THE DECISION TO SLEEP — public reproduction methods, 2026-10-25

This directory is a field-level allowlist, not an anonymized dump of a database.
Original Discord IDs, participant identity outside Blaine/Kairo/Caelum, job keys,
hostnames, internal paths, private narrative bodies and credentials are excluded.
One personal paragraph of Caelum's antecedent message is explicitly redacted.
Historical bot mentions are replaced with [Kairo mention]. Full private originals
remain in a separate permission-restricted evidence store and are not in this ZIP.

Window: [2026-10-24T04:00:00Z, 2026-10-25T15:25:00Z). EDT is UTC-4.
Capture types: SQLite mode=ro/query_only transaction; PostgreSQL BEGIN READ ONLY
and statement_timeout 30s; Discord HTTP GET pagination. No model invocation,
memory write, process restart, sleep/wake call or live intervention was performed.
Four live process environments and six module resolutions/hashes were inspected.
The raw captures are separate snapshots. Inclusion by created_at/occurred_at can
miss boundary-spanning jobs and events recorded late. Token figures are request
receipts, not unique information, all background activity, energy or dollars.

Historical search: retained platform jobs from July 20 to cutoff, SQL substring
sleep/exhaust candidates before Oct24 04:00Z, then regex terminal-only
\bsleep(?:ing)?\b. Journal prior candidates: whole-word sleep/exhausted/rest in
conversation.turn.completed, experience.note.recorded and thought.completed.
Memory episodes: sleep/exhausted whole words. Identity versions/candidates:
sleep/rest/dream/exhausted in keys/values/claims. These searches are not exhaustive
semantic retrieval. Counts of transport messages, jobs, preferences, retrieved
memories and inference calls have different denominators and are never combined.
Focal-turn coding is manual, single coder, selected conversation, not a prevalence
estimate. Criteria and all original focal questions are printed in the report.

Public reproduce.py checks the released derived data. It cannot revalidate
original server authenticity or private crosswalks. SHA-256 proves artifact
integrity only. Dream content_sha256 is the original stored deduplication key.
Dream links are validated against private imagination episode IDs and exact
titles; only the booleans are released. No behavioral benefit is inferred.

Fresh local isolated test commands (from repository unless otherwise stated):
PYTHONPATH=src:server/memory:runtime/memory-worker .venv/bin/python -m pytest -q \
 tests/test_sleep_admission.py tests/test_sleep_entrypoints.py \
 tests/test_sleep_ingress.py tests/test_sleep_runtime.py \
 tests/test_sleep_integrity_architecture.py tests/test_engineering_execution.py \
 tests/test_engineering_store.py tests/test_engineering_runtime.py
# 295 passed; original JUnit hash in tests.json

PYTHONPATH=src:server/memory:runtime/memory-worker .venv/bin/python -m pytest -q \
 tests/test_platform_store.py tests/test_metaframes_outbox.py \
 tests/test_metaframes_exporter.py tests/test_metaframes_worker.py \
 tests/test_journal_memory_lineage.py
# 111 passed; overlaps other previously reported suites

# working directory tools/kairo-console:
.venv/bin/python -m pytest -q -ra tests
# 90 passed, including browser tests; dedicated pre-existing environment

Raw logs and JUnit files are private/local, not packaged. tests.json includes
original hashes, aggregate outcomes, elapsed time and every collected test name.
No sum of overlapping suites is reported as an independent-test count.

Literature: targeted narrative review; exact-title searches plus primary-source
sleep/replay/consolidation/agent-memory/self-report searches on 2026-10-25.
19 primary works extracted in literature.json. Preprints, peer-reviewed work,
reviews and conceptual papers are labeled. NOT VERIFIED IN EXTRACTION is not
equivalent to claiming that a paper did not measure a quantity. NOT MEASURED in
the Kairo case means no such observation is available from this investigation.
No fabricated confidence intervals, causal treatment effects or significance
tests are used. Published intervals and calculated contrasts retain source units.
