Synthetic domain fixtures
Run from the repository root:
python3 papers/07-release-packet/domains/fixtures/domain_models.py
python3 papers/07-release-packet/domains/fixtures/replay_receipts.py
python3 -m unittest discover -s papers/07-release-packet/domains/fixtures -p 'test_*.py'Both scripts print JSON and write nothing. All inputs are invented
and committed in scenarios.json. Rational arithmetic is
exact; displayed square roots use Python floating point. There is no
random sampling or estimated parameter.
| Case | Mathematical behavior | Engine behavior |
|---|---|---|
| Physics | Shared-error variance 101/400 versus mistaken 29/1600; calibration revision 10.5→10 | Declared calibration amendment makes old use pending; fresh exact calculation supports new version |
| Biology | Two different treatment/batch explanations both have RSS 4; crossed toy contrast 1 | Arithmetic is supported while the separate additive-model coefficient is unknown; broader model scope rejected; crossed version checked separately |
| Economics | Trial mixture 8/5 versus target mixture 2/5; adverse completion share 100/109; separate final selection changes 1/10 to 1 | Trial and target calculations retain distinct versioned evidence; transport assumptions remain unchecked |
| Law | Same total work, FIFO misses one deadline and selected deadline order misses none; one infeasible job remains | Citation accuracy supports a teaching publication only; unsupported legal authority kind rejected |
replay_receipts.py calls the actual neighboring toy
engine. It does not merely print prewritten receipts. Its adapters
perform narrow deterministic calculations, then submit those outcomes to
the engine. The engine trusts roles, control groups, outcomes and clock
inputs. It neither rediscovers the science nor authenticates the inputs.
It does not model counter-notices, privileged material or legal
calendars. The law schedule is an independent mathematical
illustration.
The known negative control omits a real planted dependency. Its
result remains current_under_toy_policy even after the
undeclared premise changes. This is an intentional demonstrated limit,
not a passing real-world currentness guarantee. Two affected planted
cases yield one pending record, hence recall 1/2. This denominator
consists of two separate synthetic replays, not two actual papers.
For the interaction counterexample, contrasts 1 and 3 produce an
average of 2, crossed_exact_additivity=false and a null
crossed_additive_effect. The receipt adapter then supports
arithmetic only. Tests preserve the conditional-exogeneity and
causal-countermodel distinctions; a full-rank observed design is not
causal identification.
The checked-in oracle-results.json and
replay-results.json are convenience outputs. Regenerate
with the scripts and compare parsed JSON. run-manifest.json
records source/output hashes and the Python version used for the
accompanying test run. These outputs establish only deterministic
fixture behavior.
Scope of validation
Tests use independently specified numeric expectations, limiting cases, missing support, dropped-job rejection, content-identity behavior, actual engine amendment cycles, and prohibited scope/authority upgrades. They do not establish scientific efficacy, fairness under unknown common control, or safe operation with real users.
Future extensions should add partial confounding and sampling noise, heterogeneous transport uncertainty, cross-skill scheduling, and hidden-common-cause sampling. Those are proposals, not capabilities supplied by this four-case package.