What we tested—and what this cannot prove
On 31 October 2026, we sent two fictional briefs through InriGPT’s purchased workflow API: an RFP response and a content repurposing task. Both used Cloudflare-hosted Qwen3 30B-A3B. We retained the returned text and reviewed it against the supplied source. The files below reproduce those inputs and captured outputs; this article was prepared on 11 November. These are the saved final verification responses after prompt and model changes, not first-attempt samples or a complete record of every development run.
This is a close reading of two outputs, not a model benchmark, an accuracy percentage or a customer case study. It does not measure time saved, certify a proposal, or establish which model is best. Generation settings, model updates and wording can change a response. Its practical purpose is to show a review method a small team can repeat with its own permitted examples.
Case one: the missing ISO certificate
The fictional buyer requires a 24-hour support rota and an ISO 9001 certificate. The supplier notes say an approved rota exists. They contain no certificate or audit evidence, and no delivery dates have been agreed. The requested output is a requirements table, a short response draft and next steps.
The captured output correctly leaves certification unconfirmed: “no certificate or audit evidence has been supplied”. That is a useful distinction. An absent attachment is not proof that the supplier lacks certification, and an assistant should not invent a certification project or promise to obtain one.
The rota row is more subtle. The input states that the rota is approved, but the response asks for its underlying documentation. That is a reasonable verification request if the buyer needs the actual document. It becomes confusing if the reader interprets it as saying no evidence was supplied at all. There is a supplied assertion; there is no inspected rota file. Those are different levels of evidence.
The edit we would make
The supplied notes state that an approved 24-hour rota exists. The rota itself was not included in this review. Confirm the document version, coverage and project availability before attaching it to the response.
For the certification row, retain a visible placeholder: [Needs evidence: current certificate or applicable audit evidence; owner to confirm]. Leave delivery dates unresolved until an accountable project lead agrees them. The reviewer should not convert the assistant’s suggested next steps into commitments made to the buyer.
A table is useful only when its status labels are precise. “Source says”, “document inspected” and “approved for this bid” should not be merged into one green tick. In our RFP evidence worksheet, each mandatory requirement starts unverified, so the person reviewing the bid must supply evidence and make that decision.
Case two: useful social copy with an unsupported guarantee
The fictional workshop source advises starting an AI pilot with one recurring task, approved documents and a human reviewer. It says to compare each draft with its source. It contains no measurements of cost, time or accuracy. We requested a newsletter introduction, three social posts and an article outline.
The output produced the requested formats and the carousel preserved the source’s concrete steps. However, one post said the approach “ensures accuracy and control”. A review process does not ensure accuracy. A reviewer can miss a mistake, a source can be obsolete, and a draft can introduce a plausible claim that was never supplied.
The edit we would make
Begin with one recurring task, approved source documents and a named reviewer. Compare each generated claim with its source before sharing the draft. Record unresolved questions rather than treating fluent wording as verification.
The article outline also proposed discussing “the role of source materials in AI training”. That changes the subject: the supplied notes concern reference material for checking a draft, not training a model. We would replace that item with “How to trace a draft’s claims to approved reference documents”. A plausible topic heading can be wrong even when it contains familiar technical words.
The newsletter’s general efficiency language also needs an editorial decision. It is not a measured result from this test. For a source-only draft, remove broad promises and lead with the actual process: choose a task, prepare sources, review claims. If you want to add outside evidence, read and cite a relevant primary source yourself rather than asking the model to invent an authoritative-looking reference.
A repeatable claim review in five steps
- Freeze the input. Save the exact brief and approved source version. Remove personal or confidential records before public testing.
- Save the first output. Keep failures as well as useful passages. If you regenerate, record a separate attempt rather than overwriting the evidence.
- Split claims. Highlight statements about credentials, dates, quantities, commitments and cause-and-effect. Review headings and suggested next steps too.
- Link evidence. Record the passage supporting each statement. Mark an inference as an inference. Use “not supplied” when the source is silent.
- Decide what to do. Retain, qualify, remove or ask the owner. Save the revised copy and reviewer’s decision with the source.
The CSV download contains four concrete review rows from these outputs. Add your own rows rather than treating it as an automatic grader. A blank evidence cell is a question for the reviewer, not permission to generate a better-sounding justification.
How to repeat this in InriGPT
Open AI tools after signing into a purchased workspace. Select Proposals & RFPs or Content repurposing, copy the corresponding input from the JSON file, and generate a draft. The public demo can be used for a shorter trial, but its response budget and shared limits differ. Compare your actual response with the source; do not expect identical wording.
Use Projects & approvals when a team needs a saved brief and a review decision. An approval records a person’s workflow decision, not a certification that every claim is true. Keep the review focused on what the source establishes. Two examples cannot settle performance for your business: create a small test set covering direct answers, missing facts, conflicting instructions and a task outside the tool’s scope before expanding use.
Corrections and related reading
Send the page URL and a reproducible example through Contact if an instruction or observation is wrong. Do not send private records or credentials. Our editorial policy explains the distinction between captured output, fictional practice material and customer results.