papers/hanzo-temporal-assertion-graph/sections/09-limits.tex

\section{Limitations}
\label{sec:limits}

\paragraph{The measurement covers one kind of query.}
CronQuestions exercises reads of one pair at a valid instant and reads of a
pair's history. It does not exercise transaction-time reads at a past $\tk$,
corrections, retractions, one-valued pairs, declared schema, walks, paths,
diffs, communities or answers; those are covered by the package's unit tests and
by nothing measured here. Every lane question is asked with $\tk$ after the load,
so Proposition~\ref{prop:stable} is proved and tested, not measured.

\paragraph{The lane bypasses the HTTP layer.}
The driver calls the admission, record, read and resolution functions in-process.
Latency therefore excludes authentication, tenancy lookup, serialization and the
network, and the load excludes the schema check. Figures for the service as
deployed would be higher by those terms, which were not measured.

\paragraph{A single shared machine.}
Every timing comes from one laptop running other work, at one-minute load
averages between about 11 and 42 on ten cores. The comparison with semantica
survives that, because the ratio follows from the work each call performs; any
finer comparison, including between the store's two clocks and between the first
and the current run, does not.

\paragraph{Time resolution and a single connection.}
Instants are whole seconds, so Proposition~\ref{prop:stable} holds for a
transaction instant in an earlier second than the write, and two versions
written in the same second count as learned together. Each organization's file
is served through one connection, which is why walks, reads, diffs, partitions
and erasures are bounded; the bounds make some answers partial, and those
answers say so.

\paragraph{The order ranks no source.}
Because every source has the same rank, a later account wins within its moment
whoever filed it. An organization that trusts one publisher over another has no
way to say so to the store today. The shared algebra already takes a rank, so
adding one is a vocabulary for declaring it rather than a change to resolution
\proposed{}.

\paragraph{The schema is shallow.}
There is one type per entity, equality rather than subtyping, and no constraint
beyond domain, range and cardinality. Declarations bind at the knowable instant
of the checked row and are read at a valid instant after every statement begins,
so the schema has a transaction history but no valid-time history of its own.

\paragraph{Retention is incomplete.}
No operation sets the litigation hold, no operation disposes of rows under an
organization's retention policy, and erasure reaches this store alone
(\S\ref{sec:erasure}). Until those exist, the append-only table only grows, and
an erasure receipt is never complete.

\paragraph{Answers are as good as the model and the citations permit.}
The answer operation checks that citations point at rows it gave the model, not that
the rows support the text. Its output depends on the model the platform routes
to, and it has not been evaluated on any benchmark.

\paragraph{The second lane measures something else.}
LongMemEval-S measures a retrieval engine that does not use this store, on a
release its authors have deprecated, and measures retrieval rather than answers.
Its figures say nothing about the assertion store and are reported with that
scope.