How do I get an AI summary that keeps every number, date, and exception intact?

October 16, 2026 · 5 min read

SentX SI

You don't get it from the summary alone. You get it by doing three things in order: pulling every exact detail out of the source into a numbered checklist before the model reads the document, asking for a structured summary that supports each main claim with a passage or page, and diffing the result against your checklist item by item. The first step is the one most people skip, and it's the one that changes the job from "is this summary right?" — a question no one can answer in a single glance — into a set of small yes-or-no checks you can actually finish.

The reason the discipline comes before the ask is blunt. SentX's own documentation for its research summarizer states that the response can omit or misread details, and that a summary can contain errors and does not replace the source. The model card adds that answers can contain incorrect facts, citations, or calculations, and that memory is selective — repeat instructions and facts that must remain exact. None of that means the summary is useless. It means the exact details are the part of the output you verify, and verifying works far better when you already know what the exact details are.

Source: AI Research Paper Summarizer — Summarize & Ask Questions | SentX SI; Victoria AI Model Card | Capabilities, Evidence and Limits | SentX SI

Build the checklist before the model reads

Scan the source once, yourself, and write down every detail that would change the meaning if it came back wrong or missing. In practice that is five classes: figures with their units, dates and periods, sample sizes and populations, thresholds and eligibility boundaries, and conditional clauses — anything built on except, unless, only if, or excluding. Copy the wording rather than paraphrasing it, and number each item. The numbers matter: they give the later diff a stable anchor, so item 4 is the same thing from start to finish.

Two rules keep the checklist honest. First, capture the qualifier with the figure. An average handling time of 3.4 hours is a different claim from 3.4 hours excluding customs holds, and the second half is usually the half that matters. Second, capture scope. If a figure covers only one region, one product line, or one fiscal year, the checklist entry says so, because a summary that quietly widens the scope is worse than one that leaves the figure out.

Ask for structure and support

Then make the request along the lines of the workflow SentX documents: the research question, the method, the main results, and the stated limitations, with the supporting passage or page for each important claim. The summarizer page ships an adaptable prompt that asks for exactly this shape, and the checklist is what you add on top of it. Two additions pay off. Tell the model which details must come back exact — your checklist, or a compressed version of it — because the model card's own instruction is that facts which must remain exact should be repeated. And ask it to separate what the source states from what it interprets, and to flag anything it cannot verify, rather than smoothing gaps into confident prose.

Source: AI Research Paper Summarizer — Summarize & Ask Questions | SentX SI; Victoria AI Model Card | Capabilities, Evidence and Limits | SentX SI

Diff item by item

Open the original and walk the checklist. For each item you land in one of three buckets. Present and correct: mark it done. Present but shifted — a unit changed, a date rounded, a threshold softened, a scope widened — correct it against the original, matching each surviving number to its unit, population, and scope, the comparison axes the documented workflow names. Absent, or not locatable in the source: that goes in an explicit could-not-verify note, which is the documented way to end a check, not a failure. When you send corrections back, batch them into one message with the exact source wording, and restate the correct values rather than pointing vaguely at that percentage. The repair loop consumes messages, and daily chat limits apply — three without an account, ten with a free account, two hundred on Plus — so one consolidated correction pass beats six fragments.

Source: AI Research Paper Summarizer — Summarize & Ask Questions | SentX SI; Free AI Chat Online — No Signup | SentX SI

A worked illustration

To see the mechanics, take a fictional one-page quarterly logistics report — invented, non-sensitive material, because submissions to the summarizer are not confidential and may enter training or shared memory. Your pre-scan checklist comes out like this:

Source: AI Research Paper Summarizer — Summarize & Ask Questions | SentX SI; Privacy Policy | SentX SI (sentx.ai)

1. On-time delivery 94.2% in Q3 2026, up from 91.8% in Q2
2. Average handling time 3.4 hours, excluding customs holds
3. Cost per parcel $1.87, down 6% year over year
4. Penalties apply only if the monthly on-time rate falls below 90% for two consecutive months
5. All figures exclude the Gulf region until the new hub opens in November 2026
6. Data covers July 1 to September 30, 2026

Now imagine the summary comes back. Item 1 survives, though the Q2 comparison rounds to 92% — confirmable, minor. Item 3 survives clean. Item 2 loses its qualifier and reads as a flat 3.4-hour average. Item 4 softens to penalties applying when performance misses target, dropping both the 90% floor and the two-month condition. Item 5 disappears, so the reader assumes the figures cover all regions. Item 6 compresses to the third quarter.

The diff closes with two items confirmed, three requiring correction or a flag, and one accepted with a note. One correction message carries items 2, 4, and 5 back with their exact wording. Nothing in this example is a measured result — no published count exists for how often a summarizer drops a given kind of detail, and the illustration shows review steps, not a generated answer. What it does show is where the risk tends to concentrate: headline metrics and clean percentages survive, while exceptions, conditionals, and scope exclusions sit in subordinate clauses and are the class most exposed to flattening. That is my reading of how compression works, not a finding — but the checklist catches those losses whichever class fails in your document, which is the point of building it before the model ever sees the page.

Source: AI Research Paper Summarizer — Summarize & Ask Questions | SentX SI

The payoff is not a perfect summary. It is that every number, date, and exception in the document ends in exactly one of three places: confirmed against the source, corrected with the discrepancy noted, or flagged as unverifiable. A reader who relies on the summary knows which of those it is — and that is a property no amount of prompting alone can supply, because the ground truth lives in the original, not in the model.