Testing Parallel exposed a bug in our own research pipeline

Our wider evaluation covered search, document extraction and research questions. The most instructive part was a 48-question web control. It fetched fixed reference URLs, extracted their text and passed that text to our existing answer component. Parallel Task could find its own sources.

That control was deliberately limited. It did not run TickerTrac's full discovery workflow, and its performance cannot be read as our production coverage rate.

Initially, only 24 questions had usable source text. We traced the empty inputs back to their documents. Fourteen had acquisition failures, including blocked requests and broken URLs. Ten exposed parser defects.

In eight cases, an article was nested inside an HTML form. We removed forms to discard page controls and removed the article with them. Two other releases put important content outside paragraph elements, including in tables. Our selector missed it.

We repaired the parser and reprocessed all 48 controls from saved inputs, without fetching the sources again. Usable inputs rose from 24 to 34, so we generated 34 new answers; fourteen controls still lacked text and abstained without a model call. The answer model and per-source text limit stayed fixed. Agent review counted 31 answers with correct material facts and supporting citations, up from 19. A stricter interpretation of two answers' date scope lowers the repaired count to 29. Both readings are retained in the methods.

For the fourteen questions still missing source text, we tested a limited fallback: try direct retrieval, use Parallel Extract when acquisition fails, and use Search to locate a replacement on the original publisher's site when needed. Our existing answer component then reads the recovered passages.

The first version produced ten complete, citation-supported answers. Several failures came from choosing the wrong part of a successfully retrieved document.

Sandisk's separation date and Talen's cash-proceeds qualification were beyond the beginning of the page we supplied to the model. Peloton's umbrella announcement linked to the individual recall notices, but did not itself contain the hazards the answer described. Correct facts still need citations that support them.

We changed passage selection to retain relevant sections within the same text allowance and allowed a small number of links to underlying evidence. The revised run passed thirteen of the fourteen questions. None required escalation to Parallel's Core research processor.

The original Core responses also passed thirteen of these questions, at an estimated $0.35. The revised route cost about 41% less in provider calls on this set. Those Core responses were reused, and their timing included a different execution and polling schedule. We cannot make a matched speed claim.

The fallback still depended heavily on Parallel. Of 24 retained source documents, 22 came through the vendor and two through our direct fetcher. Repairing our own code helped us use those documents better.

The remaining Sandisk answer gave the correct legal separation and trading dates, then called the interval a "one-day gap." The separation occurred on February 21, 2025; regular trading began February 24, the next trading day. Western Digital's filing records both dates.

We counted the answer as a strict failure and kept it unchanged. Better retrieval had supplied the missing evidence, but it had not prevented an ambiguous, internally inconsistent summary. An escalation rule that only catches empty sources or explicit abstention will miss that kind of confident error.

These fourteen cases informed the fixes. They are a development set, reviewed by an AI agent with provider identity visible, not an independent human benchmark. The result justifies keeping the improvements and testing unseen cases. The sample cannot establish production accuracy.

We added an opt-in Parallel Extract fallback for the exact public news URL our normal fetcher could not retrieve. Direct retrieval remains first. Automatic Search and replacement-page selection remain outside this initial integration: two pages from the same publisher may describe different events. The fallback has a spending limit, reuses previously recovered material and records which service supplied the text. Returned content still has to support the answer's issuer, period and factual qualifications.

For TickerTrac, that recovered evidence must also remain connected to the investor's exact thesis version and its history. Parallel already supplies citations and a research basis; our application has to preserve how evidence was used and what the investor subsequently chose to change.

A separate 20-topic monitoring comparison runs through November 3, 2026. It has no final result yet. The acquisition experiments give us a useful implementation decision while that longer test continues.

We received $20 in signup credits. The costs above use published Parallel prices and recorded answer-model usage, rather than cash charged. They exclude engineering, review, hosting and storage. The methods supplement gives the samples, grading limits and cost breakdown.