# Specialist source, replay and collateral precision

Run the Python example or native `run_specialized_journey_review()`. `data` selects the cases; `accounts_path` and `releases_path` supply source CSVs; `governance_data` supplies the existing frozen specialist contracts; `model_path` supplies their generation artifact; `verify_reference` toggles reference checks and `return_artifacts` exposes full histories and replay inputs. Native `engine_paths` optionally names input_lineage, specialized_governance and _wholesale_journey_tools companion scripts. The default dataset uses the five existing CRE, project, aircraft, ship and equipment claims, without new rating models.

Actual loan principal and first-period operating price are read, joined by historical release, independently reconciled, and bound to the frozen contract. The original asset profile must match exactly. Existing governance selects its dated model, compiles/applies policy and executes servicing events. This example decomposes the real decision and runs the same frozen historical registry twice, preserving knowledge ordering, identical duplicate delivery and actual event IDs. Full states and actual model/policy/source snapshots are checked with replay_consistency; a changed probability must be BLOCKED. No separate PD calibrator is falsely invented: its snapshot explicitly states that this simulation recipe has none.

The convergence calculation concerns each contract's actual first collateral profile with an explicitly declared prospective 15% idiosyncratic volatility, independently of its original transaction stress nodes. The first 512 Gaussian draws from seed 4102026 are packaged unchanged as two independent 64-path replicates, four increments per path. Both languages consume the same serialized draws; native seed equivalence is not assumed. Actual asset simulation runs at 180- and 360-day steps to day720, with nested16/64-path prefixes. Coarse Brownian increments are paired sums divided by sqrt(2). An independent deterministic-ageing times lognormal formula verifies every raw terminal value before EUR-cent reporting. Expected unrounded terminal value equals deterministic ageing; rounding can move its expectation by at most0.005EUR. Reported means, sampling intervals and paired grid differences come from simulation_convergence. The exact-transition kernel has zero paired grid difference. A stringent0.5EUR MC half-width is a diagnostic request, not a guaranteed PASS.

This establishes the collateral valuation kernel's finite-grid arithmetic and sampling behavior. It does not establish continuous-time transaction-default bias. The original18 weighted transaction scenarios remain non-IID and retain their separate exact discrete contractual-ledger oracles. No forecast path average is called TTC. Default tests exercise one genuine collateral kernel; CREDITRATINGSYSTEM_EXTENDED_TESTS=true runs all five complete source/governance histories and kernels. The full installed example always runs all five.
