warn when the model does not start where the user asked

[jdeast]

When a user sets a value, the model should produce that value or say why it can't. The second half did not exist, and the failure was silent: examples/ob09020 pinned mulensevent.t_E = 76.9 and started at 74.48, and the only way to find that out was to compile the graph by hand.

Measured across the shipped examples: 19 build, 14 user values are not reproduced.

The check

ModelAuditor.check_user_starts() reads every user-set initval off the compiled graph — not p.value (a draw from the prior) and not p.initval (the ledger, which is half of what is under test). Both look right while the model starts somewhere else. inspect_start prints the misses after the startup table, beside the existing flat/unused warnings.

Two reasons, separated by the ledger rather than guessed:

ledger vs user	built vs user	reported as
same	differs	approximate — the derivation cannot preserve it
differs	—	overspecified — and _last_solved_by names the equation

Calling the first "overspecified" would send someone hunting for a conflicting pin that does not exist, so the distinction is load-bearing rather than cosmetic.

What the sweep found

All 14 misses are the first kind. Nothing in any shipped example is being trampled — every one is a value the engine faithfully kept and the derivation could not rebuild. That is direct evidence the precedence system is not the problem here, which is the opposite of where this investigation started.

The error tracks pi_rel exactly as the heliocentric/geocentric diagnosis predicts: ob09020 (pi_rel = 1.21 mas, a 747 pc lens) is ~10x the bulge events at pi_rel ~ 0.1. Two finds that were previously invisible: gj1214's transit.MIRILRS.jitter_variance (+49.5%) and galactic_model's star.BulgeTarget.mass, which starts at exactly 10**-0.3 because a mass pin never reaches logmass.

Two traps this walked into, both now regression-tested

Degrees are periodic. The prototype reported four examples as broken over bigomega/alpha pairs exactly 360 apart (352.57 vs -7.43, 210 vs -150, ...). Those are the same start. A warning block with false positives is one nobody reads.

The ledger is in INTERNAL units (config.py:890) while the params file is in user units, so comparing them raw calls every converted parameter overspecified — star.ra holds 4.6095 rad against a written 264.105 deg. The first draft did exactly that. test_the_ledger_is_compared_in_user_units fails on the pre-fix code; that was verified by mutation, not assumed.

examples/ob09020 back to published values

A params file is a template, and a reader cannot tell a transcribed number from a fitted one when every entry carries a paragraph defending it. The justifying prose is gone. pi_E_N returns to Skowron's -0.022 from a hand construction, q to the printed 0.273, and theta_E = 2.95 is now pinned.

The set does not close — marginal medians from one table plus physical parameters from another paper — and that is the normal case, not a fault. The new report says so instead of the file hiding it behind tuned numbers.

bigomega is the one entry that cannot be "the published value": Yee's Omega_node = -7.767 is in their frame, and the documented mapping consumes phi_pi, which is a model output. Re-converting against the delivered phi_pi (162.068, from 158.795) gives 340.755; the same rule reproduces the old 337.484 to 0.002 deg. Start logp +19,024 -> +36,397, with alpha_0 = 189.082 (published 189.08) and s_0 = 0.4264 (0.4294) unchanged — the published-only template starts better than the hand-tuned one did.

Measured, because it decides whether this is seedable at all: this fit cannot walk in bigomega. +/-10 deg costs >1e6 nats, 60 deg costs 7e5, and the midpoint is worse than either end — a dead mode, not a slow one, against PTDE's factor-200 flattening.

Verification

8 new unit tests. 85 passed across the acceptance, keplerian, polish, hot-chains and wrapup-guard files. The ob09020 acceptance fixture is regenerated for the deliberate seed change (81 terms); the other 12 stay bit-identical. Pre-push hook green (ruff, ruff format, full suite).

Follow-up

The heliocentric/geocentric frame fix removes the largest remaining class of misses. It is now an improvement rather than a prerequisite, since the contract is satisfied by the warning.

🤖 Generated with Claude Code