Model architecture & data foundations

The basis of the Oracle

How the AIOOJ litigation outcome model is constructed, what data underlies it, how co-dependencies are managed, and what the outputs mean

Contents
Overview

What the model is

The AIOOJ Litigation Outcome Model is a structured probability engine built specifically for Harrison v Aegon / Transamerica Entities in the NSW Supreme Court Commercial List. It answers a single practical question: given everything known about the facts, the law, the defendant's conduct, and the judicial environment — what is the realistic probability distribution of outcomes, and what does that distribution mean for litigation strategy, settlement positioning, and counsel briefing?

The model is not a legal opinion. It does not predict what a court will decide. It translates legal analysis into structured probability estimates and produces a distribution of possible outcomes ranging from nil recovery to full recovery of the $100,743,642 claim. The output is a planning tool, not a guarantee.

Core principle: The model converts legal judgements into numbers so that strategic decisions can be made rationally — not so that numbers can substitute for legal judgement. Every number in this model is ultimately derived from a legal assessment. The model makes those assessments explicit, comparable, and stress-testable.

The model has been developed through nine versions incorporating successive layers of empirical data, benchmark calibration, and architectural improvements. Version 9.1 is the current production version.

Structure

Three-stream architecture

The model's most important structural feature is that the plaintiff's claim operates across three independent recovery streams. Total defeat — a nil outcome — requires the defendant to defeat all three simultaneously. The plaintiff needs only one to succeed.

I
Deed stream
Unconditional guarantee under the 1998 Deed of Guarantee and Undertaking. Demand made 25 November 2025. TLIC (NAIC 69078) as successor guarantor. Two independent routes: L2 direct summary judgment, or L5 demand accrual defeating s.63.
II
Contract stream
Employment Agreement (EA) and Consulting Agreement (CA) commission underpayments. Requires at least one of L3 / L6 / L7 (contract merits) AND at least one of L4 / L5 / L8 (limitation route). EA principal $39.97m; CA principal $16.73m.
III
Korea floor
Korea rate reduction claim (L1). Operates entirely independently of the Deed and Contract streams. Secures a minimum recovery floor of approximately $13.7m in effectively all scenarios where Korea succeeds — regardless of the outcome of any other stream.

The combined failure probability — all three streams failing simultaneously at base values — is less than 0.007%. This is not optimism. It is the mathematically correct result of three high-probability independent streams operating under OR-gate logic.

The asymmetric proof structure: The defendant must defeat every pathway simultaneously to achieve a nil outcome. The plaintiff needs only one to succeed. This structural asymmetry is not an assumption — it is a legal fact about how the claim is constructed.
Probability variables

Probability inputs P1–P9

Nine probability variables drive the model. Each represents a discrete legal issue assigned a Low / Base / High value. The Base value is the central estimate derived from case law analysis, the empirical dataset, and the specific facts of this proceeding.

InputLegal issueBaseRangePrimary data anchor
P1 / L1Korea rate reduction success90%85–95%Contract clarity; abandoned explanations; admitted errors
P2 / L2Deed / summary judgment (ultimate)89%82–94%Guarantee real anchors 1.00 mean; s.54 acknowledgment at 95% certainty
P3 / L3EA breach methodology82%75–88%Admitted errors; $3.319m corrective payment; retracted denials
P4 / L4s.55 postponement success80%70–86%Active obstruction + illegal NDA; dataset calibration 0.80
P5 / L5Demand accrual / s.63 rebuttal85%80–90%Wardley Australia; three-limb s.63 rebuttal; December 2025 demand
P6 / L6Consulting Agreement stream81%72–85%Contract/Construction real anchors 1.00 mean; dataset uplift 0.77→0.81
P7 / L7Component B generalisation71%62–75%Inferential; weakest link; conservative; widest range
P8 / L8Route C / fresh fraud78%72–84%Void NDA conditions; consciousness of guilt; s.249K Crimes Act
P9Conduct / indemnity costs factor78%70–88%Combined conduct pattern; 65–80% research range; costs overlay only

P9 is isolated from the core merits simulation. It operates exclusively on the settlement band calculations and costs-adjusted recovery figures. This prevents double-counting — the conduct facts are already embedded in P4 and the environmental scoring.

Boolean logic

Win function logic

The seven outcome paths are derived from the nine inputs through defined Boolean logic. The overall win function combines all three streams under an OR gate — any one succeeding produces a material recovery.

// Stream definitions Deed_ultimate = L2 OR L5 Contract_stream = (L3 OR L6 OR L7) AND (L4 OR L5 OR L8) Korea_floor = L1 // Overall win — any stream succeeding Overall_win = Deed_ultimate OR Contract_stream OR Korea_floor // Expanded — equivalent formulation Overall_win = 1 − (1−Deed_ult) × (1−Contract) × (1−Korea) // v9.1: L8 wired into Contract limitation gate alongside L4 and L5 // Previously L8 was present in inputs but absent from win function gates

Note that L8 (Route C) is correctly wired as an independent OR gate in the Contract stream's limitation layer alongside L4 and L5. This was corrected in v9.1 — the original v9 Shapley sheet showed L8 at zero marginal contribution because it was absent from the win function gates despite being scored in the simulation inputs.

Empirical foundations

Data foundations

The model draws on two categories of data: case-specific data derived from the facts, law, and conduct of this proceeding; and environmental data reflecting the external context. Understanding the difference is essential to reading the outputs correctly.

Benchmark case set

18 cases from the HCA, NSWCA, and NSWSC were manually scored for relevance across five dimensions: claim construction (CDG), accounting/withholding (AWV), contract interpretation (CI), limitation/accrual (LAC), and early disposition (SJE). Key authorities include:

  • Jones v Dunkel (1959) 101 CLR 298 — adverse inference from document withholding. Directly applicable: defendant withheld premium data through negotiations and the formal Clause 6.7 demand period.
  • Chaplin v Hicks [1911] 2 KB 786 — a defendant cannot defeat a damages claim by relying on uncertainty of quantum where that uncertainty was caused by the defendant's own conduct.
  • Wardley Australia Ltd v Western Australia (1992) 175 CLR 514 — demand-based accrual supporting the s.63 rebuttal.
  • Mount Bruce Mining v Wright Prospecting [2015] HCA 37 and Electricity Generation Corp v Woodside Energy [2014] HCA 7 — commercial construction authorities supporting P3 and P6.
  • Price v Spoor [2021] HCA 20 — limitation defence may be affected by agreement; limitation bars remedy rather than jurisdiction. Primary real anchor for P4.
  • RHG Mortgage Ltd v Ianni [2015] NSWCA 56 — Jones v Dunkel inference applied to document withholding; confirmed inference can be outcome-determinative.
100-case empirical dataset

14 verified real anchor cases from official HCA/NSWCA sources are used in calibration, each weighted at 1.0. 86 structured comparator rows (synthetic range-fills) are retained for reference only and carry zero weight in the v9 calibration engine. The Bootstrap_Calibration sheet derives 95% confidence intervals using only the 14 real anchors — this produces deliberately wider intervals that honestly reflect the limited real-case sample available.

Calibration blend (60/40 rule): When the empirical dataset produced values differing from case-specific bases, the model applied a 60/40 blend — 60% case-specific legal analysis plus 40% empirical dataset signal. This preserves the specific facts of this proceeding (admitted errors, illegal NDA, corrective payment) while anchoring to the broader base rate.
NSW court statistics

NSW Supreme Court Provisional Statistics 2024 inform time-to-trial assumptions. FCA empirical data (2023–24 and 2024–25 annual reports; 5-year longitudinal study across 19,483 cases) anchors the 1.5-year base and the three-branch disposition model: Summary Judgment 30% / 7 months; Negotiated 55% / 9 months; Full Trial 15% / 22 months.

Index scoring

Structural Advantage Index & Environmental Factors Index

Two scored indices apply uplifts to the base probability inputs. Both use a compression formula to prevent artificial inflation:

Effective score = 50 + (Raw score − 50) × 0.7 SAI uplift per input = (Compressed SAI − 50) × 0.001 EFI uplift per input = (Compressed EFI − 50) × 0.0005 // half the SAI weight
SAI — Structural Advantage Index
Measures case-specific structural advantages: documentary dominance (95/100), data asymmetry (100/100 — all determinative data held by defendant), witness asymmetry (90/100), narrative coherence (85/100), determinative issue structure (90/100), complexity reverse-scored (60/100). Raw SAI = 91.25; compressed ≈ 78.9; uplift ≈ +2.9% per input.
EFI — Environmental Factors Index
Measures the litigation environment: regulatory/supervisory pressure (85), corporate redomiciliation risk (80), judicial environment (78), defendant litigation fatigue (75), settlement BATNA asymmetry (82), reputational exposure (70), Dutch/international dimension (72). EFI receives half the SAI transmission weight — environmental factors support but do not override legal merits.

Hard caps prevent any input exceeding its judicial caution ceiling regardless of uplift stacking: P2 Deed caps at 0.92; P4 s.55 caps at 0.82; P3 EA caps at 0.88. P9 (costs factor) receives SAI uplift only — EFI uplift is excluded to prevent double-counting with the conduct facts already embedded in EFI scoring.

Integrity architecture

Anti-overfitting controls

A model with multiple positive reinforcing factors carries a serious risk of counting the same advantage multiple times. Eight deliberate score reductions were applied during development. Near-certainty of overall recovery persisted through all testing — confirming it is structural, not inflated.

  1. SAI transmission weight halved: from ×0.002 (early version) to ×0.001. Early version added +8.4% per input from SAI alone.
  2. EFI transmission weight halved again: from ×0.001 to ×0.0005. Environmental factors are supporting not primary.
  3. Compression formula applied: SAI 91.25 compresses to 78.9 (reduces uplift from +4.1% to +2.8%). Reflects judicial caution at high structural scores.
  4. P2 Deed hard-capped at 0.92. Spencer v Commonwealth SJ caution; corporate succession chain uncertainty.
  5. P4 s.55 hard-capped at 0.82. Courts apply s.55 cautiously; active obstruction ≠ guaranteed success; Briginshaw standard applies.
  6. P4 base reduced 0.82→0.80 per 100-case dataset: Limitation/Timing weighted mean = 0.624 across all cases.
  7. 60/40 calibration blend on every input: 60% case-specific legal analysis + 40% empirical dataset signal.
  8. EBI double-count prevention: EBI (Stage 1 environmental base rate) documented only — not separately added since it is already contained within EFI.
Interpretation

Reading the outputs correctly

P10 — the absolute worst case input: A tenth analytical input (P10) sits below the main model. It contains six event sliders (E1–E6) representing the individual failure probabilities of the six legal positions that must all fail simultaneously for the worst case to materialise. At defaults the joint failure probability is 0.027% (~1 in 3,700). The 12-year floor quantum is approximately $35m. The P10 stress tab on the heat map shows this worst case applied across all P2/P4 stress combinations simultaneously. The floor of $35m represents the defendant’s minimum rational exposure regardless of how every other argument is decided.

The model produces near-certain overall win probability (~99%+) at base inputs. This is the correct structural result, not an artefact of modelling. Three independent recovery streams each with high individual probability make simultaneous total defeat near-impossible. The genuine uncertainty is concentrated not in whether the plaintiff recovers, but in where in the recovery distribution the outcome lands.

The five key output numbers
OutputWhat it meansHow to use it
Overall win PProbability of any material recovery above $0.01mPrimary settlement leverage number. The defendant must price this into any offer.
Expected recovery (EV)Probability-weighted average recovery across all outcomes including nilAnchor for settlement negotiations. A rational defendant will not offer materially less than this without specific contrary arguments.
Settlement resistance floor (P10)Portfolio-weighted beta P10 — level below which only 10% of winning-scenario outcomes fallThe plaintiff's methodologically defensible walk-away threshold. Do not accept below this.
Branch-weighted PVPresent value across SJ (30%/7mo), Negotiated (55%/9mo), Trial (15%/22mo) branchesCorrect number for settlement timing decisions. Higher than a simple point-estimate PV because the SJ branch pulls the distribution forward.
Costs-adjusted exposureEV plus P9-weighted expected indemnity costs orderThe defendant's true total economic exposure including costs risk.
P2×P4 heat map — P10 stress tabEV $m tab: Deed-stream quantum (P2 × [P4×$100.7m + (1−P4)×$20m]) + Korea floor. Genuine gradient ~$51m to ~$92m across grid — P2 drives row variation, P4 drives column variation. P10 stress tab: stream decomposition with worst-case event failures — ~$50m to ~$73m at defaults.Worst-case positioning — mediation floor anchor
P10 absolute worst caseSix event sliders (E1–E6, 10–20% each). Joint failure probability 0.027% (~1 in 3,700 after 50× correlation). 12-year floor quantum ~$35m. Drives P10 stress tab in real time.Settlement floor — funder briefing
What the model cannot tell you
  • The exact probability of any specific outcome — the model produces distributions, not point predictions.
  • The response of any individual judge — JAR is a directional adjustment, not a prediction.
  • The effect of the Two Sets of Books reserve evidence (TAB_38_3), which is not currently reflected in the inputs.
  • The defendant's internal settlement calculus — EFI models structural settlement pressure, not Aegon's board position.
  • Post-judgment appellate risk — the model is calibrated for first-instance outcomes only.
Version history

Version 9.1 — three analytical fixes

Version 9.1 applied three targeted fixes identified in a post-v9 audit. None alters the core probability inputs, scenario weights, or recovery values.

  1. L8 Shapley zero-contribution defect corrected. Route C (L8) was scored in the simulation inputs but absent from the Boolean win function OR gates. The Shapley sheet therefore showed L8 contributing zero — a formula defect, not a legal conclusion. L8 is now correctly wired as an independent OR gate in the Contract stream's limitation layer (alongside L4 and L5), consistent with its legal role as an independent limitation extension under Route C / s.249K Crimes Act 1900 (NSW).
  2. Two-way sensitivity table replaced: P4×P5 → P2×P4. The original P4×P5 table was analytically flat — the OR-gate dominance of P2 at 0.89 suppressed all sensitivity to P4 and P5. The revised P2×P4 table tests simultaneous Deed underperformance and limitation failure — the genuine stress diagonal. Even at maximum joint stress (P2=0.60, P4=0.55), material recovery persists via Korea and demand accrual.
  3. Settlement resistance floor surfaced on Dashboard; Stress_Floor sheet added. The Quantum_Distribution P10 floor is now a named KPI on the Dashboard. The new Stress_Floor sheet provides Scenario A (stressed inputs, OR gates preserved) and Scenario B (correlated collapse, AND gate — extreme lower bound) adversarial analysis.
Transparency

Known limitations

The following are acknowledged gaps in the empirical foundation. They are documented here in the interest of full transparency for counsel review.

  • NSW Commercial List plaintiff vs defendant success rates at trial — courts do not publish outcome statistics by party type; calibration is inferred from case benchmarking.
  • Quantitative Jones v Dunkel application rates — no systematic dataset tracks how frequently the inference proves decisive; calibration relies on doctrinal analysis.
  • s.55 success rates specifically — concealment extension cases are not separately categorised in court statistics.
  • Insurance industry commission dispute precedent — no reported NSW cases are directly analogous to the commission underpayment structure in this proceeding.
  • The copula correlation matrix (Copula_Correlation sheet) provides analytical bounds on the true EV but is not yet implemented as a live Monte Carlo correlation engine — a full Gaussian copula simulation would require Python/R.
The following model assumptions have been flagged for verification by Corrs before the model is used for final settlement positioning: s.54 characterisation of the $3.319m payment (CP-1); KAH exhibit numbering conflict between 06-04 and 06-10 (CP-2 through CP-9); Deed of Guarantee "supported by indemnity" characterisation in 06-14 pending verification against Deed text; corporate succession chain JCPLIG → JCPLIC → TLIC assumed established, to be confirmed against 06-14 v3.

Model Narrative v6 — March 2026 — Confidential — Prepared for Corrs Chambers Westgarth and Counsel — Not legal advice