# Duplicates near DL

> near DL, RPD is noise — use an absolute test

Source: https://geodb.io/guide/topic/qaqc_duplicate_detection_limit_floor/ · Category: QAQC
A pair of repeats at **0.01 and 0.02 ppm** gold is a **100% relative
difference** — and if the detection limit is 0.01 ppm, it is also
completely meaningless: that is instrument noise, not imprecision. Near
the detection limit, relative statistics like RPD lose their meaning, and
judging them anyway *"can skew our perception of the precision"*.

## The convention

Pick a practical floor — **a good rule of thumb is 10× the analytical
detection limit**, commonly called the **"line of significance"** — and
plot it on your precision charts. Bishi states it as protocol: repeat
values proximal to the detection limit are **"not considered failures."**

The implementable two-regime shape (from a production QC system,
EarthSoft EQuIS):

| Regime | Test |
|---|---|
| both results ≥ 5× DL | relative test (RPD/HARD) vs the normal threshold |
| either result < 5× DL | **absolute** test: flag only if \|a−b\| > 1–2× DL |

**Do not simply discard sub-threshold pairs** — that hides real
contamination and real sub-sampling failure at low grade. Switch the
statistic; keep the data.

(Physical basis: precision degrades asymptotically toward the DL; it is
optimised around ~50× DL.)

## In geoDB

geoDB applies the two-regime test:

- **both results ≥ 5 × the detection limit** → RPD against the normal
  threshold for that duplicate type;
- **either result below it** → the pair is judged on **absolute
  difference**, and fails only if the two results differ by more than
  **one detection limit**.

The 5× boundary is per-project (*Project settings → QAQC protocol →
duplicate detection-limit floor*); set it to 0 to judge everything on RPD.

Two things worth knowing:

- **The pair is re-judged, never dropped.** A near-DL pair that genuinely
  disagrees still fails — the floor removes noise, not evidence.
- **The floor can only ever rescue a pair.** Where the absolute test would
  be harsher than RPD, the RPD verdict stands, so raising the multiple can
  never manufacture a failure.

Every duplicate row says which test judged it, in the **Judged by** column
and in the chart hover; near-DL points plot as diamonds because the RPD
threshold line does not apply to them.

(Pairs where a value is *below* detection are handled separately:
both-below counts as agreement, one-below is not evaluable and is reported
rather than judged.)