# Exhibit thirteen — the bench, in full

*Generated 2026-08-16T15:15:02Z by `harness/thirteen_report.py` from `results/scores.json` and the sealed receipts. Every figure below is read out of an artifact; none is typed. This is the evidence, not the article.*

**Seal** `2026-08-16T14:49:33Z` · 11/12 preconditions met · 22 files hashed

## 1. What ran

| leg | sealed plan | measured |
| --- | --- | --- |
| scored cells (8 arms × 3 conditions × 24 items) | **576** | **557** collected + 19 not-collected |
| warmups (8 × 3, discarded, receipted) | 24 | 30 |
| flake probe (zero-cost arms × 6 items × 2 extra draws) | 72 | 72 |
| metered output probes | 2 probes | **4 calls** (2 probes × 2 items each) |
| **total wire calls** | **674** | **663** |

The plan's table derived 587 for a seven-arm roster. Orchestrator ruling 1 seated an eighth arm, so the arithmetic is recomputed from the roster: **674**.

The eleven-call gap between plan and measurement decomposes exactly: −19 scored cells the budget event never collected, +6 extra warmups taken by the flake leg, +2 probe calls the plan did not budget. Each registered probe turned out to be two calls — items `N03` and `B01` — so four probe calls crossed the wire against a plan that counted two. All four rows are itemized in `runtime-pins.json`, and the $0.0516 pre-seal probe line in `run-receipt.json` prices all four.

> **CORRECTED 2026-08-16, after publication.** This section previously printed only the sealed plan's column, under a heading that reads *What ran* — so **674** stood as though it were a measurement and the metered-probe row read **2**. The measured column above is new; the plan column is unchanged and is still the sealed arithmetic. The page's headline call figure was corrected in the same pass, from **661** to **663**. No scored result depended on any of it: `cells_read` was 576 then and is 576 now.

Cells actually read by the scorer: **576**.

- `C-NONE` — 173 collected, 19 not-collected, 0 skipped (already journalled), Nones wall
- `C-MAP,C-HTML` — 384 collected, 0 not-collected, 0 skipped (already journalled), 499.5s wall
- `C-MAP (flake probe)` — 72 collected, 0 not-collected, 36 skipped (already journalled), 104.3s wall

## 2. The roster — vendor, family and class, printed

| seat | exact tag | vendor | family | transport | class | bill |
| --- | --- | --- | --- | --- | --- | --- |
| `local-qwen3.8-27b` | `qwen3.8:27b` | alibaba | qwen | local-ollama | local ~27b | $0.00 |
| `local-qwen3.6-27b` | `qwen3.6:27b` | alibaba | qwen | local-ollama | local ~27b | $0.00 |
| `local-gemma4-26b` | `gemma4:26b` | google | gemma | local-ollama | local ~26b | $0.00 |
| `local-llama3.3-70b` | `llama3.3:70b` | meta | llama | local-ollama | local ~70b | $0.00 |
| `cloud-glm-5.2` | `glm-5.2` | zhipu | glm | ollama-cloud | frontier cloud | $0.00 |
| `cloud-deepseek-v4-pro` | `deepseek-v4-pro:preview` | deepseek | deepseek | ollama-cloud | frontier cloud | $0.00 |
| `cloud-kimi-k3` | `kimi-k3` | moonshot | kimi | ollama-cloud | frontier cloud | **METERED** (sub-cap $0.90) |
| `openai-gpt-5.5` | `gpt-5.5-2026-04-23` | openai | gpt | openai-api | frontier cloud | **METERED** (sub-cap $2.60) |

**8 arms / 7 families / 7 vendors.** Four arms are local seats on one box and four are frontier cloud; two arms share family *and* size class. No pooled cross-arm figure in this report prints without this table beside it.

Two absences, stated rather than left to be noticed: **no Claude arm sits** (the exhibit's author writes the map and the questions; seating Claude arms would stack a conflict on a conflict), and **no second local box** — only the-gpu-box serves the local arms, so no cross-box latency comparison is made or implied.

## 3. Runtime pins — which weights answered

*Read from the-gpu-box at `http://the-gpu-box[:port redacted]` BEFORE the first scored call. A pin recorded after a run is a re-resolution, not a pin.*

- ollama **0.32.13** · `num_ctx` 32768 · `think` False · temperature 0.0 · keep_alive never sent (house law 3)

| arm | tag | digest | params | quant |
| --- | --- | --- | --- | --- |
| `local-qwen3.8-27b` | `qwen3.8:27b` | `22130167c4c20e20…` | 27.3B | Q4_K_M |
| `local-qwen3.6-27b` | `qwen3.6:27b` | `a50eda8ed977ab48…` | 27.8B | Q4_K_M |
| `local-gemma4-26b` | `gemma4:26b` | `5571076f3d700504…` | 25.8B | Q4_K_M |
| `local-llama3.3-70b` | `llama3.3:70b` | `a6eb4748fd2990ad…` | 70.6B | Q4_K_M |

Cloud shelf re-probed at registration (19 tags visible on ollama.com): `glm-5.2` PRESENT, `deepseek-v4-pro:preview` PRESENT, `kimi-k3` PRESENT, `gpt-5.5-2026-04-23` PRESENT.

## 4. The two served blocks

Budget measured on ONE reference tokenizer: `gemma4:12b` at `num_ctx` 32768, `num_predict` 1, `think:False`, reading `prompt_eval_count` — the recon census method and the same daemon, so the two reconcile. Chat-template overhead 13 tokens, a constant on both sides.

| block | bytes | tokens | sha256 |
| --- | --- | --- | --- |
| **C-MAP** | 12080 | **3211** | `c74cc74b81e7f75b…` |
| **C-HTML** | 13120 | **3088** | `470746ad8ef5cc0d…` |

Budget band ±0.05: window [3051, 3371] tokens, C-HTML landed 3088 — **in band**.

- C-MAP part `apex/llms.txt` — 2072 B, 584 tok measured alone
- C-MAP part `research/llms.txt` — 10007 B, 2639 tok measured alone

C-HTML: **72** of 6606 available blocks, from 3 of 14 pages in the pinned order. Extractor module sha256 `a4aee74b2a6ba70f…`.

- **byte offset of the cut**: 13120
- **last block included** (`research/exhibits/01-the-open-call.html`, `<li>`): “No leaderboard, and the cards say so — figures print as counts with their denominators, and two figures closer than the registered tie band of 0.5 are within th…”
- **first block excluded** (`research/exhibits/01-the-open-call.html`, `<li>`): “Six reply cards per act, chosen by rule — the same six arms answer every ask in their act, and each act then points at the full set of twenty and names the arms…”

## 5. The answer-presence audit — computed before the first call

| stratum | count in the 20-item headline set | registered minimum | met |
| --- | --- | --- | --- |
| MAP-ONLY | 8 | — (calibration) | — |
| HTML-ONLY | 5 | 3 | yes |
| BOTH | 2 | — (calibration) | — |
| NEITHER | 5 | 4 | yes |
| MAP-STALE | 0 | 1 | **NO** |

Author predictions held on **30/30** items. The bare-key reading of "is the answer present" disagrees with the evidence-span reading on 6 items (`C03`, `C07`, `C11`, `R02`, `R04`, `R06`) — which is exactly the accident short numeric keys were going to cause, and why the stratum is assigned on evidence spans.

### MAP-STALE is VACANT, and that is a finding

- **0 of 36** URLs asserted by the two map files fail to resolve.
- **0 of 11** numeric claims the research map makes are contradicted by the page they describe.

The plan's registered worked example — *PLAN-THIRTEEN registered: llms.txt says the open call sat 'twenty models' while commit b3e096c landed a twenty-first arm, so the page publishes twenty-one.* — **does not hold against the frozen corpus.** DOES NOT HOLD against the frozen corpus. The page prints 'twenty models' too and never prints 'twenty-one': the addendum is disclosed as a twenty-first arm run AFTER the sealed round, and the page states 'the twenty arms ... are frozen exactly as they were'. Map and page agree; the drift is that the map does not MENTION the addendum, which is an omission (HTML-ONLY at best), not a contradiction.

Cause: the estate fix (s2s-web 9631d54, 'llms.txt gets a source of record, estate-wide') repaired the drift this stratum was designed to measure, in the same week and before this bench froze its corpus.

Disposition, per the plan's own registered rule for an unmeetable minimum: the reduction is printed with its cause and the claim is not silently re-scoped. The shortfall is written into the seal manifest itself.

## 6. The checker audition

**335 cases, 0 failures.** Every canonical key and every sealed paraphrase accepted; every sealed decoy rejected; shape cases (empty reply, two candidates, label prefix, abstention in both directions) behaved as registered. Checkers module sha256 `e230cc7211ae698c…`.

Items are stamped `exact` (20) or `proxy` (4) and **the two subtotals are never added together**.

## 7. C-NONE — the contamination meter and the armor gate

Rule: >=2 arms from >=2 different families correct closed-book; single-arm hits are flagged and published, excluded from the headline, and re-printed as a sensitivity row.

- items excluded: **0** []
- single-arm hits (flagged, published, out of the headline): **1** ['C14']
- frozen core excluded: **0** of 14 = **0.0%** against a 40% gate
- **ARMOR HELD**

  - *single-arm* `C14` — ['local-llama3.3-70b']

## 8. The three fences

| condition | arm | cells | transport failures | unparseable | arm median prompt tok | field median | 0.80 floor | verdict |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| C-HTML | `cloud-deepseek-v4-pro` | 24 | 0 | 0 | 2980.5 | 2999.5 | 2399.6 | RANKED |
| C-HTML | `cloud-glm-5.2` | 24 | 0 | 0 | 2972.5 | 2999.5 | 2399.6 | RANKED |
| C-HTML | `cloud-kimi-k3` | 24 | 0 | 0 | 3016.5 | 2999.5 | 2399.6 | RANKED |
| C-HTML | `local-gemma4-26b` | 24 | 0 | 0 | 3143.5 | 2999.5 | 2399.6 | RANKED |
| C-HTML | `local-llama3.3-70b` | 24 | 0 | 0 | 2975.5 | 2999.5 | 2399.6 | RANKED |
| C-HTML | `local-qwen3.6-27b` | 24 | 0 | 0 | 3119.5 | 2999.5 | 2399.6 | RANKED |
| C-HTML | `local-qwen3.8-27b` | 24 | 0 | 0 | 3119.5 | 2999.5 | 2399.6 | RANKED |
| C-HTML | `openai-gpt-5.5` | 24 | 0 | 0 | 2935.5 | 2999.5 | 2399.6 | RANKED |
| C-MAP | `cloud-deepseek-v4-pro` | 24 | 0 | 0 | 2976.5 | 3003.5 | 2402.8 | RANKED |
| C-MAP | `cloud-glm-5.2` | 24 | 0 | 0 | 2958.5 | 3003.5 | 2402.8 | RANKED |
| C-MAP | `cloud-kimi-k3` | 24 | 0 | 0 | 3028.5 | 3003.5 | 2402.8 | RANKED |
| C-MAP | `local-gemma4-26b` | 24 | 0 | 0 | 3266.5 | 3003.5 | 2402.8 | RANKED |
| C-MAP | `local-llama3.3-70b` | 24 | 0 | 0 | 2959.5 | 3003.5 | 2402.8 | RANKED |
| C-MAP | `local-qwen3.6-27b` | 24 | 0 | 0 | 3046.5 | 3003.5 | 2402.8 | RANKED |
| C-MAP | `local-qwen3.8-27b` | 24 | 0 | 0 | 3046.5 | 3003.5 | 2402.8 | RANKED |
| C-MAP | `openai-gpt-5.5` | 24 | 0 | 0 | 2947.5 | 3003.5 | 2402.8 | RANKED |
| C-NONE | `cloud-deepseek-v4-pro` | 24 | 0 | 0 | 73.5 | 82 | 65.6 | RANKED |
| C-NONE | `cloud-glm-5.2` | 24 | 0 | 0 | 74.5 | 82 | 65.6 | RANKED |
| C-NONE | `cloud-kimi-k3` | 24 | 0 | 0 | 143.5 | 82 | 65.6 | RANKED |
| C-NONE | `local-gemma4-26b` | 24 | 0 | 0 | 88.5 | 82 | 65.6 | RANKED |
| C-NONE | `local-llama3.3-70b` | 24 | 0 | 0 | 78.5 | 82 | 65.6 | RANKED |
| C-NONE | `local-qwen3.6-27b` | 24 | 0 | 0 | 82.5 | 82 | 65.6 | RANKED |
| C-NONE | `local-qwen3.8-27b` | 24 | 0 | 0 | 82.5 | 82 | 65.6 | RANKED |
| C-NONE | `openai-gpt-5.5` | 5 | 0 | 0 | 73 | 82 | 65.6 | RANKED |

**No arm tripped any fence in any condition.**

## 9. The headline table — counts, no verdict words

Denominator: **20 items** (the 20-item headline set minus 0 excluded).

| arm | family | class | C-NONE | C-MAP | C-HTML | b (MAP only) | c (HTML only) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `local-qwen3.8-27b` | qwen | local ~27b | 0 | **10** | **6** | 8 | 4 |
| `local-qwen3.6-27b` | qwen | local ~27b | 0 | **10** | **6** | 8 | 4 |
| `local-gemma4-26b` | gemma | local ~26b | 0 | **9** | **7** | 8 | 6 |
| `local-llama3.3-70b` | llama | local ~70b | 1 | **10** | **7** | 8 | 5 |
| `cloud-glm-5.2` | glm | frontier cloud | 0 | **10** | **7** | 8 | 5 |
| `cloud-deepseek-v4-pro` | deepseek | frontier cloud | 0 | **10** | **7** | 8 | 5 |
| `cloud-kimi-k3` | kimi | frontier cloud | 0 | **9** | **7** | 8 | 6 |
| `openai-gpt-5.5` | gpt | frontier cloud | 0 | **10** | **7** | 8 | 5 |

*counts only, no verdict word: this instrument does not resolve a direction at n<=20 and saying so is the honest form of the comparison.*

### Pooled discordance

- b (C-MAP right, C-HTML wrong) = **64**
- c (C-HTML right, C-MAP wrong) = **40**
- discordant pairs = **104** (clears the PERCENT_FLOOR of 30)
- larger side holds **61.5%** against a registered band of 60%
- **reading: C-MAP DIRECTION**

*Basis: the arms share items and are therefore NOT independent, so this figure describes these eight arms, not a population. It is a statement about the instrument's resolution at this N, not a p-value and not a tie between the representations.*

### The extractor's thumb — POST-HOC, not pre-registered

**the registered extractor keeps visible text only, so href attributes are dropped and the C-HTML block carries no URLs; every navigation item is unanswerable from it by construction.** C-MAP carries **50 URL occurrences**, **36 of them distinct**, and **49 markdown link targets**; C-HTML carries **0**. Counting rule: occurrences of `http://` or `https://` in the served block file, and the count of distinct addresses after trimming trailing punctuation — both reproducible from `block-c-map.txt` with one grep.

> **CORRECTED 2026-08-16, after publication.** This line previously read **`C-MAP carries 61 URLs`** — a figure no counting rule was ever attached to, and one that cannot be reproduced from the served block. The retracted **61** is named here rather than deleted. `scores.json` had already retracted it in its own `urls_in_c_map_correction` field before publication; this report shipped with the stale figure and is corrected to match. No scored result depended on it — it sizes the extractor's thumb, and the b/c/pairs figures beside it are unchanged. Note also that `scores.json` records `urls_in_c_html_raw_regex: 2` beside `urls_in_c_html_real: 0`: the 2 is a looser pattern matching a bare host string, and both hits are the same footer email address (`hello@strata2signal.com`), once from each front page. Under the rule stated above, C-HTML carries zero.

Removing the 6 navigation items (`N01`, `N02`, `N03`, `N04`, `N05`, `N06`) leaves 14 items and gives:

- b = **16**, c = **40**, pairs = **56**
- larger side **71.4%** against the same 60% band
- **reading: C-HTML DIRECTION** — the registered direction REVERSES

*POST-HOC — NOT PRE-REGISTERED, published for transparency only. removing six of twenty items is a large intervention and this figure is NOT the bench's result. It is the size of the extractor's thumb, stated so nobody has to estimate it.*

### Sensitivity — single-arm C-NONE hits put back in

| arm | C-MAP | C-HTML | b | c |
| --- | --- | --- | --- | --- |
| `local-qwen3.8-27b` | 9 | 5 | 8 | 4 |
| `local-qwen3.6-27b` | 9 | 5 | 8 | 4 |
| `local-gemma4-26b` | 8 | 6 | 8 | 6 |
| `local-llama3.3-70b` | 9 | 6 | 8 | 5 |
| `cloud-glm-5.2` | 9 | 6 | 8 | 5 |
| `cloud-deepseek-v4-pro` | 9 | 6 | 8 | 5 |
| `cloud-kimi-k3` | 8 | 6 | 8 | 6 |
| `openai-gpt-5.5` | 9 | 6 | 8 | 5 |

Denominator 19 items.

## 10. Per-stratum — published separately, never pooled

*per-stratum results publish separately and are NEVER pooled with each other; a bench that only asks about things the index names has measured the index's own claims.*

### MAP-ONLY — 8 items ['C06', 'C07', 'N01', 'N02', 'N03', 'N04', 'N05', 'N06']

| arm | C-NONE | C-MAP | C-HTML |
| --- | --- | --- | --- |
| `local-qwen3.8-27b` | 0 | **8** | **0** |
| `local-qwen3.6-27b` | 0 | **8** | **0** |
| `local-gemma4-26b` | 0 | **8** | **0** |
| `local-llama3.3-70b` | 0 | **8** | **0** |
| `cloud-glm-5.2` | 0 | **8** | **0** |
| `cloud-deepseek-v4-pro` | 0 | **8** | **0** |
| `cloud-kimi-k3` | 0 | **8** | **0** |
| `openai-gpt-5.5` | 0 | **8** | **0** |

### HTML-ONLY — 5 items ['C01', 'C02', 'C03', 'C04', 'C05']

| arm | C-NONE | C-MAP | C-HTML |
| --- | --- | --- | --- |
| `local-qwen3.8-27b` | 0 | **0** | **4** |
| `local-qwen3.6-27b` | 0 | **0** | **4** |
| `local-gemma4-26b` | 0 | **0** | **5** |
| `local-llama3.3-70b` | 0 | **0** | **5** |
| `cloud-glm-5.2` | 0 | **0** | **5** |
| `cloud-deepseek-v4-pro` | 0 | **0** | **5** |
| `cloud-kimi-k3` | 0 | **0** | **5** |
| `openai-gpt-5.5` | 0 | **0** | **5** |

### BOTH — 2 items ['C13', 'C14']

| arm | C-NONE | C-MAP | C-HTML |
| --- | --- | --- | --- |
| `local-qwen3.8-27b` | 0 | **2** | **2** |
| `local-qwen3.6-27b` | 0 | **2** | **2** |
| `local-gemma4-26b` | 0 | **1** | **2** |
| `local-llama3.3-70b` | 1 | **2** | **2** |
| `cloud-glm-5.2` | 0 | **2** | **2** |
| `cloud-deepseek-v4-pro` | 0 | **2** | **2** |
| `cloud-kimi-k3` | 0 | **1** | **2** |
| `openai-gpt-5.5` | 0 | **2** | **2** |

### NEITHER — 5 items ['C08', 'C09', 'C10', 'C11', 'C12']

| arm | C-NONE | C-MAP | C-HTML |
| --- | --- | --- | --- |
| `local-qwen3.8-27b` | 0 | **0** | **0** |
| `local-qwen3.6-27b` | 0 | **0** | **0** |
| `local-gemma4-26b` | 0 | **0** | **0** |
| `local-llama3.3-70b` | 0 | **0** | **0** |
| `cloud-glm-5.2` | 0 | **0** | **0** |
| `cloud-deepseek-v4-pro` | 0 | **0** | **0** |
| `cloud-kimi-k3` | 0 | **0** | **0** |
| `openai-gpt-5.5` | 0 | **0** | **0** |

### MAP-STALE — **VACANT**

no map/page contradiction exists in the frozen corpus — see prereg/receipts/map-stale-hunt.json.

## 11. Refusal bait — honesty, stamped `proxy`

*never averaged into accuracy; interpretable only in C-MAP and C-HTML. Items: ['B01', 'B02', 'B03', 'B04'].*

| arm | C-MAP | C-HTML | C-NONE |
| --- | --- | --- | --- |
| `local-qwen3.8-27b` | 4/4 | 4/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |
| `local-qwen3.6-27b` | 4/4 | 4/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |
| `local-gemma4-26b` | 4/4 | 4/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |
| `local-llama3.3-70b` | 4/4 | 4/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |
| `cloud-glm-5.2` | 4/4 | 4/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |
| `cloud-deepseek-v4-pro` | 4/4 | 3/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |
| `cloud-kimi-k3` | 4/4 | 4/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |
| `openai-gpt-5.5` | 4/4 | 4/4 | — (excluded from the honesty figure: with no site content, "the site doesn't say" is correct for every question) |

## 12. The flake probe

*3 draws on 6 items (['C01', 'C05', 'C08', 'C11', 'N03', 'B03']), C-MAP only, zero-cost arms.*

| arm | items | items that flipped | instability flag |
| --- | --- | --- | --- |
| `local-qwen3.8-27b` | 6 | 0 | stable |
| `local-qwen3.6-27b` | 6 | 0 | stable |
| `local-gemma4-26b` | 6 | 0 | stable |
| `local-llama3.3-70b` | 6 | 0 | stable |
| `cloud-glm-5.2` | 6 | 0 | stable |
| `cloud-deepseek-v4-pro` | 6 | 0 | stable |
| `cloud-kimi-k3` | — | — | — no flake figure: metered arm, the reason is cost |
| `openai-gpt-5.5` | — | — | — no flake figure: metered arm, the reason is cost |

## 13. The bill and the clock

- leg `C-NONE`: total **$0.4791** of $4.00 · per arm {'cloud-kimi-k3': 0.015, 'openai-gpt-5.5': 0.4642} · halted ['openai-gpt-5.5'] · Nones wall
- leg `C-MAP,C-HTML`: total **$1.3381** of $4.00 · per arm {'cloud-kimi-k3': 0.4644, 'openai-gpt-5.5': 0.8737} · halted none · 499.5s wall
- leg `C-MAP (flake probe)`: total **$0.0000** of $4.00 · per arm {} · halted none · 104.3s wall


**Grand total: $1.8688 of $4.00** (47% of the ceiling) — $1.8172 across the legs plus $0.0516 for the pre-seal metered probes.

| metered arm | spent, all legs | sub-cap | respected |
| --- | --- | --- | --- |
| `cloud-kimi-k3` | $0.4794 | $0.90 | yes |
| `openai-gpt-5.5` | $1.3379 | $2.60 | yes |

*Priced at the UNCACHED rates — an upper bound by construction (house rule).*

Pre-leg metered output probes (precondition 9):

| arm | median output tok | registered limit | projected leg $ | sub-cap | inside |
| --- | --- | --- | --- | --- | --- |
| `cloud-kimi-k3` | 15.0 | 900 | $0.67 | $0.9 | yes |
| `openai-gpt-5.5` | 58.0 | 900 | $1.186 | $2.6 | yes |

## 14. The estate probe — the confession's counts, dated

*As of `2026-08-16T14:41:23Z`. Classified on CONTENT, not status code: one host answers 200 with its HTML shell for every path.*

**14 serving · 1 phantom · 1 absent**, of 16 probed. 11 of the serving hosts byte-match their source tree.

| host | status | bytes | state | repo |
| --- | --- | --- | --- | --- |
| `strata2signal.com` | 200 | 2072 | serving | match |
| `research.strata2signal.com` | 200 | 10007 | serving | match |
| `rulesage.strata2signal.com` | 200 | 1748 | serving | no repo copy |
| `realkeep.strata2signal.com` | 200 | 2115 | serving | no repo copy |
| `amble.strata2signal.com` | 200 | 2172 | serving | no repo copy |
| `design.strata2signal.com` | 200 | 1489 | serving | match |
| `sharpsignal.strata2signal.com` | 200 | 1847 | serving | match |
| `kiln.strata2signal.com` | 200 | 1826 | serving | match |
| `memory.strata2signal.com` | 200 | 1862 | serving | match |
| `agency.strata2signal.com` | 200 | 1946 | serving | match |
| `easel.strata2signal.com` | 200 | 1789 | serving | match |
| `brain.strata2signal.com` | 200 | 1887 | serving | match |
| `keel.strata2signal.com` | 200 | 1729 | serving | match |
| `forge.strata2signal.com` | 200 | 1841 | serving | match |
| `play.strata2signal.com` | 200 | 0 | **phantom** | no repo copy |
| `asr.strata2signal.com` | 404 | 0 | **absent** | no repo copy |

## 15. The pause rule

- leg `C-NONE`: baseline `2026-08-16T14:49:xxZ (see results/c-none.log)`, None checks, probes = ['estate-alert (alerts newer than baseline)', 'the-gpu-box /api/ps (a resident that is not one of ours)', 'rulesage liveness']
- leg `C-MAP,C-HTML`: baseline `2026-08-16T14:58:55Z`, 384 checks, probes = ['estate-alert (alerts newer than baseline)', 'the-gpu-box /api/ps (a resident that is not one of ours)', 'rulesage liveness']
- leg `C-MAP (flake probe)`: baseline `2026-08-16T15:09:42Z`, 72 checks, probes = ['estate-alert (alerts newer than baseline)', 'the-gpu-box /api/ps (a resident that is not one of ours)', 'rulesage liveness']

**Mint-queue probe: UNAVAILABLE.** no mint-queue endpoint is reachable from this box: the cove serves from the-vps and exposes no queue-depth path. Recorded as a gap rather than silently dropped or invented.

The nine alerts standing on this box at run time (oldest 2026-08-02) are pre-existing and none are the-gpu-box's; the rule gates on alerts raised AFTER the run's baseline epoch, which is the only reading that distinguishes a backlog from an event.

---

*Counting rules, checkers, keys, variants, the audition table, the answer-presence audit and every reply verbatim are in the kit. CC BY 4.0.*