# Production census — final state

**Captured:** 2026-07-31 06:50–06:55 UTC, from the live production database, immediately before
wind-down. Every figure below is a direct query result, not a derived or remembered value.

## The directory (ACRA registry mirror)

| Metric | Value |
|---|---|
| Entities | **617,987** |
| Former-name aliases | 101,501 |
| Entities carrying at least one former name | 78,931 |
| Distinct SSIC industry codes | 1,021 |
| Oldest incorporation date | **1887-07-13** |
| Newest incorporation date | **2026-06-30** |
| Incorporated in the trailing 365 days | 69,717 |
| ACRA letter-files synced | 27 of 27 |

**Caveat on 617,987.** The ingest upserts with `ON CONFLICT DO UPDATE` and has no delete pass, so an
entity that ceased to be Live after its letter was last synced was never removed. The figure is
therefore an upper bound that drifts upward over time, and it reflects an as-of window of
2026-07-14 to 2026-07-31 rather than a single instant. **Newest incorporation date is 2026-06-30
against a capture date of 2026-07-31** — roughly one month stale, which is the floor rather than a
lag to be fixed, because ACRA republishes the bulk letter-files about monthly.

## The canonical store (the dedup lab)

| Metric | Value |
|---|---|
| Companies | 5,030 |
| Registry identifiers (UENs) | 5,016 |
| Deals | 18 |
| Audit-log rows | 10,128 |
| Duplicate candidate pairs detected | 58 |
| Review-queue rows | 0 |
| Published entities (`publish_state`) | **0** |
| Source runs logged | 21 |

Nothing was ever published. `publish_state` is empty because no band earned the right to publish.
The publish job ran nightly and did nothing, every night, which is what a fail-closed gate is for.

## The oracle — measured precision

Deterministic-anchor oracle, easy cases only, over detector-flagged pairs:

| Band | Precision | Wilson LB | Decisive (tp/fp) | Silent |
|---|---|---|---|---|
| fuzzy | 0.000 | 0.000 | 52 (**0 / 52**) | 2 |
| normalized_name | 0.000 | 0.000 | 1 (**0 / 1**) | 3 |

**Overall: precision 0.000 over 53 anchor-decisive flagged pairs. Zero true positives, 53 false
positives.** Five more sit in the fuzzy tail with no anchor to check them against, and are reported
separately rather than folded into the number.

Gate state at capture (`band_gate_state`, decided 2026-07-31 06:40 UTC):

- `fuzzy` — **closed**, reason `drift`, n_decisive 52
- `normalized_name` — **closed**, reason `unmeasured`, n_decisive 1

Every flagged pair that could be checked against a registry anchor turned out to be two different
companies. Nothing was broken. A registry issues one UEN per entity, so it holds no duplicates of
itself, and a duplicate detector aimed at one has nothing true left to find.

## The council

| Metric | Value |
|---|---|
| Verdict rows | 174 |
| Distinct pairs judged | 58 |
| Model families | 2 (anthropic 116 rows, deepseek 58 rows) |
| Abstains | 10 |
| Merges executed | **0** |

Per-role breakdown:

| Role | merge | keep_apart | abstain |
|---|---|---|---|
| proposer (claude-sonnet-4-6) | 4 | 53 | 1 |
| skeptic (deepseek-v4-pro) | **0** | 50 | 8 |
| judge (claude-sonnet-4-6) | 4 | 53 | 1 |

**On all four pairs where the council voted to merge, the skeptic abstained with
`MemberOutputError`** — output the parser couldn't read. Proposer and judge are the same model
family; the independent second family never returned a readable verdict on any pair that mattered.
The gate held and no merge happened, but it held on a crash rather than on an argument. Full
deliberations in `council-transcripts.json`.

## Host

| Detail | Value |
|---|---|
| Provider | DigitalOcean, SGP1 |
| Size | 2 vCPU / 2 GB / 58 GB disk |
| Uptime at capture | **31 days, 12 hours** |
| Memory in use | 1,043 MB of 1,967 MB |
| Disk in use | 11 GB of 58 GB (19%) |
| Load average | 0.41 / 0.17 / 0.17 |
| Containers | `app-web` up 2 weeks (healthy) · `app-scheduler` up 2 weeks (healthy) · `db` up 4 weeks (healthy) |

Every scheduler job reported a recent success at capture, including `directory_ingest`, which ran at
03:30 UTC on the morning of the capture.