# Persistent stack — first end-to-end coordinator query

First real-shape demand query against the long-running Go substrate
(daemons up for minutes, not harness-transient). Same shape of
test as `real_001..real_005` reality probes, but those all ran
through `playbook_lift.sh`'s spin-up-and-tear-down cycle. This is
production-shape: stack stays up, query lands, top-K returns,
operator inspects.

## Setup

- Persistent stack: 11 daemons via `scripts/cutover/start_go_stack.sh`
- Workers ingested: 500 rows from `workers_500k.parquet` (real production
  data, not synthetic)
- Index: `workers` corpus on vectord
- Embed: nomic-embed-text-v2-moe (768d)
- Distance: cosine

## Three queries

### Q1: Forklift Operators in Aurora IL

```
Need 3 Forklift Operators in Aurora IL starting at 09:00 for Parallel Machining
```

| rank | id | dist | name | role | loc |
|---|---|---:|---|---|---|
| 0 | w-325 | 0.441 | Angela Ruiz | Forklift Operator | Champaign, IL |
| 1 | w-358 | 0.464 | Lauren Evans | Forklift Operator | Danville, IL |
| 2 | w-447 | 0.465 | Raymond Alvarez | Forklift Operator | Bloomington, IL |
| 3 | w-153 | 0.490 | Sandra Phillips | Forklift Operator | Cleveland, OH |
| 4 | w-108 | 0.520 | Omar Cook | Forklift Operator | South Bend, IN |

**Read:** 5/5 role match. Top 3 in IL (semantic neighbors of Aurora
— Champaign / Danville / Bloomington are central-southern IL cities).
Next 2 are OH/IN — still Midwest, reasonable geographic falloff.
A coordinator looking at this would probably take ranks 0-2.

### Q2: CNC Operator in Detroit MI

```
Need 1 CNC Operator in Detroit MI for Beacon Freight
```

| rank | id | dist | name | role | loc |
|---|---|---:|---|---|---|
| 0 | w-175 | 0.492 | Kevin Ruiz | Machine Operator | Detroit, MI |
| 1 | w-413 | 0.521 | Daniel Hall | Machine Operator | Grand Rapids, MI |
| 2 | w-450 | 0.524 | Shirley Evans | Machine Operator | Cincinnati, OH |
| 3 | w-105 | 0.603 | Donna Sanchez | Machine Operator | Memphis, TN |
| 4 | w-162 | 0.615 | Matthew Sanchez | Machine Operator | Rockford, IL |

**Read:** Geo match perfect (top-1 in Detroit MI exactly). Role
match is "Machine Operator" not "CNC Operator" — the 500-row
sample probably has no exact-CNC-Operator role tag, so the
embedder pulls Machine Operator as the closest semantic match.
A coordinator might say "yes, Kevin's done CNC before, he counts"
or might say "no, machine operator is too generic." This is
exactly the kind of borderline case the playbook learning loop
captures: judge rates this ~3/5; first time a coordinator
approves a Machine Operator for a CNC Operator query, that
recording starts shifting the substrate's behavior.

### Q3: Warehouse Associates in Indianapolis IN

```
Need 5 Warehouse Associates in Indianapolis IN at 12:00 for Midway Distribution
```

| rank | id | dist | name | role | loc |
|---|---|---:|---|---|---|
| 0 | w-390 | 0.419 | Mary Hughes | Warehouse Associate | Indianapolis, IN |
| 1 | w-118 | 0.467 | Patricia Allen | Warehouse Associate | South Bend, IN |
| 2 | w-280 | 0.481 | Ruth Torres | Warehouse Associate | Chicago, IL |
| 3 | w-149 | 0.485 | Jennifer Nguyen | Warehouse Associate | Des Moines, IA |
| 4 | w-151 | 0.544 | Eric Diaz | Warehouse Associate | Columbus, OH |

**Read:** Cleanest of the three. 5/5 role match, top-1 in
Indianapolis IN exactly, geographic falloff (IN → IL → IA → OH)
makes operational sense — coordinator could expand the search
radius if Mary Hughes isn't available.

## What this confirms

1. **The substrate is producing useful answers on real coordinator
   queries against real production data on the persistent stack.**
   Not synthetic, not transient.
2. **Role-semantic neighbors fire as designed**: when an exact role
   match isn't in the corpus, the embedder pulls semantic neighbors
   (Q2). Operator-visible fact, not silent corruption.
3. **Geo is a load-bearing feature**: top-1 city/state matched in
   3/3 queries — the embedder treats geography as primary.
4. **Distance distribution is tight**: 0.41-0.49 across top-1 of
   all three queries. Healthy embedder + healthy index.

## What this does NOT prove

- Real coordinator demand traffic at production volumes (this was 3
  queries against 500 workers; production is potentially 5K-500K
  workers + dozens of concurrent queries).
- Latency under load (single-query latency was sub-second; concurrent
  load not measured).
- The 5-loop substrate end-to-end (this exercised retrieval only;
  observer + pathway + matrix injection + judge gate didn't fire
  because no playbook recordings exist on the persistent stack yet).

## Repro

```bash
./scripts/cutover/start_go_stack.sh
./bin/staffing_workers -limit 500 -gateway http://127.0.0.1:3110 -drop=true

curl -sS -m 30 -X POST http://127.0.0.1:3110/v1/matrix/search \
  -H 'content-type: application/json' \
  -d '{"query_text":"Need 3 Forklift Operators in Aurora IL starting at 09:00 for Parallel Machining","corpora":["workers"],"k":5,"per_corpus_k":5,"use_playbook":false}'
```

## State

- Persistent stack: still up (all 11 healthy as of 2026-05-01T03:10ish CDT)
- Workers index: populated with 500 rows
- Logs: `/tmp/gostack-logs/<bin>.log`
- Will be torn down on the next `git push` (smokes vs persistent
  pkill conflict — see header of `start_go_stack.sh`)