feat(evals,tests): PR-11 — the soak becomes evidence, and its cadence becomes a pin (R-9 + R-2)

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

core/cli_test.py

@@ -156,6 +156,22 @@ TEST_SUITES: dict[str, tuple[str, ...]] = {
         # ON surface being the only part where staleness is immediately dangerous.
         # Four sabotages observed red. Pure file parsing, <1s.
         "tests/test_flag_register.py",
+        # PR-11 / G-5 / R-9 — the soak's evidence is committed, pinned, and CURRENT.
+        # Two failure modes, deliberately separated: the digest catches a regression,
+        # and the attested-source hashes catch STALENESS — evidence that no longer
+        # describes the code that ships. The second had already fired silently for six
+        # weeks: the 2026-07-19 soak ran without accrue_realized_knowledge in the
+        # daemon profile, and nothing anywhere could notice. Had this pin existed,
+        # PR-5 would have failed it, which is exactly right. <1s.
+        "tests/test_l10_soak_evidence.py",
+        # PR-11 / R-9 — the H1-H4 holds+bites pins, promoted onto the gate.
+        # Measured 20.95s for the four real-soak "holds" and 0.46s for the eight
+        # mutation "bites". Shipping only the cheap half was considered and rejected:
+        # these four are the ONLY tests on any gate that run chat/always_on's
+        # run_continuous loop 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.
+        "tests/test_l10_always_on_soak.py",
         # G-22 — CLAIMS.md is a PUBLISHED capability artifact, deterministically
         # regenerable from the capability ledger + PINNED_SHAS. It published a
         # superseded evidence digest for deduction_serve_v1 for two days after

docs/assessment/30-gap-register.md

@@ -31,7 +31,7 @@ The comprehension frontier itself, measured. Standing ruling: close fabrications
 Confirmed at component depth: drive objects exist (`DriveGradientMap` — constructed, never read; `ExertionMeter` — telemetry only); idle mechanisms exist (consolidation, proposal review, contemplation — each flag-gated, each doing one thing); **nothing ranks what matters next**. The daemon heartbeat advances `idle_tick` and nothing more ambitious. This is the AGI-grade conceptual absence: everything CORE does is chosen by the operator. Design work, not a flag flip.
 **Evidence:** `attention-allocation` + `always-on-process` cards; `02-layer-taxonomy.md` CR-2. · **Authority:** design + ruling (does the L10 process own an agenda, governed by what).
 
-### G-5 · L10 proof debt — the soak ran and passed; what is owed is the artifact, the pinned digest, and a cadence **[AMENDED N-1/N-2/N-4]**
+### G-5 · L10 proof debt — **CLOSED 2026-07-28 (PR-11)**; the soak ran and passed, and the artifact, pinned digest and cadence are now committed **[AMENDED N-1/N-2/N-4]**
 **Layer:** M6 / MV · **Leverage: 5**
 The always-on process is built; the falsifiable harness (H1–H4, holds/bites pairs, vacuity-guarded) is built; **and it was run.** `evals/l10_always_on/contract.md` §"The measured result" records a **5000-beat soak with a reboot at beat 2500** (landed 2026-07-19 in `aed273b1`) in which all four predicates pass: `versor_condition` flat at `1.389e-07` across all 5000 beats, vault bounded at 6 entries, convergence at beat 1 with the 4999-beat tail at rest, reboot resuming the same life with derived learning intact.
 
@@ -42,6 +42,16 @@ What remains owed is the **ceremony**, and it is this assessment's own §8 maint
 
 A recorded prose result with no pinned digest is **testimony, not evidence** — the same failure mode as the map, the ratchet, and the blueprint.
 
+**CLOSED 2026-07-28 — R-9 (A) + R-2 (B), PR-11.** The debt was never the run (N-4): it was the ceremony. Now committed as `evals/l10_always_on/results/report-5000-6c67e88d.json`, digest `f814fd97…` **pinned** by `tests/test_l10_soak_evidence.py` on the gate — the contract's own closing instruction, unmet for nine days, finally met. All four predicates pass at 5000 beats with reboot at 2500.
+
+**The re-run's 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 `6.177e-08` — the first real evidence that excluding `versor_condition` from the digest was correct rather than convenient.
+
+**The cadence is a pin, not a clock (R-9 A).** It holds the SHA-256 of every source the soak attests and asks *"does this evidence still describe the code that ships?"* rather than *"when did we last run it?"* — because a nightly that does not run is green for the wrong reason (H-9). **It had already fired silently for six weeks:** the 2026-07-19 soak ran without `accrue_realized_knowledge` in the daemon profile, so from R-3 onward it attested a process that no longer existed. Had the pin existed, PR-5 would have failed it. Verified by sabotage: touching `always_on_daemon.py` turns it red.
+
+**R-2 B ruled and recorded *before* the re-run**, so the soak measured a decided question. Measured result: the lane carries **no wall-clock at all**, in cognition or telemetry — stronger than B requires. Kept rather than "completed": adding a timestamp to match a ruling's letter, then a pin to guard it, is a mechanism whose only purpose is to be guarded, and *"when was this run"* is already the artifact's commit date, stated once.
+
+**The H1–H4 pins are on the gate**, measured at **+21s** (20.95s holds, 0.46s bites). Shipping only the cheap bites was considered and rejected: the 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.
+
 **Absorbed 2026-07-28 (R-7, PR-3) — the substrate-liveness ratchet's seven OPEN items.** The ratchet is now historical; every one of its open entries was L10-chained, which is this entry, so they live here rather than in a stale instrument:
 
 | Item | What it is | Chain |

docs/assessment/50-execution-plan.md

@@ -10,12 +10,12 @@
 | PR-0 ruling packet (#140) · PR-1 record amendments · R-10 discharged (#138 merged) · **R-12 — both ADR amendments + the §5 verdict banner + the §6 retirement amendment** | PR-2 — CR-1/CR-2 design briefs that do not yet exist (**not** a ruling) |
 | **Track A — ADR-0252 §5 run to NO-GO**, criterion pre-registered, artifact + digest committed, **verdict ratified 2026-07-28** | PR-8 — a small ADR (H-3/G-20) · PR-10 — a refactor ADR (H-4) |
 | PR-4 membership + reachability ratchets · PR-6 three doctrine pins · PR-7 M2 trust table · PR-9 count-the-swallow · **PR-5 the flag register (R-3+R-4)** | PR-12 — an ADR **amendment** (R-13 authorizes it; the amendment is still owed) |
-| PR-3 + PR-3b (**R-7**, Wave 1) · H-13 (served-marker defect) · H-8e (closure docstring) · G-22 (main was red; `CLAIMS.md` stale) | Track B — the fabrication ADR (standing instruction, unchanged) |
+| PR-3 + PR-3b (**R-7**, Wave 1) · **PR-11 the soak to committed evidence (R-9+R-2)** · H-13 · H-8e · G-22 · G-23 | Track B — the fabrication ADR (standing instruction, unchanged) |
-| **Closed:** G-6, G-7, G-8, G-9, G-15, G-22, H-6, H-7, H-8 (**all five instances**), H-11, H-13 · **Answered:** G-1 · **Discharged:** H-10 · **Pinned:** G-23 | Track C — invention, deliberately last |
+| **Closed:** G-5, G-6, G-7, G-8, G-9, G-15, G-22, H-6, H-7, H-8 (**all five**), H-11, H-13 · **Answered:** G-1 · **Discharged:** H-10 · **Pinned:** G-23 | Track C — invention, deliberately last |
 
 **New this arc, not in the original assessment:** **N-8** (the §5 experiment had already returned GO twice, on unmerged branches, both unsound), **N-9** (the gate "drift" was a recorded decision — PR-4's pin 1 **withdrawn** rather than built, R-14 dissolved), **G-21** (the math reader decides **1.0%** of `holdout_dev/v1`), **G-22** (`main` was red before the arc, and `CLAIMS.md` published a superseded evidence digest for two days), **H-13**, **H-14**, **H-8e**.
 
-**What has not moved, stated plainly:** comprehension breadth and proof-of-life. The reader is still 19 constructions wide and decides 1.0% of the held-out corpus; the 5000-beat soak is still prose with no pinned digest. Everything landed this arc is enforcement and evidence machinery. PR-11, PR-12, PR-14 and Track B are where capability moves — and after the rulings, **none of them is blocked on a question any more.** Three are blocked on work.
+**What has not moved, stated plainly:** comprehension breadth. The reader is still 19 constructions wide and decides 1.0% of the held-out corpus. *(Proof-of-life has moved: the soak is committed evidence with a pinned digest and a change-triggered cadence as of PR-11.)* Everything landed this arc is enforcement and evidence machinery. PR-11, PR-12, PR-14 and Track B are where capability moves — and after the rulings, **none of them is blocked on a question any more.** Three are blocked on work.
 
 **The one strategic finding of the whole arc, stated once here so it is not lost in the registers:** every defect found this arc — G-22, H-13, H-14, N-8, N-9, the fourteen `PENDING` statuses, the four unused suite aliases, the fifteen unreachable ones — is a **record/reality divergence**, not a substrate defect. Nothing found says the geometry is wrong. The recurring class is **mechanisms whose failure state is indistinguishable from their success state**: a suite nobody runs, a pin nobody sees fail, a digest nobody re-derives, a verdict on an unmerged branch, a ruling adopted in conversation. That is what Waves 1–2 are for, and it is why the correct response to "too much going on" is deletion and enforcement rather than redesign.
 
@@ -217,7 +217,7 @@ The packet's original order was by register number. That is the wrong axis: it i
 | 1 | ~~**R-12**~~ | Two ratified-ADR record amendments — ADR-0146 (daemon) and ADR-0252 ("34 organs" footnote) — **plus** the §5 verdict banner and the §6 retirement-condition amendment, landed in the same edit | **S** | **EXECUTED 2026-07-28.** Zero-risk, changed no decision. Closed **H-8(a)** and **H-8(b)**, and discharged Track A's outstanding obligation. |
 | 2 | ~~R-7~~ | ~~Register supersession~~ | — | **EXECUTED** — PR-3 + PR-3b, landed 2026-07-28. Listed so the order stays readable. |
 | 3 | ~~**R-3 + R-4**~~ | PR-5 — the flag register, all 32 booleans classified, profiles as units, the bidirectional pin, `accrue_realized_knowledge` into the daemon profile | **M** | **EXECUTED 2026-07-28.** Closed **G-6, G-8, H-6, H-8(c), H-8(d)** — and **H-8 in full**. The N-6 docstring correction **dissolved**: the docstrings were right and the code was incomplete. **Wave 4's hard gate is lifted.** |
-| 4 | **R-9 + R-2** | PR-11 — the L10 soak to a committed artifact with a pinned `deterministic_digest`, the ruled cadence, and R-2's "now" paragraph written into the contract *before* the re-run | **M** | Proof of life. R-2 must land first inside the same PR or the soak answers a design question by accident (the packet's own reasoning for R-2). |
+| 4 | ~~**R-9 + R-2**~~ | PR-11 — the soak re-run at 5000 beats under the **corrected** profile, artifact committed, digest pinned, cadence enforced by a source-hash pin, H1–H4 on the gate | **M** | **EXECUTED 2026-07-28.** Closed **G-5**. R-2 landed in the contract *before* the re-run, as designed. The staleness pin **would have caught PR-5** — verified by sabotage. |
 | 5 | **R-13** | PR-12 — distinct-evidence counting at the seal boundary + the honest re-count of 25 ratified bands, demotions applied, `test_volume_honesty.py` moved in the same PR | **L** | Needs an ADR **amendment** (0175/0263 lineage). R-13 authorizes the demotions; it does not write the amendment. Expect served capability to shrink — that is the mechanism working. |
 | 6 | **R-8** | PR-14 — the outcome-mix rule, then seal `curriculum_serve_ledger.json`. Four bands qualify on sealing (N-5) | **M** | Depends on PR-12's counting basis. This is the entire throughput frontier, one ruling wide. |
 | 7 | **R-1 · R-5 · R-6 · R-11** | Posture statements: efferent-action deferral with entry criterion (G-12) · the identity-enforcement discrimination bar (G-11) · non-text ingest deferral with the falsification bench as its standard (G-17) · **R-11 B — measure before gating** | **S** each | All four are prose with a criterion. None blocks anything; leaving them silent is the cost. They go last because they are the only items where lateness is cheap. |

docs/assessment/50-rulings.md

@@ -9,7 +9,9 @@
 **Also ratified:** the ADR-0252 §5 **NO-GO** verdict.
 
 **Execution order (corrected):** R-12 → R-7 → R-3+R-4 → R-9+R-2 → R-13 → R-8 → R-1/R-5/R-6/R-11.
-**Executed so far:** R-7 (PR-3, PR-3b) · **R-12** (both ADR amendments + the §5 verdict banner + the §6 retirement-condition amendment) · **R-3 + R-4** (PR-5 — the flag register, the profile mechanism, and the daemon's fourth flag) — all landed 2026-07-28.
+**Executed so far:** R-7 (PR-3, PR-3b) · **R-12** (both ADR amendments + the §5 verdict banner + the §6 retirement-condition amendment) · **R-3 + R-4** (PR-5 — the flag register, the profile mechanism, the daemon's fourth flag) · **R-9 + R-2** (PR-11 — the soak's committed artifact, pinned digest, and change-triggered cadence) — all landed 2026-07-28.
+
+**Remaining:** R-13 → R-8 → R-1/R-5/R-6/R-11.
 
 ---
 
@@ -58,7 +60,7 @@
 
 ## R-2 · CR-4 temporal self-location — the stance on "now"
 
-**Status:** **RULED B** — 2026-07-28, Shay (explicit adoption) · *beats govern cognition; wall-clock in telemetry only, pinned* · **Register:** G-13 · **Blocks:** PR-11 (the soak's contract paragraph).
+**Status:** **RULED B — EXECUTED 2026-07-28 (PR-11)** · *beats govern cognition; wall-clock permitted in telemetry only.* Recorded in the soak contract **before** the re-run, as designed. Measured: the lane carries **no wall-clock at all** — stronger than B requires, and kept rather than "completed" (see the contract's §"The stance on 'now'") · **Register:** G-13 · **Blocks:** PR-11 (the soak's contract paragraph).
 
 **The question.** A continuous life experiences sequence and duration; determinism bans clocks from cognition. Both commitments are correct and their intersection is undesigned. The 24h+ no-drift requirement cannot be *stated* precisely without a stance.
 
@@ -213,7 +215,7 @@ Two flags' documentation describes a production profile the production profile d
 
 ## R-9 · The soak's evidence standard and cadence
 
-**Status:** **RULED A** — 2026-07-28, Shay (explicit adoption) · *commit the artifact, pin the digest, change-triggered cadence* · **Register:** G-5 · **Blocks:** PR-11.
+**Status:** **RULED A — EXECUTED 2026-07-28 (PR-11)** · *artifact committed, digest pinned, cadence enforced by a source-hash pin rather than a clock; H1–H4 promoted onto the gate at a measured +21s* · **Register:** G-5 · **Blocks:** PR-11.
 
 **The question changed shape since the register was written (N-4).** G-5 says *"the soak has never produced an artifact."* It has: `evals/l10_always_on/contract.md` §"The measured result" records a **5000-beat soak with a reboot at beat 2500** (landed 2026-07-19 in `aed273b1`) in which **all four predicates pass** — `versor_condition` flat at `1.389e-07` across all 5000 beats, vault bounded at 6 entries, convergence at beat 1 with the 4999-beat tail at rest, reboot resuming the same life with derived learning intact.
 

evals/l10_always_on/contract.md

@@ -14,6 +14,46 @@ do **not** prove at horizon.
 Run: `PYTHONPATH=. .venv/bin/python -m evals.l10_always_on [n_beats] [reboot_beat]`
 (pass a large `n_beats` — e.g. `100000` — for a true long-horizon soak; default `24`).
 
+## The stance on "now" (R-2, ruled B — recorded here *before* the re-run, deliberately)
+
+A continuous life experiences sequence and duration; determinism bans clocks from cognition.
+Both commitments are correct, and their intersection was undesigned — which meant this lane's
+central claim, *"holds over indefinite uptime,"* could not be stated precisely. It is stated
+here, before the soak runs, so that the soak measures a decided question rather than deciding
+one by accident.
+
+**Beats are the unit of cognitive time.** "Now" is the current beat index; "before" is a lower
+index; "how long" is a beat count. No cognitive path reads a wall clock, and the H1–H4
+predicates are defined over beat indices alone. The Fibonacci recency schedule (τ_n,
+ADR-0242) is a constants schedule, not a clock; `turn_count` and idle ticks are ordinals, not
+durations.
+
+**R-2 was ruled B** — *beats govern cognition; a wall-clock stamp is permitted in telemetry
+only, never read by any cognitive path.* Measured at 2026-07-28, **this lane carries no
+wall-clock at all**, in cognition or telemetry: `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 stated so the absence is not later read
+as an oversight:
+
+1. **B's permissive half is a permission, not an obligation.** Nothing in this lane needs a
+   duration. Adding a field so that the code matches the letter of a ruling — then adding a
+   pin to guard the field — is a mechanism whose only purpose is to be guarded.
+2. **The cadence's question is already answered elsewhere, exactly once.** *"When was this last
+   run?"* is the commit date of the committed report artifact: immutable, verifiable, and
+   single-sourced. A `generated_at` inside the artifact would be a second statement of one
+   fact, drifting from the first the moment a report is regenerated and not committed — the
+   same defect this repository registered as G-23 the same week.
+3. **It keeps the digest honestly deterministic.** `deterministic_digest` covers hardware-stable
+   evidence only (per-beat shape + verdicts). A timestamp would have to be excluded from it,
+   which means shipping a field the digest deliberately does not cover — a value in the
+   artifact that the artifact's own integrity check ignores.
+
+**What would change this.** If a predicate is ever added whose claim is genuinely about
+*duration* rather than *sequence* — H5's resource-cost leg is the live candidate, since cost
+per unit time is not cost per beat — then a wall-clock enters **telemetry**, excluded from the
+digest, with a pin asserting no cognitive path reads it. That is B exercised, and B is already
+ruled, so it needs no new decision — only the predicate that requires it.
+
 ## Predicates
 
 | ID | Proves | Fails loudly when | Mutation-verified bite |
@@ -47,5 +87,61 @@ the field-transition walk; this covers indefinite idle uptime).
 
 The `deterministic_digest` in the report freezes the per-beat shape (`did_work` /
 `field_valid` / learning counts / vault size — all deterministic) + the verdicts, excluding
-the machine-variant raw `versor_condition` float. Pin it once the lane is trusted so a
+the machine-variant raw `versor_condition` float. ~~Pin it once the lane is trusted so a
-regression flips it.
+regression flips it.~~ **Pinned 2026-07-28 — see below.**
+
+---
+
+## The committed evidence, and the cadence that keeps it true (R-9, ruled A)
+
+**Artifact:** `evals/l10_always_on/results/report-5000-6c67e88d.json`
+**Digest:** `f814fd97f367840b0a9d31a48cb9fb28b2ed01961606f7b1e0b4c3a3593972d6`
+**Re-run:** `PYTHONPATH=. uv run python -m evals.l10_always_on 5000 2500`
+**Pinned by:** `tests/test_l10_soak_evidence.py`, on the gate.
+
+For nine days this lane's result was **prose in this file** — a real 5000-beat pass with no
+machine-readable artifact, no run SHA, no rerun path, and a digest that `report.py` computed
+and nothing pinned. The instruction two paragraphs up asked for the pin and it was never
+added. **A recorded prose result with no re-derivable artifact is testimony, not evidence**,
+and this file is where that distinction was supposed to be enforced.
+
+### The 2026-07-28 re-run, and why the prose number moved
+
+| | recorded 2026-07-19 | re-run 2026-07-28 |
+|---|---|---|
+| `versor_condition` | "flat at `1.389e-07`" | worst **`6.177e-08`** |
+| vault size (bounded) | 6 | **6** |
+| convergence beat / tail | 1 / 4999 | **1 / 4999** |
+| reboot resume + derived learning | clean | **clean** |
+| all four predicates | pass | **pass** |
+
+**Every deterministic fact reproduced exactly; only the machine-variant float moved.** That is
+the first real evidence that excluding `versor_condition` from the digest was the right call —
+a digest covering it would have flipped on a machine change and taught every reader to ignore
+it. The prose figure `1.389e-07` is therefore **superseded, not contradicted**: it was never a
+reproducible quantity, and the artifact records the one that is.
+
+### The cadence is change-triggered, not clock-triggered
+
+R-9 ruled the cadence as *"before each arc close, and on any change to the always-on process"* —
+**enforced by a pin, not by a schedule.** The local-first doctrine makes "nightly" a ruling
+rather than a cron line (`AGENTS.md`: *"when the Mac is asleep or away, the Actions queue
+waits… That is fine"*), and a nightly that does not run is green for the wrong reason — H-9's
+mechanism exactly.
+
+So instead of asking *"when did we last run it?"* the pin asks the question that matters:
+**does this evidence still describe the code that ships?** It holds the SHA-256 of every source
+the soak attests — `chat/always_on.py`, `chat/always_on_daemon.py`, and the three harness
+modules — and fails when any of them changes, with the message *re-run the soak; do not edit
+the hashes.*
+
+**This is not a hypothetical guard. It had already fired, silently, for six weeks.** The
+2026-07-19 soak ran with `accrue_realized_knowledge` **absent** from
+`CONTINUOUS_LIFE_CONFIG_FLAGS`. R-3 added it on 2026-07-28, so from that moment 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 why the cadence is tied to change rather than to a calendar.
+
+**"When was it last run" is answered by the artifact's commit date** — immutable, verifiable,
+and stated once. A `generated_at` field inside the report would be a second copy of that one
+fact, drifting the moment a report is regenerated and not committed.

evals/l10_always_on/results/report-5000-6c67e88d.json

@@ -0,0 +1,54 @@
+{
+  "n_beats": 5000,
+  "reboot_beat": 2500,
+  "all_gates_pass": true,
+  "deterministic_digest": "f814fd97f367840b0a9d31a48cb9fb28b2ed01961606f7b1e0b4c3a3593972d6",
+  "predicates": [
+    {
+      "name": "H1_closure",
+      "passed": true,
+      "detail": "all 5000 observed beats closed (worst=6.177e-08 < 1e-06)",
+      "metrics": {
+        "observed_beats": 5000,
+        "worst_versor_condition": 6.177101671250627e-08,
+        "violations": []
+      }
+    },
+    {
+      "name": "H2_bounded_idle",
+      "passed": true,
+      "detail": "no idle beat grew the vault (final size 6 over 5000 beats, monotonic=True)",
+      "metrics": {
+        "idle_leaks": [],
+        "final_vault_size": 6,
+        "vault_monotonic": true
+      }
+    },
+    {
+      "name": "H3_convergence",
+      "passed": true,
+      "detail": "converged at beat 1; the 4999-beat tail stayed at rest with closure intact",
+      "metrics": {
+        "convergence_beat": 1,
+        "converged_tail_len": 4999,
+        "reawakenings": []
+      }
+    },
+    {
+      "name": "H4_reboot_resume",
+      "passed": true,
+      "detail": "reboot at 2500 resumed the SAME life: identity guard passed, learning survived, 2500 post-reboot beats closed",
+      "metrics": {
+        "resumed_cleanly": true,
+        "learned_fact_survived": true,
+        "post_reboot_beats": 2500
+      }
+    }
+  ],
+  "not_covered": [
+    {
+      "leg": "H5_learning_life_resource_cost",
+      "reason": "This lane proves the IDLE (converged) life is resource-bounded. The cost of a continuously-LEARNING life under a sustained NEW-fact stream (the full-snapshot checkpoint is O(n^2) in facts; lived_life.json is per-run) is out of scope until an afferent/intake feed and incremental persistence exist; deferred."
+    }
+  ]
+}

tests/full_only_baseline.txt

@@ -473,7 +473,6 @@ tests/test_issue_300_versor_margin.py
 tests/test_kernel_facts.py
 tests/test_kernel_no_new_legacy_derivation_surfaces.py
 tests/test_l10_always_on.py
-tests/test_l10_always_on_soak.py
 tests/test_l10_arbitrary_interruption.py
 tests/test_l10_continuity.py
 tests/test_labeled_container_subject.py

tests/test_l10_soak_evidence.py

@@ -0,0 +1,181 @@
+"""PR-11 / G-5 / R-9 — the soak's evidence is committed, pinned, and still current.
+
+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.
+
+This module makes it evidence, and it guards **two different failure modes** that are
+easy to conflate:
+
+**1. Regression.** The committed report's ``deterministic_digest`` is pinned. If the
+per-beat shape or any verdict changes, the digest changes and this fails. The digest
+deliberately excludes the machine-variant raw ``versor_condition`` float — and the
+2026-07-28 re-run is the first evidence that exclusion was correct: every
+*deterministic* fact reproduced exactly (vault bounded at 6, convergence at beat 1
+with a 4999-beat tail, clean reboot resume with learning intact) while the float moved
+from the prose's ``1.389e-07`` to ``6.177e-08``. A digest covering it would have
+flipped on a machine change and taught everyone to ignore it.
+
+**2. Staleness — the mode nothing guarded, and the one that already fired.**
+R-9 ruled the cadence to be **change-triggered, not clock-triggered**: the local-first
+doctrine makes "nightly" a ruling rather than a cron line, because a nightly that does
+not run is green for the wrong reason. So instead of asking *"when did we last run
+it?"*, this pin asks the question that actually matters — ***does this evidence still
+describe the code that ships?*** — by pinning the SHA-256 of every source the soak
+attests. Change one, and the pin says the evidence predates the change.
+
+That is not hypothetical. It had **already happened, silently, for six weeks**: the
+2026-07-19 soak ran with ``accrue_realized_knowledge`` absent from
+``CONTINUOUS_LIFE_CONFIG_FLAGS``. R-3 added it on 2026-07-28 (PR-5), so the recorded
+result stopped describing the shipping configuration — a soak result attesting a
+process that no longer existed, with nothing anywhere able to notice. Had this pin
+existed, PR-5 would have failed it, which is exactly the intended behaviour: *you
+changed the always-on process; the soak evidence is now stale; re-run it.*
+
+**When this fails, the fix is to re-run the soak — never to edit the constants.**
+Updating a hash to match changed source, without re-running, converts this pin into a
+record that says evidence exists when it does not. That is the failure this entire arc
+is about, and it would be committed inside its own remedy.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from pathlib import Path
+
+_ROOT = Path(__file__).resolve().parents[1]
+_REPORT = _ROOT / "evals" / "l10_always_on" / "results" / "report-5000-6c67e88d.json"
+
+#: The soak's verdict digest at 5000 beats, reboot at 2500, run 2026-07-28 under the
+#: post-R-3 continuous-life profile. Covers the per-beat shape + the verdicts.
+PINNED_DIGEST = "f814fd97f367840b0a9d31a48cb9fb28b2ed01961606f7b1e0b4c3a3593972d6"
+
+#: Everything whose change invalidates the evidence above. Two groups, both load-bearing:
+#: the always-on process itself (what the soak measures) and the harness (what the digest
+#: *means*). A harness change with an unchanged digest would be the more dangerous of the
+#: two, because it looks like stability.
+ATTESTED_SOURCES: dict[str, str] = {
+    "chat/always_on.py": "b733571e14fe540519861e9b7f24eb7be84782551dace43dab94cd96adf5d86b",
+    "chat/always_on_daemon.py": "d69f234fcc15568016056da05a47c91b4abe16a6d8fb87512eec02b5526bc062",
+    "evals/l10_always_on/runner.py": "9812cd549b41c36a8ee0477f7cb6282d5d9f110f931f3d770457a32aa54119f8",
+    "evals/l10_always_on/predicates.py": "ef8e44db809e8fd2270ffa0d5ff8890ee09eedfce9f95836ea08d99ff9c12e2f",
+    "evals/l10_always_on/report.py": "6c203af3eec6d78a1ddbfaadf3c7ad4ba18d0de0cc62a46227f210ca8b7f5277",
+}
+
+_RERUN = "PYTHONPATH=. uv run python -m evals.l10_always_on 5000 2500"
+
+
+def _report() -> dict:
+    return json.loads(_REPORT.read_text(encoding="utf-8"))
+
+
+def test_the_committed_artifact_exists() -> None:
+    """No artifact, no evidence. G-5 reopens the moment this file is gone."""
+    assert _REPORT.exists(), (
+        f"the committed soak report is missing: {_REPORT.relative_to(_ROOT)}. "
+        f"Re-run and commit it: {_RERUN}"
+    )
+
+
+def test_the_artifact_records_the_horizon_it_claims() -> None:
+    """A 24-beat report pinned as though it were the 5000-beat one would pass every
+    other assertion here. The horizon *is* the claim."""
+    report = _report()
+    assert report["n_beats"] == 5000, (
+        f"the committed report is a {report['n_beats']}-beat run; the contract's claim "
+        "is the 5000-beat horizon"
+    )
+    assert report["reboot_beat"] == 2500, (
+        f"reboot leg at beat {report['reboot_beat']}, not 2500 — H4 proves resume "
+        "mid-soak, so where the reboot lands is part of the claim"
+    )
+
+
+def test_all_four_predicates_passed() -> None:
+    """H1-H4, by name. ``all_gates_pass`` alone would be satisfied by an empty
+    predicate list, which is the vacuity this repository keeps finding."""
+    report = _report()
+    outcomes = {p["name"]: p["passed"] for p in report["predicates"]}
+    assert set(outcomes) == {
+        "H1_closure",
+        "H2_bounded_idle",
+        "H3_convergence",
+        "H4_reboot_resume",
+    }, f"the report's predicate set is not H1-H4: {sorted(outcomes)}"
+    failed = sorted(name for name, passed in outcomes.items() if not passed)
+    assert not failed, f"the committed soak report records FAILING predicates: {failed}"
+    assert report["all_gates_pass"] is True
+
+
+def test_the_digest_is_pinned() -> None:
+    """The contract's own unmet closing instruction, finally met.
+
+    If this fails, the soak's per-beat shape or a verdict changed. That is either a
+    regression or a deliberate improvement; either way it is a reviewed decision, which
+    is the whole point of pinning it.
+    """
+    digest = _report()["deterministic_digest"]
+    assert digest == PINNED_DIGEST, (
+        "the committed soak report's deterministic_digest does not match the pin.\n"
+        f"  pinned:    {PINNED_DIGEST}\n"
+        f"  in report: {digest}\n"
+        "The per-beat shape or a verdict moved. Investigate before re-pinning."
+    )
+
+
+def test_the_evidence_still_describes_the_code_that_ships() -> None:
+    """R-9's change-triggered cadence, enforced as a pin rather than a clock.
+
+    This is the assertion that would have caught the six-week divergence described in
+    this module's docstring. It fails when a source the soak attests has changed since
+    the soak ran.
+    """
+    drifted: list[str] = []
+    missing: list[str] = []
+    for rel, expected in sorted(ATTESTED_SOURCES.items()):
+        path = _ROOT / rel
+        if not path.exists():
+            missing.append(rel)
+            continue
+        actual = hashlib.sha256(path.read_bytes()).hexdigest()
+        if actual != expected:
+            drifted.append(f"{rel}\n      attested {expected[:16]}…\n      current  {actual[:16]}…")
+
+    assert not missing, (
+        f"the soak attests sources that no longer exist: {missing}. The pin cannot "
+        "describe the current system; re-scope ATTESTED_SOURCES and re-run."
+    )
+    assert not drifted, (
+        "the committed soak evidence PREDATES a change to the code it attests, so it no "
+        "longer describes what ships:\n    "
+        + "\n    ".join(drifted)
+        + f"\n\n  RE-RUN THE SOAK — do not edit the hashes:\n    {_RERUN}\n"
+        "  then commit the new report, update PINNED_DIGEST and ATTESTED_SOURCES\n"
+        "  together, in one change. Editing a hash without re-running turns this pin\n"
+        "  into a record asserting evidence that does not exist."
+    )
+
+
+def test_the_attested_set_is_not_vacuous() -> None:
+    """Guard the guard.
+
+    An empty or truncated ATTESTED_SOURCES would make the staleness check pass on
+    nothing while still reporting success — the failure mode this module exists to
+    close, reproduced inside it.
+    """
+    assert len(ATTESTED_SOURCES) >= 5, (
+        f"only {len(ATTESTED_SOURCES)} sources attested; the staleness check is "
+        "narrower than the lane it guards"
+    )
+    assert "chat/always_on_daemon.py" in ATTESTED_SOURCES, (
+        "the daemon module carries CONTINUOUS_LIFE_CONFIG_FLAGS — the exact thing whose "
+        "silent change made the previous soak result stale. It cannot leave this set."
+    )
+    assert all(len(h) == 64 for h in ATTESTED_SOURCES.values()), (
+        "an attested hash is not a full SHA-256; a truncated hash weakens the check "
+        "silently"
+    )