jorgecuadros-platform/docs/POLICY_OCR.md

# Insurance Policy OCR Capture

Reads an insurance policy PDF the office downloads from a carrier portal,
proposes the `Policy` row it should become, and lets staff confirm. Built
2026-08-01 (`5e9cb12`), live under `/polizas/captura`.

## Why this exists — it was not planned

This feature is **not in any spec**. It came out of building the utility
statement OCR intake in [`RECEIPT_CAPTURE_SPEC.md`](RECEIPT_CAPTURE_SPEC.md)
§2: once there was a working render → OCR → parse → match → review pipeline
for CFE/CESPT/Telnor receipts, it was obvious the same shape applies to the
*other* stack of paper this office keys in by hand every week — the carrier
policy PDFs behind every `Policy` row.

The two are the same job with a different document on the scanner. Keeping
that recognition cheap is the whole point of how it was built: the pipeline
was **reused, not copied**.

- `OcrModule` (`apps/api/src/ocr/ocr.module.ts`) was extracted out of
  `StatementsModule` in this same commit, purely so `PolicyOcrModule` can
  inject `OCR_PROVIDER` without dragging in the statement pipeline.
  `StatementsModule` now imports it and binds nothing itself. That extraction
  was **blocking**: without it the policy module could not resolve the
  provider at all.
- The engine stays Tesseract behind the same swappable seam, so a managed
  extraction API remains a one-line change in one file for both features.
- The intake screen is a *mode of* the existing policy-creation screen, the
  same way OCR receipt capture is a mode of Captura — not a new menu entry.

## What ships

| Piece | Path |
|---|---|
| API module | `apps/api/src/policy-ocr/` (service, controller, DTOs, matcher, parser) |
| Shared OCR seam | `apps/api/src/ocr/ocr.module.ts` |
| Tables | `policy_ocr_batches`, `policy_ocr_documents` (`20260801000000_policy_ocr_intake`, extended by `20260815120000_policy_ocr_ana`) |
| Web | `components/PolicyCaptura.tsx` (tab shell), `PolicyOcrIntake.tsx` (upload), `PolicyOcrReview.tsx` (review queue) |
| Abilities | `policy:ingest`, `policy:ocr-review` — both **STAFF** |

Abilities are STAFF for the same reason statement OCR is: nothing reaches the
books unconfirmed, and the review step is what makes machine capture safe at
that tier.

## The screen

`PolicyCaptura` is one screen with two ways in, mirroring `Captura.tsx`:

- `/polizas/nuevo` → **manual** tab (`PolicyForm`, every field by hand)
- `/polizas/captura` → **automática** tab (`PolicyOcrIntake`, drop a PDF)
- `/polizas/captura/[id]` → the batch review queue

Both modes end at the same place — a `Policy` row on a customer's file — so
they are modes of one screen rather than two menu entries. Either URL renders
the same component, so the tab toggle works from either entry point and old
bookmarks land on the right tab.

## Pipeline

```
upload PDF → store source → render pages → text layer? → parse → match → review → confirm
```

1. **Store the source.** `policy-ocr/{batchId}/source-N.pdf`, before anything
   else touches it.
2. **Render + read.** Every page is rendered to
   `policy-ocr/{batchId}/page-M.png`. Text-layer wins when the PDF has one
   (cheap, exact); the rendered image is OCR'd only when it does not — the
   same precedence rule as the statement pipeline. Carrier-portal PDFs are
   usually born-digital, so most of the time no OCR runs at all.
3. **Parse.** Provider detected by brand signal first
   (`GMX`, `Grupo Mexicano de Seguros`, `gmx.com.mx`,
   `JUNTOS EL RIESGO ES MENOR`), layout patterns only as fallback — the same
   ordering rule the statement parser needed.
4. **Match.** Against `Policy.policyNumber`.
5. **Review + confirm.** Nothing is written to `Policy` until a human
   confirms.

### One PDF = one policy

This is the sharpest difference from statement OCR, and it inverts that
feature's core assumption.

Utility statements arrive **bundled, one customer per page** — so there, one
page is one document and the parser runs per page. A policy PDF is the
opposite: the GMX certificate is a 2-page document where page 1 carries the
contract header and page 2 carries the per-coverage table (and the PVL
especificación runs to ten), and **every page describes the same policy**. So
the pipeline concatenates every page's text
(`\n\n` between pages, which also keeps `ocrRawText` readable for debugging)
and runs the parser and the matcher exactly **once per file**.

Consequences worth knowing before touching this code:

- `PolicyOcrDocument.pageNumber` is repurposed as the **file ordinal within
  the batch** (1, 2, 3…), not a page index. The
  `(batchId, pageNumber)` unique constraint still holds, and one batch still
  carries many policies — one per uploaded file.
- Parser regexes are anchored across the whole concatenated text (`^From$`,
  `^Currency\s+…`), which is why the page-boundary blank line matters.
- `ocrConfidence` on the row is the **mean** across the file's pages.
- A file that fails to parse produces exactly one `OCR_FAILED` row — the right
  granularity, and the page PNGs stay on disk for a re-run after a parser fix.

### `storageKey` is the source PDF, not a page image

`PolicyOcrDocument.storageKey` points at `source-N.pdf`. The review screen
embeds that file directly, so the reviewer looks at the **exact artifact the
office received** and gets the browser's native PDF scrolling, zoom and text
selection for free. Rendered PNGs are still written for future re-OCR or an
image-based audit, but nothing points at them as the document's identity.

(The statement side does the opposite — there `storageKey` is the page image,
because a page *is* the document.)

## Matching: policy number only, never the insured name

`PolicyMatcherService` matches on `Policy.policyNumber` and nothing else.

The certificate's "Insured" line is the account's registrant, which drifts
from the customer the office actually holds the file under — the same finding
the statement matcher is built around (a CESPT receipt reading
`ARNAIZ ROSAS ELSA AURORA` for a customer this office holds as `CATT, RANDY`).
Names are shown to the reviewer as a sanity check and never feed matching.

| Rows on `policyNumber` | Result |
|---|---|
| exactly 1 | `MATCHED`, confident — the only unambiguous hit |
| 0 | new policy: review offers a customer picker, confirm **creates** the row |
| >1 | surfaced as candidates, human picks |

More than one hit is never auto-resolved. Duplicate policy numbers across
customers do occur (one group policy bound by two related parties), and
picking arbitrarily would silently book the wrong coverage against the wrong
person.

## GMX ships two unrelated documents for the same policy

The office downloads both from the same portal, and either can land in a
batch. They share only the brand and the policy number, so `parseGmx` is a
two-line dispatcher over two real parsers — both returning `provider: "GMX"`,
because the matcher keys on the policy number alone and must not care which
artifact was uploaded.

| | **Caratula** (`…_Traduccion.pdf`) | **Especificación** (`…-CondicionesParticulares.pdf`) |
|---|---|---|
| Language | English (free translation) | Spanish |
| Shape | boxed header table + 4-column coverage table | 10 pages of prose, no tables at all |
| Header fields | Policy / Insured / Broker / Term / From / To / Currency | insured name, risk location, property description |
| Dates, broker, currency field | yes | **none printed** |
| Coverages | one row per risk | section heading + `Límite Máximo de Responsabilidad:` |
| Parser | `parseGmxCaratula` | `parseGmxEspecificacion` |

Selected by `isEspecificacion` on the PVL page header
(`ESPECIFICACIÓN QUE SE ADHIERE`, `PVL Hogar`, `Nombre del asegurado`).

Three things about the especificación are worth knowing before touching it:

- **The policy number's group widths differ between the two.** The caratula
  reads `007-037-07005947-0000-02` and the especificación
  `07-037-07006957-00000-01` — 2 digits in the first group, 5 in the fourth.
  The original parser pinned the widths, so it read one family and returned
  null on the other. `POLICY_NUMBER_SHAPE` now matches the shape, and since
  the especificación prints the number on all ten page headers, the ten
  readings cross-check each other (disagreement is noted, not resolved — the
  same rule the zona federal parser applies to its clave).
- **Coverages are found by anchoring on the limit label and walking backwards
  for the heading.** There is no row shape to match. A heading is a short line
  *preceded by a blank line* — that last condition is the whole trick, since
  length alone cannot tell a heading from the wrapped tail of the paragraph
  above it (`efectuados.`, `Y CADA PÉRDIDA.`), and without it coverages get
  named after the last word of the preceding prose.
- **Vigencia, agente and prima are absent by design**, not unread. The parser
  says so in a note, so a reviewer seeing three empty fields does not read it
  as a broken parse.

  **These three are keyed in by hand** — confirmed 2026-08-14 with Luz, who
  handles GMX policies at the office. The review screen already has editable
  inputs for all three, and `postPremium` enables off the *typed* premium, so
  a hand-entered prima posts to the ledger exactly like a parsed one. No code
  change was needed to support this; it is a process decision, recorded here
  because the parser's own note now instructs the reviewer accordingly.

  > **A blank vigencia is silently permanent.** `Policy.policyTo` is nullable
  > and the renewals window query filters `policyTo: { gte, lte }`
  > (`renewals.service.ts`), so a policy confirmed without one **never matches
  > and never gets a renewal notice** — no error, no warning, and nothing
  > later notices. This is why the parser's note names the consequence instead
  > of just listing the missing fields.

An excluded catastrophic risk is recorded as excluded **in the risk label**
(`Terremoto o erupción volcánica — Sección Edificio: EXCLUIDO`) with a null
amount, never as `0`: a coverage insured for zero and an excluded coverage are
the same number and very different facts, and `ParsedCoverage` has no field
for the distinction.

## A.N.A. ships two unrelated faces too

`A.N.A. Compañía de Seguros` is the Rosarito office's tourist auto book. Same
split as GMX, different reason: GMX ships two *documents about one policy*,
A.N.A. ships two *products*.

| | **AUTOMOBILE** (`SPECIAL POLICY FOR TOURISTS`) | **DRIVER´S POLICY** (the office says *licencia*) |
|---|---|---|
| Insures | a specific car | up to five named drivers, whatever they drive |
| Vehicle table | `ITEM / YEAR / MAKE / BODY / SERIAL No. / PLATES` | **none** |
| Insured | one `INSURED` cell | numbered `POLICY HOLDER` list |
| Value columns | one (`LIMIT OF LIABILITY`) | two (`SUM INSURED`, `PREMIUM`) |
| Sections | 9, numbered | 6, unnumbered, **in a different order** |
| Parser | `parseAnaAutomobile` | `parseAnaDriverPolicy` |

Selected by `isAnaDriverPolicy` on the title band.

The **four automobile products** the office sells — amplia and responsabilidad
civil, each annual or by-the-day — are the **same layout with different
numbers**. "Amplia" prints a vehicle value and `COVERED` on sections 1–2;
"resp. civil" prints `0.00` and `EXCLUDED`. That is data, not a layout, so
there is one parser rather than four.

Things worth knowing before touching the ANA parsers:

- **These are born-digital portal PDFs**, so `pdftotext -layout` returns exact
  glyphs and exact columns. The driver's policy parser uses that: `SUM
  INSURED` and `PREMIUM` print the same shape (`100,000.00 usd.` /
  `18.70 usd.`) with no per-row label, so **horizontal position is the only
  thing that separates them**. The split is computed from the header's own
  column offsets rather than hardcoded, because they shift between products.
  If a scan ever arrives without column fidelity, every amount is reported as
  a sum insured and the reviewer is told the split failed.
- **The money row is read positionally, not by finding six amounts.** An
  unused `DISCOUNT` prints as a bare `-`, so an "amounts in order" reading
  shifts every value one column left on a discounted policy. The parser
  requires exactly six whitespace-separated cells or reports the row unread.
- **Each PDF prints its face two or three times** (ORIGINAL, AGENT COPY, then
  a summary receipt and three travel ID cards), and the pipeline concatenates
  every page before parsing. The coverage walk is bounded to the first copy
  and the driver list to the first `POLICY HOLDER` block. Unbounded, the
  licencia returns the same person three times — which reads as a
  three-driver policy, not as a bug, so nothing downstream would catch it.
- **Two five-digit numbers sit in the header band** and only one is the agent
  clave: the other is the agent's own postal code
  (`ROSARITO, BAJA CALIFORNIA 22710`). Likewise the agent's street address
  reads `BENITO JUAREZ 25 No.50 INT 38`, three lines above the `No.` cell that
  holds the policy number — hence the two-space floor after `No.`.
- **Sections 6–8 print a PREMIUM where the others print a limit.** $40 is what
  legal aid *cost*, not a $40 liability limit, so it lands on
  `ParsedCoverage.premium` (a field GMX never fills) and gets its own column
  on the review screen. Adding the two together would be meaningless.
- **`coveragePeriodDays` matters here and nowhere else.** A.N.A. sells 3- and
  4-day policies. `Policy.coveragePeriodDays` defaults to 365, so a weekend
  policy left at the default sits in the renewals window a year out. The term
  is derived from the two dates and cross-checked against the printed `DAYS`
  cell; a disagreement is noted rather than resolved.

Exclusions follow the GMX rule — recorded in the risk label
(`MATERIAL DAMAGE — VEHICLE: EXCLUDED`) with a null amount, never as `0`. It
matters more here: a responsabilidad-civil policy prints `0.00` for material
damage, so the two are visually identical on the page.

### Vehicles and drivers

A.N.A. is the first provider whose face carries either, so confirm now writes
`Vehicle` and `InsuredDriver` rows alongside the `Policy`
(`applyVehiclesAndDrivers`). The parsed values are stored on the document as
`extractedVehiclesJson` / `extractedDriversJson` and shown read-only in the
review queue, so a misread VIN is catchable before it is applied.

Both inserts skip a row that already exists on the policy, matched on the
identifier the document prints — VIN then plate for a vehicle (A.N.A.'s
TRAILER and TOWING slots have no VIN), licence number then name for a driver.
The case that forces this is confirming a **renewal** onto an existing policy:
a blind insert leaves the customer with the same VIN listed twice and no way
to tell which row the renewal belongs to.

Nothing is ever updated or deleted there. A vehicle whose plate changed lands
as a second row for a human to reconcile — the safe half of the mistake, since
an overwrite would destroy the only record of what was insured last term.

The vehicle table is parsed **by token role, not by column offset**, because
`BODY` is the cell that wraps: `PACIFICA` is one token and `GENESIS SEDAN` is
two, so a fixed token count reads the VIN out of the wrong slot on the second.
The 17-character VIN is the anchor and `BODY` is whatever sits between the
make and it.

## What the parser reads, and the field it cannot

`ParsedPolicy` fields are all nullable on purpose: each carrier prints a
different subset, and the matcher and review queue both work better with
"field was read" vs "field was not" than with a guess.

Read from the GMX certificate: policy number, insured name, additional
insured, broker (→ `Policy.agentName`), legal address, ZIP, `policyFrom` /
`policyTo` / `policyDate`, currency, premium-payment cadence, and the full
per-coverage table (risk, insured amount, deductible, loss participation)
preserved verbatim.

Read from an A.N.A. face: all of the above except additional insured and
broker parens, **plus** the premium (A.N.A. prints it — see below), the policy
fee, the total, the term in days, the vehicle table, and the named drivers
with their US licence numbers. The tax and the agent clave have no column in
the schema and ride in the notes.

> **The GMX certificate carries no premium.** Not "sometimes missing" — the
> document does not have the figure. It lives on GMX's **separate `recibo`
> PDF**. The parser leaves `netPremium` / `policyFee` / `brokerFee` / `total`
> null and pushes a note onto the row —
> *"esta página no trae prima; revisar el recibo de GMX por separado"* — so
> the reviewer sees why the field is empty rather than assuming a read
> failure.

This is also why confirm never overwrites an existing `Policy.netPremium`
with null: the certificate not carrying a premium is not evidence that the
premium is gone.

A.N.A.'s faces do print one — the `DISCOUNT / PREMIUM / POLICY FEE / TAX /
LOCAL TAX / TOTAL` row is on the same page — so an ANA document reaches the
review queue with `netPremium` populated and `postPremium` already ticked.

Deductible and loss participation are stored as **strings** (`"5%"`, `"20%"`,
`"USD 1,000"`) — they are printed as a mix of percentages, currency amounts
and free text, and normalising them would lose the distinction.

## Confirm: what actually gets written

Per confirmed document, in order:

1. **The `Policy` row** — updated if a policy was matched, created under the
   picked customer if not. Only non-null `extracted*` fields are written; null
   never overwrites existing data.
2. **`Vehicle` and `InsuredDriver` rows** — for the providers whose face
   carries them (A.N.A.; never GMX Hogar), skipping any that already exist on
   the policy. See *Vehicles and drivers* above.
3. **A `PolicyDocument`** — the source PDF is streamed into the policy's
   storage namespace and attached, so the paperwork stays with the policy. Its
   `documentType` is named after whichever parser claimed the page
   (`ANA_POLICY`, `GMX_POLICY`).
4. **Optionally a `Transaction`** — `INSURANCE` domain, negative amount
   (a charge), `captureSource: "OCR"`, `captureRef` = the document id.

The ledger write is **opt-in twice over**: staff must tick `postPremium`
*and* a premium must have parsed to a positive number. Without that gate the
premium-less certificate above would silently book a $0 charge on every
confirm.

`createdPolicyId` and `postedTransactionId` are unique columns on the
document row, so a double-confirm cannot re-apply — and a `POSTED` document
is refused outright.

Discarding a batch is refused once any page is `POSTED`: a partly-applied
batch has already written `Policy` (and possibly `Transaction`) rows, and
hiding the paperwork behind a "discarded" label would leave those rows
unexplained. Reject the remaining pages individually instead.

## API surface

| Method | Route | Ability |
|---|---|---|
| `GET` | `/policy-ocr/status` (is OCR + storage available) | authenticated |
| `GET` | `/policy-ocr/batches`, `/batches/:id`, `/batches/:id/documents` | authenticated |
| `GET` | `/policy-ocr/documents/:id/page` (streams the source PDF) | authenticated |
| `POST` | `/policy-ocr/batches` (upload) | `policy:ingest` |
| `PATCH` | `/policy-ocr/documents/:id` (edit the extracted fields) | `policy:ocr-review` |
| `POST` | `/policy-ocr/documents/:id/reject` | `policy:ocr-review` |
| `POST` | `/policy-ocr/batches/:id/discard` | `policy:ocr-review` |
| `POST` | `/policy-ocr/batches/:id/confirm` | `policy:ocr-review` |

## Requirements

Same as statement OCR: object storage (`S3_ENDPOINT` + credentials) for the
source PDFs and page images, and `tesseract-ocr` / `tesseract-ocr-data-spa` /
`poppler-utils` in the API image. `GET /policy-ocr/status` reports both; if
either is missing the feature reports itself unavailable and only this
feature is disabled.

## Tests

`apps/api/src/policy-ocr/parsers/policy-parser.spec.ts` — 53 cases against
verbatim text extracted from five real documents, indentation and blank lines
included (the column positions are what the parser reads, so a cleaned-up
fixture would test nothing — and on A.N.A.'s driver's policy the offsets are
literally the only thing separating two columns).

From `HC_Folio_000767_Traduccion.pdf` (caratula): provider detection from the
wordmark and from the footer URL, the header fields, every coverage row off
the second page, the deductible/loss-participation strings, the
missing-premium note, the broker line with the agent-number parens absent,
and a page with no GMX signal at all (which must yield no provider rather
than a bad guess).

From `007_LGS-HGMX_07006957_01_0-CondicionesParticulares.pdf`
(especificación): the differently-grouped policy number, the risk location
read across its wrapped line, the empty `Asegurado Adicional` cell that must
not capture the next line, the absent-by-design fields, currency taken from
the USD limits rather than the M.N. sublimits in the body prose, a limit
split under `Edificio` / `Contenidos` sub-labels, a limit printed on the
label's own line, a deductible stated as a sentence *above* its limit, a
sublimit block whose amount sits after both a blank line and a page break,
the excluded earthquake coverage, and the hydrometeorological deductible and
coinsurance pulled from their own per-zone block.

From the three A.N.A. PDFs: brand detection (and that GMX's layout rules
cannot claim an ANA page), the header band, DD MM YYYY read out of three
separate column cells, the six money cells with `DISCOUNT` printed as a bare
`-`, the vehicle row with a one-word and a two-word `BODY` cell, the empty
TRAILER/TOWING slots, the agent street number and postal code that must *not*
be read as the policy number and clave, the declared values labelled by item
slot, the `$500.00` inside the deductible sentence that is not a sum insured,
the per-person/per-accident split in both of its printed forms, an add-on's
figure recorded as a premium, section 9's parenthesised limit, the by-the-day
term, the excluded sections, the column-position split on the driver's policy,
its different section order, and — for both faces — that a doubled or tripled
input yields one set of coverages and one driver rather than one per copy.

Four of the GMX cases are regression tests for ways the parser can silently attach
the *wrong* value rather than none — a neighbouring coverage's prose read as
a deductible, the page-level `DEDUCIBLES:` paragraph read as one, a coverage
named after a wrapped prose tail, and one section's per-zone deductible
adopted by the coverage above it. Each was a real defect caught by running
the parser against the full ten-page document.

## Not built

- **GMX and A.N.A. only.** The dispatcher (`detectPolicyProvider`) is a table
  of `[provider, pattern]` pairs plus a `parsers` map, so adding Qualitas is a
  parser function and two entries — but no other carrier's layout has been
  seen yet, and guessing at one produces a parser nobody can verify.
- **The `recibo` PDF.** Reading the premium off GMX's separate receipt
  document, and pairing it to the certificate it belongs to, is the obvious
  next piece. It is what would let `postPremium` stop being a manual tick.
- **Renewals from OCR.** A re-issued policy arrives as a new certificate with
  the same number; confirm updates the existing row rather than versioning
  it. Nothing tracks "this is the 2027 issue of that policy".

## Related

- [`STATEMENT_OCR.md`](STATEMENT_OCR.md) — the utility statement pipeline this
  was lifted from, as built. [`RECEIPT_CAPTURE_SPEC.md`](RECEIPT_CAPTURE_SPEC.md)
  §2 is its design and the measured evidence behind it. Between them they are
  the origin of three rules the policy parser applies: detect the provider by brand before layout, only
  ever apply the Tesseract digit-confusion map (`O→0`, `S→5`, `B→8`, …) to
  fields known to be digits, and parse amounts by separator *position* rather
  than assuming `,` is thousands.

  Those last two are **duplicated on purpose**, not imported: the module is
  kept self-contained, since sharing a helper would couple two unrelated
  domains through it. If you fix a bug in one, check the other.
- [`INSURANCE_FEATURES_SPEC.md`](INSURANCE_FEATURES_SPEC.md) — the four
  insurance features that *were* planned. This is not one of them.