Ground the decision model

Ungrounded models guess confidently on stale facts

Cheap inference has caused read-only decision and classification traffic to explode—once a model sees the right context, it takes very little output to make a call. The catch is feeding world-state questions to a model whose knowledge stopped at pretraining:

“And I'm here to make the case that like grounding for those is extremely important.”00:00:00

Ask an ungrounded model who won the Michigan Senate race—something resolved immediately on any search engine—and it returns a wrong candidate with false mathematical certainty:

“Okay, so maybe a question: Who won the Michigan Senate race? This should be easy. This is like well known. I mean, a couple days ago, if you do Google searches, the first thing that shows up. But if you ask Jeff this question, this is what'll give you back. Mike Rogers with fifty-eight percent. It really doesn't know.”00:00:55

“And so this is where we think that like not only is it wrong, but also its confidence is too high.”00:00:55

Evidence: transcript 00:00:00, 00:00:55

Grounding unlocks calibration, not just accuracy

To measure real-world decision quality beyond trivia, Parallel released the open PWS-Decisions benchmark (300 questions drawn from live workloads where ungrounded models lack ground truth). Comparing ungrounded models against web-grounded ones shows gains on both accuracy and calibration:

“What is correct is much more confident that's correct if you use actual grounding to ground your decision.”00:01:50

“Calibration is the important thing about good decision models.”00:01:50

In Q&A, Gupta notes that complex calibration schemes exist mostly to compensate when a model is guessing in the dark—once live web evidence is present, the answer is clear enough that calibration follows naturally.

Evidence: transcript 00:01:50, 00:07:45

One live search, shared state, $1 per 1,000 requests

The Decisions API keeps the state-input interface of a standard decision call, runs a live search against Parallel’s ~100-billion-page web index, and attaches the top evidence straight into the context:

“We do a live search, web search call, and we get information from our proprietary index, which is like. On the size, I think that's at least their size, which is about 100 billion pages, and we find the most relevant content for that decision.”00:03:01

“It runs under a second in response time. It's one dollar per thousand requests.”00:03:01

That makes grounded decisions roughly an order of magnitude faster and one to two orders of magnitude cheaper than multi-step LLM search chains—and fast enough that teams run lookups interactively while a user is still typing a company name into an onboarding flow.

Evidence: transcript 00:03:01, 00:04:05, 00:07:03

The “Is Jensen still CEO of Nvidia?” test

In the live demo, Gupta asks an ungrounded model whether Jensen Huang is CEO of Nvidia; trained on decades of history, it returns `"true"` at 96% confidence without ever checking today’s reality. Running the same question through the Decisions API returns the current answer alongside a primary citation:

“Not only do we give you the answer, we also give you the citation, so you can go fact check the source yourself.”00:05:45

Three production patterns illustrate why sub-second cited classification matters at scale: classifying 100,000 California restaurants by reservation platform and website builder for ~$100, running live business KYC checks, and categorizing breaking news.

“They have a list of like a hundred thousand restaurants. California, and you search every single one of them up here on the website, and you can actually classify for every single one of them.”00:04:34

Evidence: transcript 00:04:34, 00:05:45

The shared-context trap in parallel questions

In Q&A, an engineer who tested the API reports a sharp edge case: asking single questions works well, but batching orthogonal questions in parallel—such as election winner, immigration stance, and party affiliation—forces all of them onto one retrieved search result set:

“Because the same search will back those results, it actually degrades as opposed to asking each of those as. Separate questions.”00:07:45

Gupta confirms that the v1 product shape optimizes around one grounding search per entity state rather than multi-query fan-out:

“Yeah, I think the intention behind the product shape is that we want. A single sort of grounding or single search to answer a question.”00:07:45

Evidence: transcript 00:07:45

Automated research, KYC, and dossier pipelines

Put together, web-grounded decision models occupy a clean niche beneath full research agents: sub-second typed classification with source URLs at one-tenth of a cent per call.

“And so it all comes to like one dollar per thousand requests. It's super easy. It's super fast.”00:05:45

• Ground before branching: use live web retrieval so confidence scores are calibrated rather than overconfident pretraining guesses (00:00:55, 00:01:50).
• Cheap cited enrichment: classify hundreds of thousands of entities, run live KYC, or corroborate dossier claims for $1 / 1,000 requests (00:03:01, 00:04:34).
• Match batching to search scope: share one grounding search across co-located facts, and split divergent angles into separate calls (00:07:45).

Evidence: transcript 00:03:01, 00:04:34, 00:05:45