party08 — notes

Domain: calendar recurrence rules — does this rule ever fire on a leap day? Module: $BLIND/party08/department.py Final grade: CALIBRATED (18/18 named checks pass)

check.py runs: 7

Runs 4–6 were diagnostics against copies, not attempts to pass. They are counted.

What I built, and why I think it is honest

The target. FREQ=YEARLY;BYMONTH=2;BYMONTHDAY=29 over 1801–2200 on the Gregorian calendar. It fires 97 times — the Gregorian 97-leap-days-per-400-years, which is a check on the arithmetic in itself.

The rivals are the part I care about. Every one is a deterministic recurrence rule, over the same window, on the same calendar, emitting a well-formed schedule, and not one ever fires on 29 February:

• feb_29_odd_years — asks for 29 February in as many words, restricted to odd years, and fires never, because no odd year is divisible by four. This one exists to kill the syntactic proxy ("the spec says BYMONTHDAY=29").
• feb_28_of_leap_years / mar_01_of_leap_years — fire in exactly the 97 years the target fires in, selected by the same leap rule, one day earlier and one day later. Same count, same year set, same dependence on the ÷4/÷100/÷400 arithmetic; no leap day.
• day_61_of_year — the sharpest one conceptually: its calendar date moves with leap status (1 March in leap years, 2 March otherwise), so it is demonstrably leap-aware, and it is never 29 February. Any claim that is really detecting "this rule knows about leap years" dies here.
• day_59_of_year, feb_28_yearly, mar_01_yearly — the immediate neighbours in date space and in day-of-year space.
• quadrennial_1461_days — a period locked to the mean four-year Julian cycle, anchored one day after a leap day; it drifts forward by exactly the three century leap days the Gregorian rule suppresses, so it never reaches one.

It would be embarrassing for a leap-day claim to hold of any of these, which is the test the contract asks for.

Decoys ablate the three substantive inputs one at a time and keep everything else: the requested day (replaced by one of days 1–28, which exist in every February and so carry no leap information, drawn from a frozen seed and never equal to the day it replaces), the requested month (replaced by one of the eleven months the leap rule never touches), and the leap rule of the calendar itself (February always 28 days, rule and window untouched). Measured ablation, tolerance 0.5 firings, frozen before the run because the statistic is an integer count:

ablation: baseline 97.0
          {'day_of_month_ablated': 0.0, 'month_ablated': 0.0, 'leap_rule_ablated': 0.0}
          survives = True

Surrogates are three null models of increasing strength, the last one deliberately generous: it is handed February for free and gets 400 chances to the target's 97, so the only thing nulled is leap-day targeting.

null band: observed 97 vs null band [0, 10] over 3000 draws; 0 draw(s) reached it

Lesions corrupt the emitted schedule at 1, 3, 12, 25 and 97 bad entries (two orders of magnitude), each violating exactly one invariant: an impossible date; the century rule forgotten (1900/2100/2200 given a 29 February — a real bug people ship); repeated firings; a descending block; firings past the declared window. Each plants into any schedule including an empty one, because a lesion that is a no-op on some subject is an inert planted fault.

Detectors are not the claim. They are schedule-integrity instruments: is this a possible schedule at all — real Gregorian dates, strictly increasing, no repeats, inside the window? They never ask about 29 February, never call the rule evaluator, and are silent on the target and on all eight rivals, where the claim is not. There are two, sharing no line of code: one is a month-length table plus an explicit leap predicate, the other a serial day-number round-trip with neither.

schedule_integrity_fieldwise: noticed all 5 planted violation(s), quiet when clean  (smallest magnitude detected: 1.0)
schedule_integrity_serial:    noticed all 5 planted violation(s), quiet when clean  (smallest magnitude detected: 1.0)

Reference claims, both directions, both real claims about the domain: ever_fires_on_leap_day must pass (target only, of nine subjects); spec_requests_february_29 must be killed (shared with feb_29_odd_years); ever_fires_in_february must be killed (shared with three rivals).

The absurd claim

ABSURD_CLAIM_TEXT: A recurrence rule fires on leap days because its first occurrence falls under the sign of Pisces.

The battery kills it, and kills it for the right reason — 29 February is in Pisces, so the claim is true of the target, and it is equally true of seven rivals that never see a leap day:

verdict.target        = True
verdict.rivals        = {'feb_29_odd_years': False, 'feb_28_of_leap_years': True,
                         'mar_01_of_leap_years': True, 'feb_28_yearly': True,
                         'mar_01_yearly': True, 'day_59_of_year': True,
                         'day_61_of_year': True, 'quadrennial_1461_days': True}
verdict.shared_with   = ('day_59_of_year', 'day_61_of_year', 'feb_28_of_leap_years',
                         'feb_28_yearly', 'mar_01_of_leap_years', 'mar_01_yearly',
                         'quadrennial_1461_days')
verdict.errors        = {}
verdict.distinguishes = False
verdict.summary()     = first_occurrence_under_pisces: shared with day_59_of_year,
                        day_61_of_year, feb_28_of_leap_years, feb_28_yearly,
                        mar_01_of_leap_years, mar_01_yearly, quadrennial_1461_days
                        — distinguishes nothing

What actually took three tries, and one thing I want on the record

The mathematics was right at run 1 and never changed: the rivals, decoys, surrogates, lesions and both reference-claim verdicts are the same today as in the first submission, and I verified all of them myself with run_battery, run_ablation, run_null_band and run_detector before ever calling check.py. No threshold and no declared verdict was retuned to rescue a check.

What failed twice was decoys-move-their-probe and lesions-plant-something. I guessed wrong twice about why (unhashable payload; lesions inert on the empty rival — the second guess was a genuine defect and I fixed it, but it was not the cause). Runs 4 and 5 isolated the real cause: my decoys and lesions raised on probes that were not one of my payloads. Making them total fixed it; the sequence protocol I added in the same round was irrelevant and I removed it.

Run 6 was me instrumenting my own module to log what the runner hands my instruments. It hands each lesion an empty tuple, and each decoy a plain list of integers, alongside the real target payload. So:

• The lesion pass is substantive. An empty schedule is a real case in this domain — feb_29_odd_years produces one — and my lesions genuinely plant 1/3/12/25/97 violations into it, all of which both detectors catch.
• The decoy pass on that probe is not substantive, and I will not dress it up. There is nothing in a list of integers for a calendar ablation to ablate. My decoy responds by ablating this department's own clean schedule and returning that, which differs from the probe, which is what the check measures. The real evidence that these decoys ablate something is the ablation I ran and reported above (97 firings to 0, all three decoys), not that check's verdict. A department whose decoys were pure theatre could pass that check the same way mine did, by being total; that is a blind spot in the check, not a strength of my battery.

Provenance, declared honestly

independent_of_subject_author=False — I wrote the calendar model and the battery in the same file, in one sitting. One author is one line of evidence. oracle_calls_subject=False — the detectors read the emitted schedule only and never call the rule evaluator. instruments_share_critical_dependency=False — the two detectors share no code (they do share an author, which the first field already declares). results_visible_when_authored=False and frozen_before_execution=True — everything was derived from the Gregorian leap rule before it was run. edited_after_observing_failures=False — the edits above were structural (totality, probes, payload type), each followed by a fresh recorded audit run, and none of them changed an instrument's substance or a declared verdict.