Release model card
Version 3.3 · 14 September 2026 · engine/API 0.7.0 guided MCP · specification v1.14. Owner: Marc Dumpff. Authorized for research release by Marc on 14 September 2026. Independent model approval has not been recorded.
This card describes the 0.7.0 source contract. The repository’s release record separates local checks, deployment verification and Claude conversation acceptance; authorization alone does not establish which build is running. Read the engine version, build/source revision and historical-data digest from the completed answer or engine health. Preserve the frontend calculation version and exact numerical request too. Older saved results keep their original definitions and versions.
Supported use
Test explicit theories about Agreement cash flows, Buyer behavior and hedges. Dollar reports USD outcomes; Coin reports BTC-equivalent investment flows. BTC cash-flow IRR is a useful annual timing measure and should be read with purchases, contributions, recoveries, total return and the stated holding reference. The model computes conditional outcomes; it does not forecast the market or execute trades.
The funding convention introduced in 0.6.0 and retained in 0.7.0 funds each Agreement at its own entry price: one BTC per par Agreement. It separates contractual BTC, gross deployment, receipt offsets, realized net cash and signed horizon value, and provides unhedged cohort reconciliation. Coin-specific IRR diagnostics remain separate from USD diagnostics. Purchase-normalized multiples use gross BTC funded at entry; net economic and realized recovery ratios are separately named. Holding BTC uses each strategy’s matching dated contributions, not one common starting budget. See Reading BTC performance.
Guided AI analysis
The MCP connection now guides the assistant through the objective, comparison, Agreement-flow capital basis, timing and assumptions. Before calculating, it prepares a brief and asks about missing essentials. Dollar and Coin starting cases use the same versioned presets as the browser. A fixed-wallet experiment remains unsupported. This changes the AI workflow, not the model mathematics or the browser’s starting assumptions. A prepared brief records what the caller supplied; it does not independently establish human approval.
MCP calculation calls require a signed preparation that expires after one hour. To replay an analysis, prepare its saved numerical inputs again. Connecting exposes guidance and an optional starting prompt; it does not install a skill. See the MCP guide.
Corrections retained from 0.6.1
Engine 0.6.1 calculates performing Coin sensitivity at every monthly payment age. It corrects the last eligible historical vintage and refuses an explicit Coin objective on Dollar-only API routes. Detailed simulation responses retain their memory reservation while the response is held or sent. Dollar Risk includes numerical loss frequency and severity, and the Dollar cash multiple explicitly uses gross Agreement purchases. The BTC funding convention introduced in 0.6.0 remains unchanged. Re-run older hedge comparisons to apply the corrected sizing.
These corrections do not complete the audit’s remaining calibration, pricing-domain, execution or production-capacity checks.
Starting assumptions
New workspace cases use illustrative GBM with continuous annual drift 0 and volatility 43%; the median price can fall even when expected price is constant. API/legacy defaults use an endpoint-pinned bridge. Saved scenarios retain their own construction. Neither setting is a calibrated forecast. The exact saved SimConfig, surface, costs, market seeds and fixed valuation seed define an experiment.
Limits that remain
Buyer behavior parameters are priors. Monthly paths and price history do not calibrate how Buyers would perform. Option surfaces and execution, basis and funding assumptions are supplied inputs. Sampled p5, tail loss and minimum outcomes are not guaranteed floors; valid-IRR counts and sampling uncertainty stay visible.
The model does not simulate a complete funded BTC wallet, conversion policy, daily margin, option/swap collateral, forced closure or equal-budget portfolio ranking. The USD cash screen and analytical monthly BTC conversion are separate conventions. Reporting in BTC does not assert BTC settlement. A purchase-yield attribution mark is not certified financial-reporting fair value or auditor acceptance.
Evidence and AI access
Completed evidence retains the objective, conventions, request, seeds, engine/data identity and calculation version. The verifier checks recognized package structure and integrity, not investment suitability. Automated tests check their stated accounting, numerical and lifecycle properties; they do not constitute independent model approval.
Local 0.7.0 checks passed 442 backend tests, including 27 MCP integration tests, plus the frontend, documentation and acceptance-harness checks. These verify the connection protocol and calculation safeguards. Actual Claude conversation acceptance remains pending manual testing; the repository also provides a seven-case conversation evaluation harness.
The AI tool capability table names supported objectives. A single Coin scenario returns the complete research observation; Coin Monte Carlo and hedge comparisons use the research tool. The existing purchase-price solver is USD-only and refuses a Coin objective. See MCP reference.
The repository’s MODEL_CARD.md contains the full release card, evidence paths and approval status. The prior v2.1 card is archived under docs/model-cards/; its dated example outputs are not current workspace results. Limits and assumptions remain part of every research interpretation.