> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gdeltcloud.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Limits and error bars

> What the four metrics do not claim, and the measured noise floor on each scale.

Read this before putting these scores in a model, an alert threshold, or a customer-facing product.

## They are judgements, not measurements

Nobody measured an event's systemic importance with an instrument. A language model read the coverage,
answered a fixed set of questions against a published rubric, and a fixed formula turned those answers into a
number. That is more auditable than a score with no stated method — you can read every input and disagree
with a specific one — but it is still judgement.

## They do not predict anything

`propagation_potential` is not a probability that something will spread. `market_sensitivity` is not a
forecast price move or a direction. Both score conditions **present in the event as reported**.

## The frameworks are borrowed structure, not endorsement

We adapt the *shape* of established methods — Basel's equal-weighted G-SIB categories, the ERCS barrier
model, Richardson's log-fatality scale, the reasonable-investor materiality test. That structure is what
makes these scores legible to an analyst who already knows them. It does **not** mean the results carry those
institutions' authority.

<Warning>
  **None of these institutions endorses, reviews, or is connected to this work.**
</Warning>

## Coverage against hard registries is thin

Across 42,099 events, **2.2%** touch a node matchable to a hard registry and **1.5%** attach to an
identifiable traded instrument. Treat all four as **ordinal ranking signals** — good for sorting and triage,
not for asserting one event is 1.4× another.

## CAMEO+ only

The four metrics exist only on CAMEO+ events. **Conflict events omit them** rather than returning `0` — a
`0` would assert "measured, and it is the minimum", which is false. A filter like
`min_magnitude=6&event_family=conflict` therefore returns an empty `200`.

## The error on each scale

We coded the same 59 events three times and measured how far each score moved when nothing changed.

| Metric                  | Mean     | Run-to-run σ | **Smallest defensible difference** |
| ----------------------- | -------- | ------------ | ---------------------------------- |
| `magnitude`             | 3.5 / 10 | ±0.36        | **0.72**                           |
| `systemic_importance`   | 0.26     | ±0.036       | **0.07**                           |
| `propagation_potential` | 0.31     | ±0.029       | **0.06**                           |
| `market_sensitivity`    | 0.27     | ±0.030       | **0.06**                           |

<Warning>
  **Never read a single-event difference smaller than the last column.** It is indistinguishable from the
  instrument talking to itself.
</Warning>

Magnitude is the least stable, and nearly all of that sits on the `verbal` path — verbal events swung up to
4 points between identical runs, while `hazard` averaged σ 0.14.

## Reliability is not validity

Three runs agreeing proves the instrument is **consistent**, not that it is **right**. A scorer that was
reliably wrong would look identical in the table above. Establishing correctness needs a rubric-blind human
gold set graded by domain experts, which **we have not built**.

`propagation_potential` in particular has **no outcome validation** — we do not currently measure whether
high-scoring events are followed by more downstream events than low-scoring ones.

## Effective date and backfill

|                     |                                                             |
| ------------------- | ----------------------------------------------------------- |
| Contract version    | `v2-2026-07-24`                                             |
| In effect from      | **28 July 2026**                                            |
| Older events        | Previous scores, empty `metric_version` — a different scale |
| Sub-factor evidence | Returned only for events coded from 28 July 2026            |
| Backfill            | **Planned, not yet run**                                    |

`metric_version` is stored but **not currently returned by the API**. Use the presence of
`metrics.metric_inputs` as the boundary marker — it appears only on v2 events.