Finding Canada's naval suppliers with LoQuery

The header with the extension status and your balance, and the two ways to ask: name the things yourself, or describe them and let LoQuery find the matches.

Checking that the extension in this browser is live, because that is where the searching will happen.

The searches and page reads run from your own browser, not a shared scraping service. Your balance is yours to watch: a run spends from it and nothing recurs.

The two text boxes, the list of sites LoQuery keeps out of its sources when it can (Wikipedia, Reddit, social media by default), and a plain sentence about how discovery runs.

Nothing yet. It waits for a question. Start Research stays off until one is in.

"What to find out for each" is the contract: those facts become the columns of your table. The sites-to-avoid list is yours to edit.

The model choice with its price printed beside it, the discovery budget (20 rounds, up to 300 pages), and three reasoning switches that are on by default.

No hidden pricing: the standard model is about $0.59 per research, the fast one about $1.22, and you can change the discovery budget before or during a run.

A header that names the current stage, an elapsed clock and the agent at work. Below it, a board that will hold one card per company, and the steering controls.

Reading your question, setting the discovery budget you chose, and preparing the first searches.

You are never looking at a spinner. The panel says what is happening, and "Tell LoQuery in your own words" lets you redirect it in plain language while it runs.

What research costs per item (about three cents, sometimes six or more), a note that discovery itself costs money even when it finds little, your balance, and three buttons.

Estimating from its own past runs. The run has already begun by the time the dialog shows, and Stop Research ends it at once.

You see the shape of the bill before the work is done, in dollars, next to your balance.

How LoQuery understood the question (as tags), a funnel line that says how many queries and pages produced the list, and every candidate with its source and an × to drop it.

Four waves of searching: a broad search, a focused expansion, a deep dive and a last pass, up to the 20 rounds you allowed. This run used all 20 and read 210 pages across 49 sites.

Nothing is hidden. You can see the yards and primes arrive (Irving, Seaspan, Davie, Lockheed Martin Canada, Thales Canada, Babcock Canada, TKMS Canada, Ultra Maritime, OSI Maritime, BMT Canada, Kraken Robotics) and you can see where each came from.

Names that LoQuery is not sure about, with the one-line description it found and two buttons. Below them, the steering actions you can use right now.

Verifying the uncertain names on its own (in this run, Gastops was held, then confirmed and added), and rejecting off-topic ones with a written reason, for example "aerospace supplier, not a naval solutions provider".

A rejection with a reason can be found and reversed. A guess cannot.

The end message, a four-minute timer you can extend, the full candidate list with a remove control on each name, and the choice: research the list, run more discovery, or cancel.

Waiting. Names it could not confirm are marked "held, not checked" and research skips them unless you keep them.

This is the moment that keeps the thinking human: the list is yours to prune before any research money is spent on it.

Your ten criteria have become 21 named columns: headquarters city and province as two, the contract's awarding body, value and date as three, recent news as a title and a date, plus a summary, the website and the source links.

Deciding the order it will fill the fields in, and the sources it will prefer for this topic: defence.gc.ca, the yards' and primes' own sites, navsource, SEDAR.

The shape of your answer is fixed and visible before the work starts. What it will not be able to fill is listed too, as field limitations.

All 21 fields for this company. Each filled one shows its value and, under it, the page it came from. Ten of the fourteen evidence fields are "Not Found" here, and the summary says plainly that no public link to either programme was found.

Writing down what it could prove and what it could not, instead of filling the gaps with plausible text.

A blank you can see is cheaper than a guess you have to catch. The verdict for this company is "uncertain", which is the right answer for a consultancy with no naval work on record.

The exact search queries for each field, the pages that were read for the answer, and every page considered, ranked Authoritative, Reliable, General or Low.

A precision pass first, then a broader one, then an exhaustive one with site-specific searches (the company's own site, SEDAR, the contract registers), stopping when the quality bar is met or the attempts run out.

You can audit the method, not just the result. A separate check then reviews the verdict and states whether it agrees with the scoring rules, and the fields are checked against each other for coherence.

Which engine asked for a human check and on which search, with three choices: open and solve it, say you solved it, or dismiss it.

Moving that search to the engines that are not blocked and carrying on. Nothing stalls.

Because the searching runs from your browser, a challenge is yours to answer if you want that engine back sooner. The run does not depend on it.

Filters that count the verdicts (confirmed clean, with caveats, uncertain, off-topic or rejected, failed, cancelled), three row densities, a column chooser, and rows that carry a score, a verdict, a one-line why, notes and a source link per filled cell.

Filling the table as each company finishes, and saving progress as it goes.

This is the deliverable: a supplier table you can sort, filter, check cell by cell, and hand to someone else.

Partway through this run: two companies judged off-topic (zero evidence fields filled, a structural mismatch) and several marked uncertain with four or five of fourteen evidence fields found.

Fix the verdict if you know better, re-run the flagged fields, teach a rule for the items still to come, or leave it as an honest gap.

The run marked Stopped with its elapsed time, the count of answered items, a CSV of them, and the note that any item can still be re-run from its card.

Closing the books: the first run's close-out recorded what the run actually used and what was charged, against the balance shown at the start.

Nothing you paid for disappears when you stop, and the charge is written down next to the usage.

3 confirmed clean, 18 with caveats, 95 uncertain, 29 a mismatch, 1 rejected and 3 off-topic. The rest of the board is the table itself, one card per company.

Few companies came back clean, because ten facts per company is a high bar. Cells such as Industrial and Technological Benefits commitments and programme role often read Not Found: the run did not find a page that said them, and the table says so instead of filling the gap.

The uncertain rows are still a lead list with receipts. Each one is a company with the facts that were found, the page each came from, and the gaps marked, ready for someone to check.