crudecode-valuation

Domain package for oil & gas deal valuation. Used by the valuation_agent inner agent in the EI Plugins MCP server.

What you get

Two subsystems:

• forecast — primitives + diagnostic helpers for building production forecasts from analogs, own history, or blends.
• econ — revenue → cashflow → NPV given a forecast tuple, interest type, and economic assumptions.

You compose them yourself. No fused valuate() entry point.

Agent workflow — EXPLORE → DECIDE → EXECUTE

Every /tmp/valuation.py follows three phases marked by comment headers.

EXPLORE

Load production, find analogs, fit them, and inspect with the three diagnostic helpers:

from crude_valuation import (
    load_production, find_analogs, fit_curve,
    cohort_summary, fit_quality, target_vs_cohort,
)

prod = load_production(WELL_APIS, run_sql)
analog_apis = find_analogs(ANALOG_FILTER, run_sql, exclude=WELL_APIS)
analog_curves = []
for a in analog_apis:
    try:
        analog_prod = load_production([a], run_sql)
        q = np.array(analog_prod["oil_bbl"])
        months = np.arange(len(q), dtype=float)
        meta = peek_well(a, run_sql)
        analog_curves.append(
            fit_curve(months, q, stream="oil", lateral_norm_ft=meta.lateral_ft, b_fixed=0.8)
        )
    except ValueError:
        continue   # skip thin / un-fittable analogs

print(cohort_summary(analog_curves))
cohort_median = percentile_curves(analog_curves, pct=0.5, norm_to_lateral_ft=10_000.0)

for api in WELL_APIS:
    meta = peek_well(api, run_sql)
    target_prod = load_production([api], run_sql)
    target_q = np.array(target_prod["oil_bbl"])
    target_months = np.arange(len(target_q), dtype=float)
    print(api, target_vs_cohort(cohort_median, target_months, target_q, meta.lateral_ft))

DECIDE

A free-form comment block. Document the call per well in plain English:

# === DECIDE ===
# Cohort: 15 analogs after dropping 3 bound-riders. b median 0.92, IQR 0.31.
# Target 42-329-12345 (Wolfcamp A, 9,400 ft):
#   36 months observed, qi_ratio 1.04 -> in line with cohort
#   Choice: pdp; fit own decline with b=0.9 (cohort median, dropped outliers)
# Target 42-329-12999 (PERMITTED, 10,200 ft):
#   no history; planned spud 2026-08
#   Choice: pure_analog with cohort curve at target lateral

EXECUTE

Build Forecasts inline per-well — no router, no module constants:

# === EXECUTE ===
forecasts_oil = []
forecasts_gas = []
for api, choice in CHOICES.items():
    if choice["strategy"] == "pdp":
        forecasts_oil.append(_build_pdp(api, b=choice["b"], stream="oil", ...))
    elif choice["strategy"] == "thin_blend":
        forecasts_oil.append(_build_thin(api, cohort_median_oil, ...))
    else:  # pure_analog
        forecasts_oil.append(_build_pure_analog(api, cohort_median_oil, ...))
    # same for gas

# Econ + persist
revenue = compute_gross_revenue(forecasts_oil, forecasts_gas, price_deck=...)
cashflow = compute_net_cashflow(revenue, interest_type=INTEREST_TYPE, ...)
npv_dict = npv(cashflow, discount_rates=[0.08, 0.10, 0.12, 0.15])

spec = build_minimal_briefing(headline=..., commentary=..., npv=npv_dict, ...)
persist(spec, persist_url="$persist_url")

Commentary must reference at least one concrete diagnostic you looked at.

Playbooks

crude_valuation/forecast/examples/ ships annotated playbooks showing the EXPLORE → DECIDE → EXECUTE pattern end-to-end. Today only forecast_mixed.py is in the new playbook form; the other two are still recipe-style.

Diagnostic helpers in detail

• cohort_summary(curves) — distribution stats per fitted parameter (qi, di, b, terminal, lateral). Plus notes flagging fits that rode the b bound.
• fit_quality(curve, months, q) — R^2, RMSE, bound-riding flags, max residual %, early/late signed-% bias windows for a single fit.
• target_vs_cohort(cohort, target_months, target_q, target_lateral_ft) — qi_ratio of target to cohort, peak detection, per-month signed-% residual vs cohort placed at target lateral.

None of these decide anything. The agent reads them and chooses.