Evidence, findings and reports
Your agent writes the assessment. Ducket keeps the sources, saved work and mechanical checks. Argus by Ducket supplies dated market evidence.
Three different checks
- Source integrity: saved bytes, passages and dates match the evidence handle.
- Mechanical checks: identity, attribution, brief binding and required content have passed the applicable checks.
- Independent review: a separate substantive review actually occurred. Your agent's own self-review does not establish this.
A quotation can match a page while naming the wrong company, edition or claim. Matching bytes alone establishes no funding, buyer, availability or intent claim. Uploaded organizer text remains attributed to you; it cannot impersonate a server-fetched page.
Facts and unknowns
Facts link to saved company-attributed passages. Inferences are labelled readings of evidence. Unknowns record what the searches did not establish. A failed tool means checking could not finish. Neither an unknown nor a technical failure is a confident rejection.
Read the assessment
Native research reports use these sections: Fit thesis, Signals, Deal recommendation, Risks and Contacts. The outreach draft is separate. A complete assessment may recommend against approaching the company. Missing or malformed work cannot become a completed report.
Resume without repeating purchases
Connected company reports use Commercial fit, Relevant buyers, Participation angle, Supporting sources, Uncertainties and Next steps. Open View report on the company, or View draft for partial writing. Continue with my agent preserves saved versions. This writing uses bounded connected access separately from native paid research and does not establish qualification or inventory.
Give a new conversation your saved session reference. Ask it to resume, then recover saved operations and evidence handles through source_operations, following next_offset. Recovery reads the saved answers without another supplier request. Uncertain purchases still need reconciliation.
Old rosters can support dated history. They cannot prove current sponsorship or inventory. An unknown finding cannot support a categorical rejection. Supported month dates need no invented day.
If qualification or its evidence changes, the workspace shows Reassessment needed. The old report remains history. Ask your agent to reassess using the current evidence, update the research and save a new report. Requalification alone leaves earlier conclusions stale.
Return to the company in Argus or give its address to a fresh connected conversation. Saved reads remain available. A stale revision requires a fresh read before editing. An uncertain supplier request needs reconciliation; the same request under a new ID is also refused until reconciliation.
See Troubleshooting and the tool reference. For the separate dated market layer, see Argus signals.
The corpus-only examples below use the separate Argus read connector. Source qualification and research use the Source connector described above. Existing market-reading prompts remain available.
What you can ask
Argus by Ducket reads event pages and keeps what they said: which companies sponsored which event, in which year, and in which role. Once your assistant is connected you can ask it those questions the way you would ask a colleague. Who was on last year's list and is missing from this year's. Who has started turning up in a market. Whether a claim in a media pack matches what the page actually said. How much of a market the record covers before you lean on it. Every answer comes back with the page it was read from and the date it was read, so you can check it yourself before it goes into a deck, a call, or an email.
Copy-paste example prompts
These are written to paste as they are. Swap the vertical, the event name, or the year for your own. Where a question calls for a judgement (which of these lapsed sponsors is worth a call this week), ask your assistant for it in the same breath: Argus supplies the receipts, the model supplies the opinion, attributed to itself. Argus never ranks companies for your event.
In Claude
I run a mid-size cybersecurity conference in Singapore. Using Argus, list companies that sponsored a comparable APAC cybersecurity event in the last two years and are missing from that event's most recent roster on file. Give me the event, the edition each company last appeared in, and the source page.
You get the lapsed companies with the edition each last appeared in and the dated page behind every row, and it will decline to say which of them suits your event, because Argus holds no view on your event.
Using Argus, which companies made their first appearance as a sponsor at an APAC fintech event in 2026? Group them by what they sell and cite the first-appearance page and date for each.
You get the entrant cohort with each company's dated first appearance, and the grouping by category is your assistant's own work rather than a corpus fact.
A media pack claims Acme Corp was a gold sponsor of GovWare 2025. Check that against Argus and tell me what the corpus actually holds for that edition, with the page and the date.
You get the receipted rows for that edition or a plain statement that the corpus holds none, and an absent row means the capture did not show it rather than that the claim is false.
Argus recorded a sponsor leaving an event I follow. Pull the receipts: the earlier page that listed them, the later page that no longer does, and the observed date on both.
You get both captures with their dates and the roster row that changed between them, and the reason for the departure is nowhere in the corpus, so any explanation is your assistant's own read.
Before I put any of this in a pitch, ask Argus how much of the APAC cybersecurity market it actually covers: how many events are tracked, how many carry a roster, and how fresh the newest capture is.
You get the coverage census with the date of the newest capture, and a thin count is a statement about what the corpus has seen rather than a claim about the market.
Give me everything Argus holds on Palo Alto Networks: which events, which editions, which role on each roster, with the source page and observed date on every row.
You get that company's receipted history at role grain, so an exhibitor row is never folded into a sponsor row, and anything the corpus never captured comes back as absent rather than as a guess.
How many editions of GovWare does Argus have on file, and how many companies were on the roster in each one? Flag any edition where the capture was too thin to compare with the one before it.
You get the per-edition roster counts with the dated capture behind each and a coverage note where an edition is too thin to compare, and roster size is an observed count on file that says nothing about how good an event is.
Pull the Argus receipts for APAC cybersecurity sponsorship over the last two years, then give me your own reading of what they suggest. Keep the two apart and label which sentences rest on receipts and which are yours.
You get the receipted rows first and the assistant's interpretation second, marked as its own: Argus supplies the receipts, the model supplies the opinion, attributed to itself.
In ChatGPT
Use the Argus connector to find APAC developer-tools events with at least two editions on file, and summarise who sponsored each edition, with the source page and date.
ChatGPT reads the corpus through its deep research pair, so results arrive as event records carrying their receipts, and a short list means the first page of matching rows rather than everything that matches.
Using Argus, find companies that sponsored an APAC retail-technology event in 2024 or 2025 and do not appear on that event's latest roster on file. Give me the last edition each appeared in and the page it was read from.
You get the lapsed rows with dates, and it will decline to order them by how likely they are to buy, because that judgement belongs to you and your assistant rather than to the corpus.
Which companies show a first sponsor appearance at an APAC logistics event during 2026 in Argus? Cite the edition and the dated page for each.
You get the entrant cohort with dated first appearances, and a first appearance is the earliest edition on file, so it means new to the corpus rather than new to the market.
Which tracked APAC cybersecurity events have the largest rosters in Argus, and what is the count for each edition?
Largest here means the most sponsors observed on file, which is a count of what one capture showed rather than a verdict on an event, and every count arrives with its edition and its dated page.
Our client says they held a title role at Cybersec Asia in 2024. Check Argus and tell me exactly what the corpus recorded for them that year, with the page and the date.
You get the role the corpus recorded for that edition or a plain statement that it holds no row, and roles are read from the page as published, so an exhibitor row and a title row are never folded together.
Show me every sponsor departure Argus recorded in the APAC cloud-infrastructure market in the last six months, with the earlier page, the later page, and both observed dates.
You get the dated change rows with both captures, and why a company left is nowhere in the corpus, so treat any reason your assistant offers as its own.
Before I quote any of this to a client, ask Argus what it does not cover in APAC fintech: how many tracked events carry no roster, and how old the oldest capture in that slice is.
You get the gaps stated plainly with dates, and a gap describes the corpus rather than the market, which the answer should say out loud.
Which companies appear on the same rosters as Fortinet across the APAC events Argus tracks, and how many shared editions does each have?
You get the shared-edition counts with the receipts behind them, on a full-catalog key, and the count is an observed overlap on file, so reading a competitive story into it is your assistant's work.
Connect your assistant
The examples in this section require the corpus-only Argus connection and an existing Argus license. Client and workspace support still applies. Choose the door your client supports.
- claude.ai and ChatGPT. Add a custom connector by URL, using https://api.ducket.io/mcp, sign in to Argus, and approve one screen. There is no key to paste.
- Claude Desktop, Claude Code, Codex CLI, Gemini CLI. Add the same URL as a remote connector, or run the local server with an API key from the Organisation tab in Settings. In Claude Code that is one command:
claude mcp add ducket-argus --env DUCKET_API_KEY=dk_example_9hK2mQ4vX8wL1pR7 -- python -m ducket.mcp
Access is read only and bound to one organisation: the assistant reads exactly the slice your organisation is licensed for and can see nothing outside the corpus, including your account settings and billing. Remove the connector in your client to disconnect, or revoke the connector key on the Organisation page to cut every connection riding it.
Optional workflow Preview.
The optional Preview workflow is selected by Ducket for named accounts. A connection alone keeps Current. After a version change, refresh capabilities or reconnect. Your saved event, reports and consumed allowance remain attached.
For a requested contact search, your assistant must inspect its actual tools and permissions. A supplier name does not prove a working connection. ChatGPT and Claude control their own connector setup and approval steps. Argus receives no sibling connector tokens. If setup fails, continue the company assessment or an exploratory unsent approach and recover the existing work before trying another action.
External charges remain with your connected account under existing purchase authority. A matching APAC role is a contact hypothesis. It does not verify budget ownership or a warm introduction. Keep private CRM and contact databases outside Argus. Team handoffs use deliberately shared reports after recipient access is checked.
What your assistant is handed
Your assistant is handed the record and nothing else. Each answer carries the company, the event, the year, the role and tier the page used, the page it was read from and the date it was read. Bookkeeping we keep for ourselves, like the internal number of a page capture, does not go to your assistant, because it means nothing outside our own system and would only give a model something to guess at.
The words in the record were written by other people, on their own event pages. So every result now arrives with a short note telling your assistant to read that text as a quote and never as an instruction, and the text itself is cleaned of invisible characters before it is handed over. A sponsor page can still contain misleading text. Your assistant must treat it as untrusted quoted data. Real names are untouched: Japanese, Korean, Thai, Chinese and accented spellings come through exactly as published, because a name spelled wrong in a pitch is its own kind of error.
The technical reference
Source workflow setup lives on Connect your agent. The separate Argus contract is documented in the Argus API reference. Read that page when you are wiring Argus into software of your own rather than asking questions in a chat window.