Methods: TickerTrac's Parallel evaluation

All questions used public sources. We deliberately selected known historical cases with factual distinctions such as announced versus completed transactions, annual versus quarterly results, and conditional versus received payments. The sample was not random customer traffic or a breaking-news discovery test.

The wider completed evaluation included 200 queries in each of Parallel's Fast and Advanced search modes, 100 document extractions including 23 PDFs, 100 Core research questions and a 30-question Pro subset. TickerTrac controls covered 52 retained-filing questions and 48 web questions.

The retained-filing control searched material already acquired and indexed. The web control fetched fixed reference URLs and passed extracted text to TickerTrac's existing answer component. Parallel Task could obtain alternative sources. The web control therefore measured a limited acquisition/extraction/answering path, not TickerTrac's full discovery stack or a matched comparison of answer models.

Twenty-four of the 48 web controls initially lacked usable text: eleven HTTP 403 failures, two HTTP 404 failures, one redirect outside the permitted host and ten parser failures. Eight parser failures removed article text enclosed in forms; two missed non-paragraph content.

We reprocessed the same 48 controls using frozen inputs. Previously acquired HTML bytes were re-extracted; PDF inputs and failed-fetch records were unchanged. No sources were fetched again in this parser rerun. The 34 controls with usable text received new model answers; fourteen empty-input controls abstained without a model call. The answer model, answer contract and 6,000-character allowance per source stayed fixed. Generation randomness remained uncontrolled. The original Parallel responses and grades were reused.

Separate AI agents graded the original answers with provider labels hidden. A separate agent review evaluated the 48 repaired control outputs, including abstentions, against supplied passages and frozen references. Neither was independent human review, and concealed labels do not guarantee concealed provider identity.

A pass required correct material facts and supporting citations. Partial answers, missing inputs and abstentions remained in the denominator. A stricter date-scope interpretation changes two repaired passes to partial: a Federal Reserve decision/effective-date distinction and a J&J answer using a closing release for an announcement-time question. That reduces the repaired total to 29/48, including 8/10 parser-recovered questions. The primary count and sensitivity result are both reported.

We retained the original fourteen acquisition-failure questions and URLs. The first fallback passed 10/14 after two explicit abstentions escalated to Core. The revised policy retained relevant passages instead of only document openings and could follow limited same-publisher links to underlying evidence. Requests, selected sources and elapsed work were bounded; paid calls reserved cost before dispatch. Evaluation grades did not determine routing.

The revised run used one answer generation per question, with the same answer model and contract. All fourteen obtained source text; thirteen passed unblinded primary-agent review. Sandisk remained a strict failure because its answer listed February 21 and February 24 while describing a "one-day gap." No answer was manually corrected or selectively regenerated to improve the score.

This is a known-case regression result. All fourteen cases had informed the changes. Combining it with the earlier 34 parser-repaired cases gives 44/48, or 42/48 under the date sensitivity check, but that is a composite of separate runs and reviews. It is not a fresh matched 48-question trial.

The policy retained 22 vendor-derived documents and two native fetches. Native acquisition conditions also changed: a Mastercard PDF that previously returned HTTP 429 succeeded in this run. The gain cannot all be attributed to code changes. Link selection sometimes acquired unnecessary general or later pages, so historical source selection remains a limitation.

The experimental policy included Search and linked pages. The initial product integration is narrower: Parallel Extract for an exact public news URL after native retrieval fails. Automatic replacement search is outside that integration because a common publisher does not establish that two pages describe the same event.

The answer component used Claude Sonnet 5 and recorded token usage. Parallel estimates use $0.001 per Extract URL, $0.001 per Fast Search and $0.025 per successful Core run from its published pricing, checked October 20, 2026. These are provider-call estimates, not reconciled invoices; engineering, review, hosting and storage are excluded. TickerTrac received $20 in signup credits.

Median elapsed time for the fourteen uninterrupted serial runs was 5.3438 seconds, with a 2.8039 to 14.7152-second range. This includes acquisition, passage selection and answering. Fourteen observations do not establish a service-level guarantee. Reused Core jobs cost $0.35 for 13/14 passes, but used different scheduling and polling; their latency is not comparable.

The first fallback's successful path cost an estimated $0.225886. A lost answer receipt retained a conservative $0.25 reservation, raising that experiment's upper estimate to $0.475886. The revised run is separate. Its first launch stopped before paid work; the completed run above made no Core calls. Reporting the successful run alone should not be read as the cost of the entire evaluation program.

The separate comparison of 20 frozen monitoring topics ends November 3, 2026. Publication time, event time and first observation are evaluated separately. Failed or capped collection is incomplete evidence. Comparative recall will remain inconclusive if the trial yields fewer than 30 supported event clusters. No final monitoring result is claimed here.