Rank 4 on the docket. The L10 always-on lane has passed a 5000-beat soak since
2026-07-19. That result lived as PROSE IN A CONTRACT FILE — no machine-readable
artifact, no run SHA, no rerun path, and a deterministic_digest that report.py
computed and nothing pinned. The contract's own closing line asked for the pin
("Pin it once the lane is trusted so a regression flips it") and it was never
added. A recorded prose result with no re-derivable artifact is testimony, not
evidence, and that file is exactly where the distinction was meant to be
enforced.
R-2 first, before the run — as the plan required, so the soak measured a
decided question rather than deciding one by accident.
R-2 was ruled B (beats govern cognition; wall-clock permitted in telemetry
only). Measured: this lane carries NO WALL-CLOCK AT ALL — runner.py,
predicates.py and report.py contain no time, datetime, or timestamp of any
kind. That is stronger than B requires, and it is KEPT rather than "completed",
for three reasons now recorded in the contract:
1. B's permissive half is a permission, not an obligation. Adding a field so
the code matches a ruling's letter, then adding a pin to guard the field,
is a mechanism whose only purpose is to be guarded.
2. "When was this last run?" is already the artifact's COMMIT DATE —
immutable, verifiable, stated once. A generated_at inside the report would
be a second copy of one fact, drifting the moment a report is regenerated
and not committed. That is G-23's defect class, registered this same week.
3. It keeps the digest honestly deterministic. A timestamp would have to be
excluded from it — a value in the artifact that the artifact's own
integrity check ignores.
The contract also records what WOULD change this: a predicate whose claim is
about duration rather than sequence (H5's resource-cost leg is the live
candidate). B is already ruled, so that needs no new decision — only the
predicate that requires it.
The re-run: 5000 beats, reboot at 2500, under the POST-R-3 profile. All four
predicates pass. Committed as
evals/l10_always_on/results/report-5000-6c67e88d.json, digest f814fd97….
The most useful result was not the pass. Every DETERMINISTIC fact reproduced
exactly — vault bounded at 6, convergence at beat 1 with a 4999-beat tail,
clean reboot with derived learning intact — while the machine-variant float
moved from the prose's 1.389e-07 to a worst of 6.177e-08. That is the first
real evidence that excluding versor_condition from the digest was correct
rather than convenient: a digest covering it would have flipped on a machine
change and taught every reader to ignore it. The prose figure is superseded,
not contradicted — it was never a reproducible quantity.
tests/test_l10_soak_evidence.py, on the gate, guarding two DIFFERENT failure
modes that are easy to conflate:
REGRESSION — the digest is pinned, plus the horizon (a 24-beat report pinned
as though it were the 5000-beat one would pass everything else), the four
predicates by name (all_gates_pass alone is satisfied by an empty predicate
list), and a vacuity guard on the attested set.
STALENESS — the mode nothing guarded, and the one that had ALREADY FIRED.
R-9 ruled the cadence change-triggered, not clock-triggered: local-first
doctrine makes "nightly" a ruling rather than a cron line, and a nightly that
does not run is green for the wrong reason (H-9). So instead of "when did we
last run it?", the pin asks the question that matters — DOES THIS EVIDENCE
STILL DESCRIBE THE CODE THAT SHIPS? — by holding the SHA-256 of every source
the soak attests: chat/always_on.py, chat/always_on_daemon.py, and the three
harness modules.
That is not hypothetical. It had already fired silently for six weeks: the
2026-07-19 soak ran with accrue_realized_knowledge absent from
CONTINUOUS_LIFE_CONFIG_FLAGS, so from R-3 (PR-5, this same day) onward the
recorded result attested a process that no longer existed — and nothing in the
repository could notice. HAD THIS PIN EXISTED, PR-5 WOULD HAVE FAILED IT. That
is precisely the intended behaviour, and it is verified by sabotage: touching
always_on_daemon.py turns it red.
The failure message says RE-RUN THE SOAK — DO NOT EDIT THE HASHES, because
updating a hash to match changed source without re-running converts the pin
into a record asserting evidence that does not exist. That is the failure this
whole arc is about, committed inside its own remedy.
Three sabotages observed red: a touched daemon (the staleness mode), a
regressed verdict in the artifact, and a shorter horizon pinned as if it were
5000 beats.
H1-H4 promoted onto the gate, and the cheap option was deliberately refused.
Measured 20.95s for the four real-soak "holds" and 0.46s for the eight mutation
"bites" — 98% of the cost is the holds. Shipping only the bites, or parking the
file in a curated-but-unreachable suite, would have satisfied R-9's letter and
bought nothing; a suite nobody runs is the non-guarantee this arc spent PR-4
closing. Those four holds are the ONLY tests on any gate that run
chat/always_on's run_continuous end-to-end, and that loop just shipped a
six-week silent divergence. 21s per push to actually run the process that broke
is the most defensible time on this gate. Removed from full_only_baseline.txt
in the same edit; the baseline shrank 746 -> 745.
Closes G-5. Proof-of-life has moved from prose to committed, pinned,
change-guarded evidence.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wcw2pnMBwyvmNyQg4uPEt4
The always-on PROCESS shipped (run_continuous + daemon + CLI), but the claim "it holds
over uptime" was UNFALSIFIED: the existing evals/l10_continuity soak drives the TURN loop
and is disjoint from run_continuous; the daemon path was covered only by ≤5-beat unit
tests. This lane converts "built" → "proven over horizon" for the idle path.
evals/l10_always_on/ mirrors the l10_continuity predicate/mutation harness but drives the
IDLE heartbeat: it seeds a real continuous life (held self + a cognitive turn to excite the
field), runs N beats with NO user turn, and gates four falsifiable predicates, each with a
*_holds (real soak) AND a *_bites (mutation) test per CLAUDE.md schema-as-proof:
- H1 closure — every observed idle beat versor_condition < 1e-6 (held over uptime, READ
never repaired); bites on a breached beat AND on a vacuous no-field soak.
- H2 bounded idle — a no-work beat adds NOTHING to the vault (no idle leak — invisible at
5 beats, fatal at 100k); bites on idle vault growth.
- H3 convergence — a saturated idle life SETTLES and stays settled (no re-awakening), at
rest at the end, closure intact on the tail; bites on never-settling + on re-awakening.
- H4 reboot resume — a reboot mid-soak resumes the SAME life (strict identity guard passes,
pre-reboot DERIVED learning survives, post-reboot closure holds); bites on failed
resume + on lost learning.
Measured (5000 idle beats, reboot@2500, ~20s): ALL gates pass — versor_condition flat at
1.389e-07 (no drift, no repair), vault bounded at 6 (no idle leak), converged at beat 1
with a 4999-beat at-rest tail, reboot resumed the same life with learning intact. This is
the empirical resolution of the L10 riskiest-unknown FOR THE IDLE PATH (the closure-by-
construction ruling covered the field-transition walk; this covers indefinite idle uptime).
Honestly NOT covered (recorded in the report, no silent skip): the continuously-LEARNING
life's resource cost under a sustained new-fact stream (O(n²) snapshot; per-run lived_life)
— out of scope until an afferent/intake feed + incremental persistence exist (the next
frontier).
Run on demand: PYTHONPATH=. python -m evals.l10_always_on [n_beats] [reboot_beat]. Path-
scoped out of CI smoke (like l10_continuity); runs in post-merge full-pytest. Additive — a
new eval lane, no existing files touched. 12/12 lane tests + architectural invariants green.
Claude