Provenance and limitations

This author was started as a new delegated agent with fork_turns=none, according
to the parent assignment, and was provided no prior conversational history.
The author received the current environment/system instructions, the parent
allowlist assignment, and one verbatim research specification from
../source-instruction.txt. ORIGINAL-INSTRUCTION.txt preserves those source bytes.
The assignment explicitly described forbidden semantic classes and boundaries;
it supplied no concrete former cases, generator structures, implementation,
taxonomy, old labels, results or concrete interface. Those boundary descriptions
are control instructions, not exemplars used to construct this corpus.

Actual direct semantic file input: solely ../source-instruction.txt.
Additional file access: the byte-exact archival copy and files created by this
author below author-artifacts; specific access events are in ACCESS-LEDGER.json.
No old chat, epistemic pilot, previous 48 cases, former development case,
generator, diagnose.py, reference.py, evaluator, ground truth, raw result, audit
result, former E006 draft, Colony/worker/scheduler implementation, example or
external dataset was inspected. No workspace discovery, listing or searching
outside this authored allowlist was performed. No web, networking, connector,
thread tool, SEARCH, computer use, diagnosis call or target output was used.
No third agent was spawned. The only program infrastructure used is existing
bash, Python with isolated mode and site disabled, and its standard library.

There are shared filesystem permissions, including technical permission to read
files outside the behavioral allowlist. This is application-level audited
semantic separation, not OS/container isolation and not a cryptographic proof
of absence of every possible file access. The ledger records explicit file
reads/writes, shell commands and tool actions. Interpreter/stdlib dependency
reads and ordinary shell execution infrastructure are described separately;
their internal OS accesses were not individually traced. The initial cat command
used the default shell login setting; later shell commands explicitly disabled
login. No shell-startup contents were inspected or used as semantic inputs.
Python isolated mode prevents importing workspace modules or user site hooks;
only explicitly authored files are loaded as script/code inputs. No OS isolation
claim should be inferred from the use of -I/-S.

The author's fresh-agent/fork status is reported from the parent assignment,
not independently attested by an inaccessible orchestrator audit. The local
access ledger is author-generated evidence, not a hardened external monitor.
The current local UTC clock supplies sealing time; it is not externally signed.
Hashes ensure byte integrity from sealing onward, not historical independence.

The finite Boolean problem is mathematically exact but intentionally limited.
It does not cover probabilistic belief, natural language ambiguities, causal
interventions, infinite domains, noisy evidence, time-varying worlds, paraconsistent
reasoning, numerical floating point or retrieved sources. There are two-valued
complete worlds and a three-way epistemic partition only. Labels concern claims,
not general completeness of the entire world. Formal validity in this domain
does not establish broad diagnostic competence or generalization guarantees.

The 60 sealed records include 20 distinct evidence/query cores, grouped as ten
minimal pairs and represented three ways. Dependence is intentional; do not
treat them as 60 independent statistical samples or use naive significance
claims. Development has 32 separate generated cores. Ordinary logical operators
are shared as part of the task semantics, while structural organizations, bounds,
claims and seeds differ across partitions. No claim is made that no conceivable
logical equivalence exists between projected theories of different sizes.

The verifier is authored here and independent in its evaluation algorithm, but
it is not third-party formal assurance. Exhaustive finite checks and witness
certificates are independently inspectable. Main-instance compatibility has
not been checked here and may require STOP. This author neither knows nor
accommodates the target interface. The artifact set supplies no diagnosis score,
no E006 PASS and no permission to execute the sealed test.

Known implementation requirements: existing standard-library Python supporting
dictionary union (Python 3.9+), filesystem byte preservation, and enough memory
for at most 2048 possible worlds per case. SHA-256 and JSON UTF-8 are portable.
Generated payload bytes are deterministic; audit timestamps and seal metadata
are intentionally not part of deterministic corpus generation. The reproducer
never overwrites an existing directory. Any detected failure after sealing
requires STOP rather than creative repair.
