Energy Economics · September 2026 · 5 min read
Eliminating Hallucinations in Upstream Valuation: Deploying Type-Safe Data Contracts for Energy Asset Due Diligence
Jean Pierre Assiana
Founder & Principal Architect, ChemData & AI Consulting Ltd
The economics of upstream oil and gas are fundamentally unforgiving. A single miscalculated parameter in a Production Sharing Contract (PSC)—an unverified cost-recovery cap, incorrectly phased royalty rate, or unindexed corporate tax tier—can shift an asset's Net Present Value (NPV) by tens of millions of dollars, transforming an apparently accretive acquisition into a stranded liability.
As energy operators and private-equity allocators increasingly use Large Language Models (LLMs) and autonomous agents to accelerate commercial due diligence, a dangerous technical bottleneck has emerged: stochastic drift.
Generic generative AI is probabilistic by design. Upstream corporate finance requires absolute mathematical determinism. In complex 400-page Joint Operating Agreements or Host Government Contracts, standard retrieval-augmented generation can produce subtle, silent hallucinations: a 12% royalty assigned to the wrong year, a negative tax rate, or a CAPEX schedule misaligned with the field decline curve.
At ChemData & AI Consulting Ltd, we address this vulnerability by decoupling semantic extraction from quantitative execution through type-safe data contracts.
The fragility of upstream fiscal waterfalls
A modern PSC is a chain of interdependent fiscal rules. An error at one boundary propagates through every downstream calculation.
Gross Hydrocarbon Revenue
│
▼
[ Royalty Deduction ] ────────────► Government Take
│
▼
[ Cost Recovery Ceiling ] ────────► Contractor CAPEX / OPEX
│ subject to cap and carry
▼
[ Profit Oil Pool ]
│
┌─────┴──────────────────┐
▼ ▼
Government Profit Contractor Profit
Share Share
│ │
│ ▼
│ [ Corporate Income Tax ]
│ │
└────────────┬───────────┘
▼
Total Government TakeThe cost-recovery bottleneck. Capital recovery is restricted by a dynamic ceiling, often 50% to 70% of gross revenue. Unrecovered balances must be tracked precisely and carried forward.
The profit-oil split. The remaining pool is divided using negotiated sliding scales or cumulative production tiers.
The tax layer. Contractor profit oil is taxed under host-country rules, often with complex ring-fencing provisions.
If an ingestion pipeline drops a boundary condition or misreads an unrecovered-cost carryforward, the resulting DCF, Internal Rate of Return (IRR), and payback period are invalid. In M&A or a licensing round, executing a model on unvalidated output is unacceptable.
Enforcing type-safe validation at pipeline boundaries
Autonomous extraction agents should never interface directly with a financial engine. Every payload must pass through a programmatic gateway that enforces strict types, field boundaries, lifecycle parity, and cross-variable checks. Non-deterministic outputs are intercepted and audited before they can alter a valuation.
from typing import List
from pydantic import BaseModel, Field, model_validator
class PSCFiscalContract(BaseModel):
"""Type-safe data contract enforcing structural and fiscal bounds
on LLM extraction payloads for upstream asset valuation."""
asset_name: str = Field(description="Target concession or license block")
oil_price_deck: float = Field(gt=0.0, description="Base-case price deck ($/bbl)")
royalty_rate: float = Field(ge=0.0, le=0.40, description="Royalty 0%–40%")
cost_recovery_cap: float = Field(ge=0.10, le=0.90, description="Cost-recovery ceiling 10%–90%")
govt_profit_share: float = Field(ge=0.0, le=1.0, description="Government profit-oil split")
tax_rate: float = Field(ge=0.0, le=0.60, description="Corporate income tax 0%–60%")
opex_per_bbl: float = Field(ge=0.0, description="Direct lifting cost ($/bbl)")
production_profile: List[float] = Field(description="Annual production (MMbbl)")
capex_profile: List[float] = Field(description="Annual CAPEX schedule ($M)")
discount_rate: float = Field(default=0.10, ge=0.05, le=0.25, description="Asset WACC")
@model_validator(mode="after")
def validate_temporal_parity(self):
"""CAPEX and production vectors must share one lifecycle horizon."""
if len(self.capex_profile) != len(self.production_profile):
raise ValueError(
f"Lifecycle Horizon Mismatch: CAPEX vector length "
f"({len(self.capex_profile)}) does not match production vector "
f"length ({len(self.production_profile)})."
)
return selfAutonomous self-correction: the adversarial feedback loop
Type safety becomes more powerful when combined with stateful agent orchestration. If an extraction agent parses an ambiguous stabilization clause or malformed parameter, the boundary layer turns the validation exception into precise diagnostic feedback and routes it back to the agent.
[ Unstructured PSC PDF ]
│
▼
[ Legal Extraction Agent ]
│ extracts cited parameters
▼
[ Pydantic Gateway Node ] ◄──────────────┐
│ │
├── Fails bounds ──────────────┘
│ "tax_rate: -0.30 violates ge=0.0.
│ Re-parse Section 14.2."
│
▼ Passes bounds
[ Deterministic DCF Valuation Engine ]
│
▼
[ Verified NPV, IRR and Payback Metrics ]The model re-evaluates the cited source section, corrects the parameter, and submits a clean data structure for deterministic execution. The system does not ask a language model to perform the valuation; it uses the model to extract evidence under controlled, testable constraints.
The commercial outcome: precision at speed
For independent E&P companies evaluating farm-ins, private-equity teams auditing reserve filings, and trading houses assessing asset credit risk, this architecture delivers three immediate benefits.
- Auditability and governance. Every financial input maps to a cited clause in the underlying contract and an explicit, versioned code constraint.
- Rapid multi-regime benchmarking. Investors can compare bidding rounds across jurisdictions in minutes rather than weeks.
- Mathematical integrity. Deterministic financial algorithms—not probabilistic prose—drive NPV, IRR, payback, and government-take decisions.
Engineering your commercial edge
ChemData & AI Consulting Ltd builds the architecture that bridges frontier AI capability with the exactitude demanded by physical energy and capital markets.
Whether you need to automate commercial due diligence, stress-test late-life abandonment liabilities, or deploy auditable multi-agent infrastructure, we deliver software grounded in engineering rigour and deep energy-domain expertise.
Discuss your energy architecture
Explore a bespoke commercial due-diligence, fiscal-modelling, or risk-mitigation engagement.