# Reproduce the two turns

Every live number in *The engine, on the record* came from one of these reads. They need no key and no
account. Capture date: **2026-10-18, 20:11–20:13 MDT** (2026-10-19, 02:11–02:13 UTC).

The engine restarts. During the capture window the origin returned **502** twice, and a read taken 30
seconds later was 200 with the same store counts (the status `uptimeMs` was 74 s). The store numbers
are stable across a restart; the uptime is not. Retry a 502 before treating it as a finding.

## 0. When you ran it

Write down the wall-clock time and the `date` header of the first response. A live read without a time
is not a receipt.

```bash
date -u +%Y-%m-%dT%H:%M:%SZ
```

## 1. The status read

```bash
curl -s https://metatronmind.com/api/system/status | tee public-status.json | jq '{
  genesisAt, storeAgeMs, learningMs, uptimeMs, reading,
  dictionarySize, specialistCount, totalForwardPasses,
  activity, capabilities, calibration
}'
```

Expected at capture:

```json
{
  "genesisAt": "2026-09-28T22:11:53.889Z",
  "storeAgeMs": 1742359117,
  "learningMs": 1742359117,
  "uptimeMs": 78858,
  "reading": "at-rest",
  "dictionarySize": 145565,
  "specialistCount": 21,
  "totalForwardPasses": 145750,
  "activity": { "trainer": "stopped", "concern": null, "tick": 0 },
  "capabilities": { "ingest": false },
  "calibration": {
    "at": "2026-10-13T04:44:54.859Z",
    "dictSize": 145179,
    "suggestedDecline": 0.74,
    "currentDecline": 0.76,
    "youdenJ": 0.85,
    "falseDecline": { "n": 0, "of": 27 },
    "ungroundedCatch": { "n": 51, "of": 60 },
    "smearCatch": { "n": 11, "of": 12 }
  }
}
```

What to check:

- `learningMs == storeAgeMs`. That is the legacy alias. The status handler's comment says *"Nothing in
  the engine 'learns'; do not build on it."*
- `dictionarySize` is the roster total (generalist + specialists), **not** including off-roster cubes.
- `capabilities.ingest` is `false` for an anonymous remote caller: the public box will not take a
  document from you.
- The `about` block is the set of facts a public page may use. It includes `writer`, `decline`,
  `ladder`, `rereads`, `memory`, `xPosts`, `screening`. Read it directly with
  `jq '.about' public-status.json`.

## 2. The store's shape

```bash
curl -s https://metatronmind.com/api/cubes/topology | tee cubes-topology.json | jq '{
  totalDictionary, generalistDictionary, totalCubes, specialistCount: ([.nodes[].specialists[]] | length)
}'
```

Expected at capture: `totalDictionary: 145565`, `generalistDictionary: 760`, `totalCubes: 22`,
`specialistCount: 21`.

The check the article makes:

```bash
jq '[.nodes[].specialists[].dictionarySize] | add' cubes-topology.json   # 144805
jq '.generalistDictionary' cubes-topology.json                            # 760
# 144805 + 760 = 145565, the public dictionarySize
```

Off-roster cubes are not in that sum. The registry read shows them:

```bash
curl -s https://metatronmind.com/api/cubes/registry | jq '[.offRoster[] | {id, role, dictionarySize}]'
# archaeology-docs 10938, resident-memory 107
```

145,565 + 10,938 + 107 = **156,610** — the number the public count deliberately is not, because chat
cannot reach those two cubes.

## 3. The answered turn

```bash
curl -s -X POST https://metatronmind.com/api/chat \
  -H 'content-type: application/json' \
  -H 'origin: https://metatronmind.com' \
  -d '{"message":"What causes sodium channels to open during a nerve impulse?"}' \
  | tee /tmp/chat-covered.full.json \
  | jq '{mode, finalCubeId, routedCubeIds, cascadeDepth,
         attribution, "writerAsked": .ladder.writer.asked,
         "top5": [.blendedCascade[:5][] | {cubeId, sourceId, querySim}],
         "cascadeRange": ([.blendedCascade[].querySim] | {min: min, max: max})}'
```

Expected at capture: `mode: "specialist:llm"`, `finalCubeId:
"specialist-medicalphysiology-musrga3w"`, `attribution.verdict: "covered"`,
`attribution.top: 0.828`, `attribution.threshold: 0.76`, `writerAsked: true`, every `blendedCascade`
entry from one document, range **0.7833–0.8278**. Wall-clock: **10.3 s**.

## 4. The refused turn

```bash
curl -s -X POST https://metatronmind.com/api/chat \
  -H 'content-type: application/json' \
  -H 'origin: https://metatronmind.com' \
  -d '{"message":"What did the Zorblat-9 mission find on Europa in 2031?"}' \
  | tee /tmp/chat-declined.full.json \
  | jq '{mode, finalCubeId, routedCubeIds, response,
         ladder: {verdict: .ladder.verdict, rung: .ladder.rung, top: .ladder.top,
                  threshold: .ladder.threshold, search: .ladder.search,
                  writerAsked: .ladder.writer.asked}}'
```

Expected at capture: `mode: "decline"`, `finalCubeId: "generalist"`, `routedCubeIds: []`,
`ladder.verdict: "not-covered"`, `ladder.rung: "none"`, `ladder.top: 0.717`,
`ladder.threshold: 0.76`, `ladder.search: {ran: true, returned: 8, floor: 0.8, best: 0.748,
offTopic: true}`, `ladder.writer.asked: false`. Wall-clock: **1.7 s**.

The refusal sentence is `response`; the same sentence is `attribution.sentence`. The engine wrote it;
the writer was not asked.

## 5. The two rules the audit depends on

**What the shipped copies are.** `chat-covered.json` and `chat-declined.json` in this pack are the
decision fields of the two responses — mode, route, scores, threshold, verdict, rung, attribution and
the ranked ids — with the retrieved passage text removed. The commands above return the full response
live, passages included. The store's documents are third-party sources, so the pack records the
decision and the numbers, not a copy of what was read; the live call is one command away for anyone who
needs to see the passages themselves. The trim changed no score, id or verdict.

**`writer.asked` is the line between a computed refusal and a styled one.** If it is `true`, the model
had the turn and could have said "I don't know" in its own words. The engine's claim is stronger:
under the ladder, when nothing clears and the web does not, no model is called.

**A web search is not a web answer.** `returned` is how many results came back; `best` is the highest
similarity of any of them to the question; `floor` is what one had to reach to count. At capture,
eight results came back and the best scored 0.748 against a floor of 0.800, so the search counted as
nothing. If a system reports only "searched the web", it has reported nothing.

## 6. The code, not the prose

The pack ships `grounding-ladder.ts`. The rest of the map is in `architecture-2026-10-18.md` §7. The
three constants to read for yourself:

- `GROUNDING_DECLINE_THRESHOLD` — 0.750 default, 0.760 in production (`server/chat/retrieve.ts`).
- `GROUNDING_RELATED_MARGIN` — 0.010 default; capped at 0.05 so a wide band cannot become a lower
  threshold by another name.
- `GROUNDING_LADDER_WEB_FLOOR` — 0.800 default.

## 7. What is not reproducible from here

- The ingestion path (the validator, the work ledger) is operator-only and not exposed to an anonymous
  caller. The article says so; it does not claim a visitor can add a document.
- The **75–160 ms** retrieval measurement was taken on the production box and published with *A
  Forward Pass on the Graph*; it is not in this pack.
- The **Epoch 1** ablation numbers (0 of 33 edges lockable, state erased in ~5 passes, 13.94 min/h of
  engine-write) are published with *Eight subsystems, deleted*, which ships the boundary document and
  the constraints ledger.
- The store is live. The counts in this file are a snapshot at the timestamp above; the durable facts
  are the genesis date, the architecture, and the two response shapes.