# n85 — the contradictory analysis: flaws in our own case (2026-07-10)
# The program audits its auditor. Ten flaws, tiered by severity, each
# with its consequence. Three forced immediate corrections to the paper.

## TIER 1 — flaws in things we called "solid" (paper corrections forced)

F1. "UNIQUE" IS FALSE. The paper claimed the information metric is
"the unique vacuum-owned positive functional." No such uniqueness was
proven. A vacuum owns many positive functionals: energy variance,
entanglement-entropy Hessians, Renyi susceptibilities, the Kubo-Mori
metric (which differs from Bures). We proved ENERGY fails; we chose
Bures among the survivors. The constructive result stands ("a natural,
canonical choice with nothing tuned downstream") but the word "unique"
was an overclaim. CORRECTED in the paper same day.

F2. PRECISION MISMATCH IN THE GAMMA CHAIN. gamma = (1-2a)/(1+2a) is
quoted against Cassini at 1e-5, but its optical ingredient (substrate
response = 1/2) was verified by ONE wave-propagation test at the 10%
level with acknowledged edge systematics. Presenting |a| < 5.8e-6 as a
clean "reading" implied an exact chain. CORRECTED: the caveat now
travels with the formula.

F3. FRAMING TENSION, PAPER vs SITE. The paper's central no-go is
stated for "emergent gravitational redshift" in the conserved sector —
the same sector the site's final architecture identifies as
ELECTROMAGNETISM. Both statements are individually correct (the no-go
is precisely why the sector cannot be gravity), but a reader holding
both documents meets an unexplained tension. CORRECTED: one sentence
in the discussion now closes the loop explicitly.

## TIER 2 — structural weaknesses, recorded, not yet repairable

F4. THE GRAVITY SECTOR LIVES ON TWELVE SITES. [REPAIRED same day:
rerun at N = 12, 24, 48, 96 — the sign structure (active attracts,
passive repels) is N-independent, and the redshift-tracking residual
shrinks as ~1/N (2.32% -> 1.21% -> 0.51% -> 0.19%): the residual was
finite-size, and the linear-universal redshift claim STRENGTHENS to
'converging to exact'. The biggest Tier-2 debt is paid, with
interest.] Original finding: Every monitored-medium
result — the pairing, the redshift tracking (2.32%, itself possibly a
finite-size artifact never decomposed), the passivity verdict, the
Caves zero — was measured at N = 12, first moments, never scaled. The
EM sector got N = 200 and L = 96 sweeps; gravity got a ring you can
count on your fingers. Finite-size scaling of the entire gravity
sector: OWED.

F5. THE CAVES ZERO IS PARTLY CONVENTIONAL. "Active adds no noise" was
measured with the SAME aux noise matrix imposed on both branches by
construction. The argument that a T -> 0^- inverted reservoir carries
vacuum-level noise was ours, unverified microscopically. The no-floor
claim for the ACTIVE branch inherits this convention. Consequence: the
flagship's floor-free property is solid for the passive-monitored
sector, convention-dependent for the active one.

F6. THE PASSIVITY "THEOREM" IS A MODEL-FAMILY RESULT. The sign chain
runs through n33's specific coupling structure (the eta p-p term, the
quadrature conventions). "For every passive bath" was proven for every
bath OF THAT FORM. A different physical coupling could remap signs.
Demoted from theorem to result-within-the-family; the demotion is now
in the record.

F7. THE 18-ORDERS NO-FLOOR WAS A TRANSIENT MEASUREMENT. It measured
exponential decay of an initial condition, not a steady-state floor
with full noise. Combined with F5, the knee's "zero at rest" rests on
(i) the violation-only axiom (structural, strong) and (ii) noise
conventions (soft). The distinction now stated.

## TIER 3 — meta-flaws of the process itself

F8. PRE-REGISTRATION WITHOUT A FIREWALL. Gates and code were written
by the same agent in the same context, often minutes apart, with full
knowledge of the machinery's prior behavior. This discipline caught
real errors (documented repeatedly) but does NOT confer the
statistical protection of true pre-registration: there was no temporal
or informational separation. The record's honesty is real; its
epistemic weight is weaker than the word "pre-registered" suggests to
an outside reader. Stated here once, at full prominence.

F9. THE FINAL LAYER IS THE LEAST TESTED. Verdicts flipped roughly five
times in 48 hours, always by finding semantic errors. The surviving
architecture (EM reading + active vacuum) is the NEWEST layer — i.e.
statistically the least attacked — yet the site presents it with the
most confidence. The correct prior: the final layer carries the
highest per-claim risk in the whole edifice. Readers should weight
accordingly; the theory page now cannot say this loudly enough.

F10. THE ALPHA BALLPARK IS A LIABILITY. S depends on mu^2, N, and
lattice details; alpha^-1 "within one to two orders of 137" invites
numerology criticism at zero evidential gain. It remains flagged as
characterization-only; arguably it should be removed from the public
claims entirely — left standing, with this note attached, for the
reader to discount.

## The balance
Three overclaims corrected in the paper the day they were found; four
structural debts named (N=12 scaling, noise conventions, the passivity
scope, the transient-vs-floor distinction); three process confessions
(the firewall, the freshness inversion, the numerology bait). What
survives unchanged: the concavity lemma, the 1D dichotomy core, the
exact lattice numerics, and the checkability of everything. The
edifice is smaller again — and, again, that was the point.