Methodology
PhaseFolio publishes the valuation framework behind each signed export: rNPV assumptions, probability-of-success calibration, Monte Carlo simulation, IRA terminal-value treatment, development-path semantics, backtest validation, source references, and permanent version history.
- 01
Backtest Methodology
AUC, Wilson confidence intervals, calibration, and per-indication cohorts (RA, NSCLC, antimicrobial).
Validation of probability-of-success predictions against historical outcomes. Discrimination (AUC) versus absolute calibration, Wilson score intervals on accuracy, selection-bias caveats on calibration plots, and the limits of each cohort.
Read section →
- 02
Probability of Success Calibration
Sources behind the indication × modality × biomarker matrix and the multipliers.
The three-dimensional benchmark matrix derived from BIO/QLS 2021 (N = 12,728+ stage transitions), modality-specific Citeline 2024 adjustments, and the seven evidence-based multipliers applied through the log-odds path. Includes the overfitting-governance gate: a multiplier scores the engine only if a held-out cohort with both approvals and failures can validate it, otherwise it is demoted to a non-scored risk flag.
Read section →
- 03
IRA Terminal-Value Framework
Year 9 / Year 13 cliff modeling under the Inflation Reduction Act.
How the engine accounts for Medicare Drug Price Negotiation. The Year 9 small-molecule cliff, the Year 13 biologic cliff, MFP discount mechanics, and what the closed-form sensitivity slider is fitting.
Read section →
- 04
Development Path Templates
Path tiers, automatic-effect policy, and the no hidden multiplier rule.
How development-path labels are treated inside indication plans: baseline, structural-template, analyst-confirmed, and display-only tiers. The section makes explicit that path labels do not change PoS, revenue, cost, duration, or launch timing by default; only visible, cited scenario assumptions can alter the math.
Read section →
- 05
Evidence Standards
What counts as a citation, freshness rules, and the audit chain.
The Evidence Register schema, accepted source tiers, citation freshness decay, and how the substrate links each numeric assumption to the source that supports it.
Read section →
- 06
Network Benchmarks
Q4 2026 anonymized data publication — methodology preview.
A public commitment device. The first annual anonymized benchmark dump ships Q4 2026; the methodology, anonymization approach, cohort cuts, and citation format are visible now.
Read section →
- 07
JHTV Public-Data Portfolio Scoring
Four-dimensional rubric used in jhtv-portfolio@2026-Q2 case study.
Scoring method for 602 publicly-listed JHU Tech Ventures Therapeutic Modalities inventions: clinical relevance (LLM-assessed vs SoC), modality fit (cumulative LoA), competitive whitespace (CT.gov + FDA counts), and IRA exposure proxy. Locked weights 30/25/25/20.
Read section →
- 08
JHTV Deep-Dive Production
How the top-10 asset deep-dives were produced — model, prompts, audit pass.
Two-tier scope: engine outputs (rubric, rNPV, Monte Carlo, dossier) versus AI-augmented analysis (mechanism plausibility, comparator status, asset thesis). Generated by per-asset Claude Opus 4.7 agents at maximum thinking effort with the 1M-token context window. No MD review; every finding source-anchored. Hyper-critical CMO-lens audit pass surfaced material findings on 8 of 10 top-10 assets.
Read section →
- 09
Modality COGS Benchmark
Constructed cost÷price ratios for 7 modalities — the prior behind the COGS% wizard default.
Per-modality COGS% as a constructed, disclosed ratio: steady-state fully-burdened manufacturing cost divided by a stated reference net price. Covers Small Molecule, Peptide, Monoclonal Antibody, Bispecific, ADC, Cell Therapy, and Gene Therapy. Includes the price-dominance caveat (COGS% is set by price far more than manufacturing cost), the gene-therapy bimodal band, confidence flags for lower-sourced entries (bispecific, peptide), and an explicit "prior not estimate" disclosure. Added at methodology@2026-05-20.
Read section →
- 10
Commercial-Model SG&A Benchmark
SG&A% by commercial model (orphan / specialty / primary care) — the prior behind the SG&A% wizard default.
Per-commercial-model SG&A% as a share of net revenue, keyed on the go-to-market model rather than the modality (salesforce size, prescriber breadth, DTC, patient-services intensity drive the ratio, not the molecule). Three tiers: orphan/rare ~32% (the counterintuitive HIGH tier — small denominator plus dedicated rare-disease infrastructure), specialty ~25%, and primary care ~25% (specialty and primary care share a central rather than fabricating a gap). Includes the model-not-modality caveat, the steady-state-vs-launch-year disclosure, and an explicit "prior not estimate" framing. Added at methodology@2026-05-21.
Read section →
- 11
Tax Rate Benchmark
US federal statutory 21% as a transparent, conservative default — the prior behind the tax-rate wizard default.
The default corporate tax rate is the US federal statutory 21% (26 U.S.C. §11(b)) — a transparent, jurisdiction-neutral, conservative anchor that is deliberately NOT keyed on modality or archetype. Explains effective-rate reality (real pharma runs ~0–16% via IP domicile, GILTI, FDII, and R&D credits), the §172 80% NOL cap (so "clinical-stage means 0% tax" is wrong), and how to adjust the rate down for an acquirer with offshore IP. Added at methodology@2026-05-21.
Read section →
- 12
SEC EDGAR Deal-Term Extraction
Real comparator deal terms from SEC filings — the substrate behind engine 2.3 auto-populate.
Extraction of upfront, milestones, royalty %, and equity from 8-K Item 1.01 filings (and 6-K for foreign private issuers) for a curated biotech sponsor cohort. Includes the engine 2.3 cascading-match auto-populate (indication × deal_type → indication-only, N≥3 gate) that fills empty deal structures from aggregate-eligible SEC comparator medians while analyst-entered deal terms remain authoritative. Aggregate medians exclude unreviewed needs_review rows per counsel; customer-facing surface is the auto-populate disclosure (10-K Exhibit 10.x and 6-K-derived FPI cohort claims remain separate scopes). Added at methodology@2026-Q3.
Read section →
Use PhaseFolio From Your Agent
PhaseFolio exposes a Model Context Protocol server. Any MCP-compatible client — Claude, Codex, Cursor, your own agent — can call the same engine, methodology, and signed dossier exports your human analyst uses. Three public tools (query_benchmarks, verify_export, get_methodology) require no auth.
Claude Code
HTTP-native, one CLI command
claude mcp add --scope user --transport http phasefolio https://app.phasefolio.com/api/mcpCodex CLI
OpenAI Codex; HTTP-native
codex mcp add phasefolio --http https://app.phasefolio.com/api/mcpClaude Desktop / Cursor
Stdio bridge via the npm wrapper
npx -y @phasefolio/mcpAdd to claude_desktop_config.json (or your client’s MCP config) under mcpServers.phasefolio with command "npx" and args ["-y", "@phasefolio/mcp"].
Six additional org-scoped tools (get_project, get_scenario, get_evidence, get_dossier, query_landscape, list_scenarios) unlock with a bearer token. Contact us for access. Server discovery at /.well-known/mcp.json.
Citation provenance is part of the methodology
Each methodology subsection renders its own verified references, favoring DOI links for journal articles and official report pages for industry sources. The current methodology version reconciles the BIO/QLS 2021 probability-of-success anchor, peer-reviewed clinical-success literature, and the source links used by signed-export captions.
What this methodology is for
PhaseFolio is decision-support for portfolio triage, asset stress-testing, and term-sheet negotiation by institutional investors, research institutions, biotech BD teams, and licensing analysts. It is not calibrated for IFRS-13 / US-GAAP fair-value reporting, Section 409A valuations, fairness opinions, or litigation-grade damages calculations. For those uses, treat any output as a working model that a qualified valuation professional must independently challenge and recalibrate.
Material limitations. Several PoS modifiers (genetic validation 2.6×, biomarker, orphan, CAR-T, gene therapy) are reported in the source literature as relative success ratios but applied here through the odds-ratio (logit) path — a deliberately conservative approximation. Per-cell sample sizes and confidence intervals are not yet surfaced in the UI. Sensitivity analysis is one-at-a-time (tornado), not global. Validation evidence to date: three published cohorts across three sub-indications. See all back-tests →
Overfitting governance. Each PoS multiplier adds a degree of freedom, so a multiplier is allowed to score the engine only if a held-out cohort containing both approvals and failures can validate it; a signal that fires only on a cohort's failures is demoted to a non-scored risk flag. The antimicrobial backtest publishes the full ablation — the defensible 0.629 AUC from the one validatable multiplier, not the uncheckable 0.797 the unvalidated pair would have shown. See the governance gate →
An expanded version of this model card — including the per-modifier estimand table, full validation evidence, and update cadence — is available on request. Contact contact@phasefolio.com.
This document set describes the computational methodology used by PhaseFolio. It is provided for transparency and does not constitute financial or medical advice. All benchmark data derive from published, peer-reviewed sources cited in each section. Every methodology version is permanent and verifiable — see the full version history.