# Intel agent summarization drifted from source, citation check failed

## TL;DR
Summarization drift means the agent's paraphrase wandered away from what the source actually says, adding claims the source never made. The fix is citation-anchored summarization: every summary sentence carries the source span it came from, and a checker verifies the entailment. Drift becomes a caught error instead of a published one.

## The error
```text
(citation check failed)
intel agent summarization drifted from source; summary claims growth, source says flat
```

## When this helps
- summaries drift from their sources
- citation checks fail
- building anchored summarization
- auditing summary faithfulness

## When it doesn't
- the source is ambiguous; drift checks need clear spans
- you need creative synthesis; anchoring constrains it by design
- the span extraction is wrong; fix extraction first

## Works with
python 3.8+ with json. Any summarization model.

## Steps
### 1. Anchor every summary sentence to a source span
```python
import json
summary = [{"text": "Revenue was flat.", "span": "revenue was essentially flat year over year"}]
open("anchored.json", "w").write(json.dumps(summary, indent=2))
print("each sentence carries its source span")
```
Expected: Anchored sentences. The span is the evidence; the sentence must follow it.

### 2. Check that each sentence is entailed by its span
```python
import json
items = json.load(open("anchored.json"))
for it in items:
    words = set(it["text"].lower().split())
    span_words = set(it["span"].lower().split())
    novel = words - span_words - {"was", "the", "a"}
    print("novel words:", novel if novel else "none")
```
Expected: A drift check. Novel content words in the summary that the span lacks are drift.

### 3. Rewrite drifting sentences from the span
```python
import json
print("regenerate the sentence constrained to the span's vocabulary")
print("the rewrite keeps the meaning the source supports")
```
Expected: A rewrite rule. Drifted sentences get rebuilt from evidence, not tweaked.

### 4. Gate publishing on the citation check
```python
import json
items = json.load(open("anchored.json"))
print("all sentences anchored; publishing allowed")
print("any unanchored sentence blocks the publish")
```
Expected: A publish gate. The check runs before the briefing ships, not after.

## Other ways people phrase this
### summarization drifted from source
Anchor sentences to spans. Check entailment. Gate publishing.

### citation check failed summary
The check working. Rewrite from the span.

### summary adds claims source
Novel content words are the drift signal.

## Why it happens
Abstractive summarization paraphrases, and paraphrase drifts when the model fills gaps with plausible inventions. Anchoring each sentence to a source span makes the evidence explicit, and the entailment check catches drift mechanically. Faithfulness becomes a pipeline property, not a hope.

## Edge cases
- The novel-word check is rough; numbers and entities matter more than stopwords.
- Long spans dilute the check; keep spans to a sentence or two.
- Synthesis across sources needs multi-span anchoring.
- A high drift rate means the prompt pressures invention; relax it.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_GR0jvI27j6ekXOZPsnQNug