*English edition (en-US), September 29, 2026. This English text is the reference version.*

# Reading Error Matrix — measuring our own margin of error

_Last updated: September 8, 2026 · tolerance: ±0.5% · reproduction: `python3 -m pipeline.hata_matrisi` (0 tokens)_

> **Why do we publish this?** The claim "we are the accuracy layer" should itself be testable.
> When you ask an audit firm or a data provider "what is your margin of error?",
> you get no answer. Ours is here.

## A. Independent second-reading round (machine-to-machine agreement)

Every medium-confidence extraction was compared against a second reader that had never seen the record.

| Measure | Value |
|---|---|
| Pool (medium-confidence extractions) | 3,046 records |
| Deterministic cross-verification | **1,688 records in agreement** (the rest were passed to the adjudicator round) |
| Adjudicator reader round | 1,193 records · agreement 719 · discrepancy 51 · no verdict 423 |
| **Discrepancy rate among adjudicated records** | **6.6%** (51/770) |

⚠️ **Note (in the interest of honesty):** This is machine-to-machine agreement. If the two readers make the SAME error,
no discrepancy appears — so this rate is the **lower bound** of our margin of error,
not its upper bound. An upper bound can only be given with a human-labeled gold set (below).

### What did the discrepancies turn out to be? (the first 10 cases that went through human review)

| Class | Count | What it means |
|---|---|---|
| Reading error (ours) | 6 | The report was correct, our extraction was wrong → record corrected, company not penalized |
| Company-internal disclosure inconsistency | 1 | Contradiction within the report itself → flag on the company (Accuracy penalty) |
| False alarm | 1 | The difference was merely rounding → closed |
| Unverifiable reading | 2 | Could not be confirmed from the text → score not published |

In the first batch, **60% of the discrepancies turned out to be our own reading errors** — all were corrected and re-sealed
(versions are public; no record is deleted).

## B. Human-labeled gold set (the real error matrix)

- Target: **200 reports**, a sample stratified by source pipeline (`data/altin/altin_set.json`)
- Labeled: **0** · status: **labeling in progress**

_Labeling is in progress. Once it is complete, the field-by-field matrix (incorrect / missed / fabricated) and the UPPER bound of the margin of error will be published here._

## What do we do with errors?

1. **Our own reading error** → the record is corrected (append-only new version + seal), the company is not penalized.
2. **A contradiction in the company's own report** → flag; in a severe case, the Accuracy cap is applied.
3. **Unverifiable reading** → the score is not published ("reading under verification").

_Companies may always file a dispute (10-business-day SLA); dispute outcomes and corrected
seals are on the public record._