# Royalty-at-Risk LATAM — Forensic Methodology v1.0

**Version**: v1.0  
**Date**: 2026-04-27  
**Owner**: AgroPatent Index — Forensic Models Team  
**Status**: Production · pending Big-4 methodology review (FTI / Deloitte FAS / KPMG Forensic / EY Fraud Investigation eligible)  
**Next public review**: 2026-Q3 (Bayer H1-2026 + Corteva Q1-2026 10-Q)

---

## 1. Purpose and scope

This document describes the methodology used to estimate **USD royalty pool, effective collected revenue, and royalty-at-risk** for biotech traits (INTACTA, RR2, Enlist E3, HB4, VT3P, Bollgard II, etc.) across LATAM jurisdictions (BR, AR, UY, PY, BO, MX, CL, CO).

It is **not an audit**. No external entity — Bayer, Corteva, Syngenta, BASF, Bioceres, ABRASEM, ASA, Kynetec, AgEvidence, ISAAA — produces an audited royalty-at-risk number for LATAM, because the universe to be measured (semilla guardada ilegal, saqueo MT/GO, enforcement gap) by definition is not reported in any public ledger.

What this methodology *does* deliver:

1. **Range-bounded estimate** (low / mid / high band, ±10% per parameter).
2. **Explicit citation per parameter** — every adoption %, royalty rate, area, yield, FOB, leakage factor traces to a public source.
3. **Backtest reconciliation** vs filings registered with SEC, BaFin, CVM, CNV, B3 — anchors any third party can independently verify.
4. **Versioned assumption file** (`data/royalty_assumptions.json`) — drift over time is auditable via git history.
5. **Honest disclosure of model limitations** — drift between model and anchor is published, not hidden.

This is the same standard applied by **Hindenburg Research, Muddy Waters, Bernstein Sector Reports, Spruce Point Capital, Wolfpack Research** for short-side / activist publications: methodology disclosed, range-bounded, public anchors, no fudging to match a target answer.

---

## 2. Mathematical model

### 2.1 Per-cell calculation

For each (jurisdiction × event × crop):

```
production_t              = area_planted_ha × yield_t_per_ha
royalty_pool_usd_per_cell = production_t × adoption_pct × royalty_rate_usd_per_t
effective_collected_usd   = royalty_pool_usd_per_cell × leakage_factor_collected
atrisk_usd                = royalty_pool_usd_per_cell − effective_collected_usd
```

Where:
- `area_planted_ha` is the country-level area for that crop (safra 2024/25)
- `yield_t_per_ha` is country-average yield (safra 2024/25)
- `adoption_pct` is the fraction of total crop area planted with that specific event
- `royalty_rate_usd_per_t` is the **post-leakage tech fee** disclosed in semilleras' annual reports (calibrated as net-of-leakage)
- `leakage_factor_collected` ∈ [0, 1] is the fraction of theoretical pool that effectively reaches the rights holder

### 2.2 Band construction (low / mid / high — 90% interval)

Three bands are computed by applying ε-perturbations to three orthogonal multipliers:

| Parameter             | ε default | Justification                                                          |
|-----------------------|-----------|------------------------------------------------------------------------|
| `adoption_pct`        | ±10%      | ISAAA Brief year-on-year variance for major LATAM events 2018–2024     |
| `royalty_rate`        | ±10%      | Tech fee dispersion across contracts (multiplier × territory × renewal) |
| `leakage_factor`      | ±10%      | Historical enforcement collection variance (Bayer + Syngenta data 2018–2024) |

**Low band**: all three multiplied by `(1 − ε)`.  
**Mid band**: central values from `royalty_assumptions.json`.  
**High band**: all three multiplied by `(1 + ε)`.

This produces a band approximating the 90% confidence interval under the assumption of independent, log-normal noise on each parameter. **It is NOT a statistical confidence interval** — it is a *deterministic sensitivity band*. We refer to it as "banda 90%" as a calibration heuristic, not a probabilistic guarantee.

### 2.3 Aggregation

After per-cell computation, totals are aggregated by:
- **By country**: pool / effective / at-risk per jurisdiction (BR, AR, UY, PY, BO, MX, CL, CO)
- **By owner**: pool / effective / at-risk per technology owner (BAYER_GROUP, CORTEVA_GROUP, SYNGENTA_GROUP, BASF_GROUP, BIOCERES_GROUP, etc.) — using an event→owner canonical map
- **By owner × country**: matrix used for backtest reconciliation
- **By crop**: soja / maíz / algodón / trigo
- **By event**: INTACTA_RR2_PRO, ROUNDUP_READY_2, ENLIST_E3, HB4_SOJA, etc.

---

## 3. Data sources

### 3.1 Public sources (citable per parameter)

| Parameter type      | Primary source                                                              | Where to verify                                         |
|---------------------|------------------------------------------------------------------------------|----------------------------------------------------------|
| Area planted        | CONAB (BR), Bolsa de Cereales BA (AR), USDA FAS PSD                          | Monthly CONAB releases · weekly Bolsa BA · USDA quarterly |
| Yield t/ha          | CONAB, USDA FAS, Bolsa BA                                                    | Same as above                                            |
| FOB price USD/t     | CBOT, Matba-Rofex, B3 (BR), spot quotes                                      | CME / Matba public quotes · refreshed quarterly          |
| Adoption % per event| ISAAA Brief 55 (2024) — *Global Status of Commercialized Biotech/GM Crops*    | https://www.isaaa.org/                                   |
| Royalty rate USD/t  | Bayer Crop Science Investor Relations · Corteva 10-K Note 6 Revenue · Syngenta Group annual report | SEC EDGAR · BaFin · Investor Day decks                   |
| Leakage factor      | Bisang & Anlló 2014 (CEPAL/INTA) · Trigo 2018 (ArgenBio) · ABRASEM saqueo studies · Kynetec range-band public commentary | Academic literature · industry association reports      |
| HB4 specifics       | Bioceres 20-F (SEC) FY2025 · pre-bankruptcy filings                          | SEC EDGAR CIK 1769484                                    |

### 3.2 Sources we deliberately exclude (and why)

| Source                              | Why excluded                                                                  |
|-------------------------------------|-------------------------------------------------------------------------------|
| Kynetec FarmTrak panels             | Private (USD 200k/yr) · NDA prohibits redistribution · using would create circularity if buyer also subscribes |
| Phillips McDougall AgChem reports   | Proprietary · can't cite in defamation-defensive publication                 |
| Equity research notes               | Tertiary / opinion-laden · not primary filings                                |
| Internal Bayer/Corteva ledgers      | Not public · would require leak / illegal disclosure                          |
| AFIP SISA / SIO Granos productor-by-productor declarations | Confidential fiscal information · only AFIP can disclose nominally |

**This is intentional.** A defamation-defensive forensic publication cites only filings that any third party can independently retrieve and verify. Equity research and private panels create asymmetric-information risk in an SEC enforcement defense.

---

## 4. Backtest protocol

### 4.1 Anchor selection criteria

An anchor qualifies for inclusion in `data/public_backtest_anchors.json` if and only if:

1. The figure appears in a filing **registered with a securities regulator** (SEC 10-K/20-F/8-K, BaFin Annual Report, B3 ITR, CVM IPE, CNV AR, BMV MX) **OR** in a press release on a regulated stock exchange disclosure channel.
2. The **page reference** and **verbatim excerpt** are recorded in the anchor entry.
3. The **scope mapping** to the model is unambiguous (which owner, which countries, which fiscal year, which segment).
4. A **reconciliation target** specifies an `expected_match_low` and `expected_match_high` band, with a written **rationale** for why the model output should fall in that band.

### 4.2 Reconciliation verdict

For each anchor:

```
if model_mid is within [expected_match_low, expected_match_high]:
    verdict = "reconciled"
elif drift_pct ≤ threshold_pct:
    verdict = "reconciled"   # within tolerance
else:
    verdict = "drift"
```

Where `drift_pct` is the % distance from the nearest edge of the expected band, and `threshold_pct` is anchor-specific (typically 15–20%).

### 4.3 Drift handling (anti-p-hacking discipline)

When an anchor drifts, **we do not adjust assumptions to make it reconcile**. That would be p-hacking and would invalidate the entire methodology. Instead:

1. Document the drift in `forensic_model.json → backtest.results[].verdict = "drift"`.
2. Hypothesize the cause in writing (anchor includes seed price premium not modeled · cross-licensing not captured · currency translation gap · safra year ≠ fiscal year).
3. If the cause is structural (e.g. our model excludes seed premium intentionally), document it as a known limitation in §6 of this methodology and leave the drift visible.
4. Only update assumptions when **independent new evidence** (a new anchor, a new academic paper, a new regulator filing) justifies the revision — and log the change in §8 review log.

### 4.4 Current backtest status (as of 2026-04-27)

3 anchors evaluated:

| Anchor                              | Model mid (USD)  | Expected band   | Verdict    | Drift % |
|-------------------------------------|-------------------|-----------------|------------|---------|
| BAYER_2024_CROPSCIENCE_LATAM        | $1.31B            | $1.50B – $2.10B | reconciled | 12.7%   |
| CORTEVA_2024_SEED_LATAM             | $0.23B            | $0.30B – $0.55B | drift      | 23.3%   |
| BIOCERES_FY2025_HB4_REVENUE         | $9.08M            | $18M – $32M     | drift      | 49.6%   |

**Drift hypotheses (not yet adjusted)**:
- **Corteva drift**: Model captures Corteva via ENLIST_E3 / CONKESTA_E3 / HERCULEX_I but excludes Pioneer brand seed premium and inbred line licensing income, both of which are inside Corteva's "Seed segment LATAM" disclosure. The drift is consistent with seed price markup (~25–30% of seed segment revenue) not being a royalty per se.
- **Bioceres drift**: The 20-F figure ($24.5M) is a global HB4 revenue line that includes RAGT (FR), Florimond Desprez (FR/EU), and Limagrain (FR) cross-licensing inflows. Our model only captures BR + AR cultivation — not licensing fees from EU/Asia partners. The drift is consistent with the geographic scope mismatch.

Neither drift will be "fixed" by parameter tuning. The right resolution is either (a) document permanently as scope difference, or (b) extend model scope (e.g., add cross-licensing income line) — which is a v2 decision pending Big-4 review.

---

## 5. Versioning and review

### 5.1 Model version

Format: `vMAJOR.MINOR — description`

- **MAJOR** bumps when the mathematical model changes (new term added, formula altered).
- **MINOR** bumps when assumptions are recalibrated against new data (new ISAAA Brief, new fiscal year anchors).

Every published version is committed to git with a tagged release. Historical `forensic_model.json` outputs are preserved in `data/snapshots/forensic/` (planned v1.1).

### 5.2 Review cadence

- **Quarterly**: ingest new public filings (Q earnings calls, 10-Qs, semestral B3 reports). Update anchors. Re-run backtest. If drift on any anchor exceeds threshold by >5pp from prior quarter, escalate.
- **Annually**: full assumption review. New ISAAA Brief, CONAB/Bolsa BA annual cierre, USDA FAS year-end PSD. Refresh `royalty_assumptions.json` with version bump.
- **Ad-hoc**: regulatory shock (Bioceres bankruptcy, Bayer divestiture, M&A) triggers immediate review.

### 5.3 External review eligibility

This methodology document and the underlying JSON assumption files are designed to be reviewable by:

- **Big-4 forensic & dispute advisory practices**: Deloitte FAS, KPMG Forensic, EY Fraud Investigation, PwC Forensic
- **Independent forensic consultancies**: FTI Consulting, AlixPartners, Kroll, BRG, Berkeley Research Group
- **Academic econometricians**: agricultural economics departments at INTA, Embrapa, IICA, Cornell, UC-Davis

A **methodology review** (not audit) by one of the above firms typically costs USD 30–80k and takes 4–8 weeks. The deliverable is a sign-off letter stating the methodology is internally consistent, the data sources are appropriate, the calculations match the documented formulas, and the limitations are disclosed.

**This is NOT the same as audited financial statements.** No firm will sign off on the *numbers* themselves because the underlying transactions are not auditable. They will sign off on the *process* — which is what serious institutional buyers need to defend the figure internally.

---

## 6. Known limitations

1. **Seed price premium is excluded.** Model captures only royalty per ton harvested. Seed sale margin (typically 25–40% of seed segment revenue for Bayer/Corteva) is NOT in pool — buyers must add separately if they want top-line "Seeds & Traits LATAM" comparison.

2. **Cross-licensing income is excluded.** Bayer ↔ Corteva ↔ Syngenta cross-license INTACTA, ENLIST, dicamba traits across territories. Model attributes 100% to single owner; cross-license splits would require contract-level disclosure (private).

3. **Stack double-counting.** A field with INTACTA_RR2_PRO will trigger BOTH INTACTA_RR2_PRO adoption AND ROUNDUP_READY_2 adoption. Model intentionally counts both because royalty contracts typically charge per stack component. If buyer wants single-event attribution, divide by stack-multiplier (typical 1.3–1.6 for soja BR).

4. **Currency translation.** All USD figures use FX rates implied at filing dates (EUR/USD = 1.073 for Bayer 2024, BRL/USD = 5.385 for ABRASEM 2024). Intra-year FX swings ±15% AR and ±10% BR not modeled.

5. **HB4 transition.** Bioceres bankruptcy March 2026 → Natal Seeds successor. HB4 royalty collection during transition is highly uncertain. Model uses pre-bankruptcy rates; flag this in any HB4-derivative figure.

6. **Mexico cultivation restriction.** OGM maize cultivation banned by 2023 decree; model excludes MX maize from royalty pool. Soja MX is import-only grain; no downstream royalty.

7. **No probabilistic uncertainty quantification.** Bands are deterministic ε-perturbations, not Monte Carlo. v2 may add MC sampling if a buyer requests probabilistic confidence intervals.

8. **No event-level tax / royalty regime nuance.** Argentina retenciones 33% soja affect cobranza efectiva; not separately modeled — folded into AR leakage factor 0.55.

---

## 7. How a buyer should cite this number

**Recommended citation format**:

> Royalty-at-risk LATAM 2024/25 cycle: USD 0.93B (mid) within USD 0.75B–1.13B band (low–high). Source: AgroPatent Index Forensic Model v1.0 (2026-04-27). Methodology: deterministic ±10% sensitivity over adoption %, royalty rate, leakage factor. Assumptions: ISAAA Brief 55 + CONAB BR + Bolsa Cereales BA + USDA FAS + semilleras' SEC/BaFin filings. Backtest: 1 of 3 anchors reconciled within tolerance; 2 anchors drift documented in §4.3.

**NOT acceptable citations**:

- ❌ "Audited royalty-at-risk: USD 0.93B" — we do not claim audit.
- ❌ "Royalty leakage in AR is exactly 45%" — banded estimate, not point.
- ❌ "Bayer is losing USD X to Uso Propio" — we estimate at-risk pool, not a counterparty's specific loss; that requires their internal ledger.

---

## 8. Review log

| Date       | Reviewer      | Change                                                                 |
|------------|---------------|------------------------------------------------------------------------|
| 2026-04-27 | Model owner   | Initial release v1.0. 3 anchors. ε = 10% across the board. 1 reconciled / 2 drift documented. |
| 2026-Q3 *(planned)* | Model owner   | Refresh anchors with Bayer H1-2026 + Corteva Q1-2026 10-Q. |
| TBD        | External Big-4 | Methodology review (sign-off letter or finding letter).               |

---

## 9. Contact

For methodology questions, anchor additions, parameter challenges, or external review coordination:

**model owner**: AgroPatent Index — Forensic Models Team  
**publication channel**: this repository (`docs/forensic_methodology_v1.md`)  
**JSON sources**: `data/royalty_assumptions.json`, `data/public_backtest_anchors.json`, `data/forensic_model.json`  
**rebuild command**: `python scrapers/build_forensic_model.py`

---

*End of methodology v1.0.*