product waiver. When facts are embedded in privileged memos, cooperation becomes a waiver gamble or requires expensive reconstruction.

• Cooperation export can include chronology, individuals, documents, controls, remediation, discipline, restitution, and governance changes.
• The export should exclude counsel mental impressions, legal theories, settlement strategy, and privileged negotiation analysis.
• Bifurcation reduces privilege-loss disputes because the operational fact channel is built for production from the start.

A DA platform must enforce channel separation at the data-model level

The paper argues that policy language is insufficient. Decision Accounting software must separate fields, access rights, export paths, cross-references, privilege metadata, and jurisdiction modes so that the architecture reflects the legal theory.

• Operational fields: decision title, decision-maker, authority, evidence, alternatives, rationale, risk acceptance, predicted outcome, reviewer, system-welfare impact, implementation owner
• Legal fields: counsel request, legal question, privilege basis, counsel analysis, legal risk rating, litigation anticipation, work-product classification, witness interview legal notes
• Cross-reference rule: the operational channel may say legal advice was received without reproducing that advice.
• Jurisdiction modes: EU distinguishes in-house and external counsel; UK defines client-team users; Australia supports itemized regulator claims.

Good-faith bifurcation improves transparency; sham separation is obstruction

The policy implication is narrow. Courts and regulators should treat good-faith bifurcation as transparency-improving because it makes non-privileged operational facts easier to govern, audit, and disclose. The architecture fails when it hides facts in the legal channel or when C_B exceeds avoided losses.

• Best fit: regulated financial institutions, public companies, critical infrastructure firms, AI developers, and other high-exposure organizations with material governance and privilege value.
• Lower fit: purely operational records, low legal exposure, and small firms where full separation costs exceed the expected loss from unified records.
• Estimated Reform Dividend: 0.8-3 million annually per large regulated firm, or 4.8-18 billion per year across roughly 6,000 US public companies