The record model separates evidence, interpretation, prediction and evaluation so that one cannot silently inherit the authority of another.

## The nine record types

| Record | Purpose | Example |
|---|---|---|
| `resource_snapshot` | Identify retained source material and capture limitations | A permitted article excerpt or statistical release |
| `source_assertion` | Preserve an attributed statement and selectors | A quoted claim about future demand |
| `interpretation` | Record a separately attributed reading | An extractor's interpretation and its confidence |
| `assessment` | Retain an explicit assessment | A review of the available evidence |
| `question` | Define the target and resolution contract | Whether monthly throughput falls below a threshold |
| `forecast` | Record declared structured output against a pinned question | An unconditional binary probability of 0.70 |
| `outcome` | Record resolution and supporting observations | The defined event occurred: 1 |
| `evaluation` | Pin the forecast, outcome, rule and result | Brier loss of 0.09 |
| `artifact_descriptor` | Identify an artifact and its content | A simulation result or dataset artifact |

## Identity is not truth

Two identities serve different purposes:

- **Content digest:** SHA-256 of exact bytes. Line endings or whitespace changes alter it.
- **Record digest:** SHA-256 of the entire canonical JSON body using RFC 8785. All record members contribute.

A reference pins type, id, revision and digest. It identifies the particular record used, not merely its latest name. Canonicalization and hashing do not validate the record schema; use `load_record` or `validate` too.

There is no independent timestamp or signature assurance in the 0.1 core. A locally declared `recorded_at` time remains a declaration. [Security boundaries](/sdk/docs/security) explain what verification can and cannot establish.

## Corrections preserve history

<div class="diagram" role="img" aria-label="An evaluation remains pinned to forecast revision 1 and outcome revision 1. A corrected outcome revision 2 creates a new evaluation; the original remains reproducible."><div class="card-grid"><div class="flow-node"><strong>Original evidence</strong><span>Forecast r1 + outcome r1 → evaluation A</span><small>Retained and reproducible</small></div><div class="flow-node"><strong>Correction</strong><span>Outcome r2 names predecessor r1 → evaluation B</span><small>Append a successor; preserve A</small></div></div></div>

An identical append is idempotent. Different content at the same type/id/revision is refused. A correction is a new revision with a predecessor reference; it does not overwrite the earlier record.

## Time has several meanings

| Time | Question it answers |
|---|---|
| Source publication | When did the source publish the material? |
| Capture / retrieval | When did this application obtain these bytes? |
| Information cutoff | What information was allowed when forecasting? |
| Target period | Which real-world interval is the forecast about? |
| Resolution / grading deadline | When and how is the outcome determined? |
| Local recording / evaluation | When does this recorder say it saved or evaluated the record? |

These meanings are not interchangeable. A provider's build date is not a publication timestamp. A day-only date cannot establish an exact ordering within that day. An absent cutoff can prevent scoring even when a number looks like a probability.

## Missing, withheld and absent

Supported fields can use explicit markers such as `{"unknown": "source_does_not_declare"}` or `{"withheld": "restricted_information"}`. Use them only where the corresponding schema allows them. `null` can represent known absence in a particular field; it is not a universal missing-value convention.

Preserving an unknown is usually more useful than fabricating a default. The evaluator reports named exclusions rather than quietly repairing the source's meaning.

## Four numbers to keep apart

1. **Extraction confidence:** confidence that an article was read correctly.
2. **Declared event probability:** a forecaster's stated chance of a defined event.
3. **Scenario result:** model output conditional on assumptions.
4. **Observed measurement:** a reported value for a defined population and period.

Only eligible declared binary probabilities are handled by the current Brier evaluator. A market index of 0.29 is not thereby a 29% probability. An interval is not automatically a confidence interval. A model's simulated event frequency is not automatically a calibrated real-world probability.