Repairing misleading news clips: blind rules vs LLMs vs a gated fix

Result

Extending to the story edge restores 94% of lost context but changes 98% of fair clips. Gemini 4 Argon restores 74% and changes about a third of fair clips; gating with GLM-5.2 first leaves 96% of fair clips untouched and restores 44%.

Data

162 clips

54 misleading, 108 fair

Compute

Hosted APIs

rules need none

Truth

Exact

the edge that was moved

Metric

Restored %

no bootstrap: 54 misleading items

Misleading clips with their context restored (%) · higher is better

Dot = measured, dashed line = best. Hollow dot = reference row (agreement ceiling or baseline), not a contender.

Model	Context restored	Judged fair after %	Added s, misleading	Fair clips changed %	Added s, fair
Rule: extend to story edgebaseline	94.40	90.7	40.3	98.1	29.7
Model repair: Gemini 4 Argonclosed	74.10	94.4	15.8	31.5	5.7
Rule: extend 2 sentencesbaseline	64.80	92.6	24.8	99.1	20.1
Model repair: GLM-5.2open	64.20	100.0	18.5	82.7	14.7
Model repair: DeepSeek V4.1 Flashopen	48.10	79.6	17.3	67.6	13.0
Gate GLM-5.2, repair Argon	44.40	75.9	9.9	3.7	0.8
Gate GLM-5.2, repair GLM-5.2	44.20	80.8	11.7	4.7	1.1
Rule: extend 1 sentencebaseline	35.20	87.0	13.9	99.1	11.5
Gate GLM-5.2, extend 1	20.40	72.2	8.7	5.6	1.0

The hookLink to this section

Stage 10 found a judge that can flag a clip that misleads out of context. Flagging is half the job. The other half is fixing the clip: putting the lost context back without turning every 30-second short into a 90-second one.

Because each trap was built by moving one edge of a fair clip, we know exactly where the context went. So repairs can be scored without a judge: did the edge come back?

The setup, in one pictureLink to this section

stage-10 clips: 54 misleading · 108 fair (54 faithful, 54 trimmed)
   ▼
9 repair policies, all in whole sentences, at most 75 s:
   rules   extend 1 or 2 sentences each side · extend to the story edges
   models  see the numbered story + clip, return a new span (or keep it)
   gated   repair only when the GLM-5.2 verifier flags the clip
   ▼
context restored %  (exact: the moved edge is back)
fair clips changed % · seconds added · second check by a judge

What the numbers sayLink to this section

Blind rules work only by making everything long. Extending to the story edge restores 94.4% of lost context, but adds 30–40 s and changes 98% of fair clips. Extending by one sentence restores 35%: lost context is usually more than one sentence away.
A model that reads the story repairs better per second. Gemini 4 Argon restores 74.1%, adding about 16 s to misleading clips and changing roughly a third of fair clips by about 6 s on average.
GLM-5.2 fixes but fidgets. It restores 64% but changes 83% of fair clips: it doesn't leave good clips alone.
Gating is the precise option. With GLM-5.2 deciding when to repair, fair clips are almost untouched (3.7% changed, +0.8 s) and 44% of misleading clips get their context back. The ceiling is the gate: GLM-5.2 flags 65% of traps, and Argon restores most of those.

The dirty laundryLink to this section

54 misleading items, no bootstrap. Differences under about 10 points aren't reliable.
"Restored" is strict. A repair that fixes the meaning another way, such as moving the other edge, counts as not restored. The judge column covers that case, but the judge is lenient: it already calls 28% of the unrepaired misleading clips fair.
Model-made traps, text only, as in stage 10.

The recipeLink to this section

Pick by the cost of a miss. For a newsroom, run the ungated model repair (74% restored). For volume, gate first so fair clips stay short and untouched.
Don't fix context with a fixed padding rule. It either misses (one sentence) or bloats everything (story edge).
Score repairs against the edge you removed if you can construct the error. It's exact and needs no judge.

What's nextLink to this section

Raising the gate's recall is the main lever: every trap the verifier misses is a trap nobody repairs.

Method & environment

Question: When a clip misleads because context was cut, which policy restores the context at the lowest cost in added seconds, while leaving fair clips alone?
Compute: Hosted APIs (Vertex AI)
Method: A repair restores context when the moved edge is back at or past the faithful clip's edge, which needs no judge. All repairs work in whole sentences, at most 75 s. Differences under about 10 points are not reliable at this size.
Numbers are imported directly from the run's result files; nothing is typed by hand.