AI Rewrites Your Research Note Summary: How to Keep the Numbers Intact

It is 6:40 on results day. The conclusion of a quarterly note has to come down to 300 words for the sales desk and a handful of client chats. The original runs 600 words with three tables attached. You want it shorter and smoother, nothing more.

You paste it in and add one line: change the wording only, leave the numbers alone.

Thirty seconds later it hands back a draft. The sentences are tighter, the repetition is gone, "we" has become "the company." It reads better than the original. You drop it into the email and send it to thirty people.

It did not miscalculate anything. The problem is the instruction. You asked it to change the wording, and in its reading, how a number is written is part of the wording.

All companies, dialogues and figures in this piece are constructed examples. They do not correspond to any real security, research note or disclosure.

Start with one rewrite: eight differences, not one typo

This is the only example in the piece. Every section returns to it.

Exhibit A｜Conclusion of a quarterly note, original Final, 29 Oct 2026, 08:12｜version 3｜HuaChen Precision (ticker 688XXX, constructed)

HuaChen Precision reported 9M26 revenue of RMB 4.26bn, up 18.3% YoY, and net profit attributable to shareholders of RMB 512m, up 12.7% YoY, slower than revenue growth, mainly on a 1.8 percentage point decline in gross margin. Q3 revenue was RMB 1.51bn, up 9.4% YoY and down 3.2% QoQ. We cut our 2026E net profit forecast to RMB 690m (from RMB 740m), implying 21x 2026E P/E. Maintain Overweight. Risks: weaker downstream capex; raw material price volatility.

Exhibit B｜The rewrite HuaChen Precision's 3Q26 report shows 9M26 revenue of RMB 4.26bn, up 18.3% YoY, with net profit attributable to shareholders of RMB 520m, up 12.8%. Revenue growth held up while profit growth lagged slightly, as gross margin fell 1.8% YoY. Q3 revenue was RMB 1.51bn, up 9.4% YoY and down modestly QoQ. On this basis we cut our 2026E net profit forecast to RMB 690m, implying a P/E of about 21x. Risks: weaker downstream capex; raw material price volatility. Maintain Overweight.

Eight changes. Not one typo. The sentences read better than the original.

Some of that smoothness has a cause. The original carries causality, qualifiers and comparisons inside its sentences, which makes it heavy. The rewrite strips those parts out and the sentences turn light. Part of the ease you feel is the thing that was removed.

Send this draft and the client comes away with three facts. Profit up 12.8%. Gross margin down 1.8%. A modest sequential decline.

The first figure is RMB 8m too high. The second states one-third of the actual decline. The third has no size at all.

Four kinds of change, ordered by how hard they are to catch

These eight changes are four problems. What separates them is how much work it takes to notice one after it has happened.

In six of the eight, the number itself survived. What moved was everything around it: the decimal place, the unit, the bracket, the adjective. Rewriting wants evenness, and evenness and accuracy run on different tracks.

1. Precision: one rounding step, and the growth rate follows

Why it happens. Beside 4.26bn, which sits on a 10m grid, 512m falls between the lines. The rewrite rounds it to 520m. Once that happens, 12.7% becomes 12.8%, and the two figures now look like they belong to the same dataset.

It did not recompute. It aligned. Alignment and accuracy are separate things. This one is easy to miss because 520m and 4.26bn look like one series, while 512m does not.

There is a related trap in the same family: earnings per share. If the original states that EPS uses a weighted average share count, the rewrite may treat that qualifier as decoration and drop it. What is left is an EPS with no stated basis. Recompute it on the latest share count and it can move by a tenth.

How to check. Compare every figure against its counterpart in the source, digit by digit. A figure that has been rounded is an edit, and the eye can do this without going back to the model.

The second layer is the derived set: any ratio built from a stated figure. Growth rates, margins, EPS, ROE. Go back to the source for all of them at once. In a research note those figures usually sit within a paragraph of each other, so one pass covers the group.

1. Units and basis: the same 1.8, two different declines

Why it happens. A percentage point and a percent are both written "1.8". The only difference is the word after the number. The rewrite reads sentences, not dimensions, and deleting "percentage point" leaves nothing missing as far as the sentence is concerned.

Basis means the conditions under which a figure holds: unit, period, denominator. Take any of the three away and the number is still there while the meaning is gone.

The consequence is not small. Gross margin of 30.2% falling to 28.4% is a decline of 1.8 percentage points, or about 6.0% in relative terms. Written as "fell 1.8% YoY," the reader receives one-third of the real decline.

Two more places this shows up. The first is single quarter versus cumulative. A 9M report carries both figures, so the rewrite merges them into one clause to save space. The reader then cannot tell whether RMB 1.51bn covers three months or nine.

The second is a lost forecast year. A P/E of 21x is 21x against 2026E. Remove the year and that multiple can end up applied to 2027 EPS.

How to check. Attach three labels to every figure: unit, period, denominator year. Unit means RMB, percent or percentage point. Period means single quarter or cumulative, YoY or QoQ. Denominator year means which year's earnings the multiple sits on.

If any one is missing, go back to the source.

Labelling is faster than reading because the fields are fixed. You are not deciding which sentence looks suspicious, and that is where the prompt template in Section 3 begins.

1. Comparison base: the prior value in brackets is the weight of the sentence

Why it happens. "(from RMB 740m)" reads like an aside. Take it out and the sentence gets cleaner, so the rewrite removes it. But the whole weight of the word "cut" lives in that gap.

Leave only "cut to RMB 690m" and the reader knows the direction, not the size. Whether that is a 6.8% cut or a 1% cut depends entirely on a number that is no longer on the page.

What these deletions share is that the sentence improves. That is why they are not read as errors. They are simply passed over.

It is the same mechanism as the few words that get dropped in compression, applied to a different document. Rewriting optimises for flow. What it leaves behind is not its concern.

How to check. Wherever cut, raised, maintained or initiated appears, two figures must follow: the new one and the old one. Only one figure means going back to the source.

Ratings and target prices behave the same way. "Maintain Overweight" is only meaningful if you can state the previous rating. "Cut target price" is only meaningful with the previous target.

One more item is easy to miss: consensus has versions. A summary that cites "market expectations" has to say which day's vintage it came from, because two vintages can be a whole note apart.

1. Magnitude and position: every sentence survives, the paragraph does not

Why it happens. Rewriting carries an unwritten objective: make the paragraph feel balanced. So "down 3.2% QoQ" becomes "down modestly", the causal clause behind slower profit growth collapses into "lagged slightly", and the rating slides from the first line to the last.

Each move holds up alone. "Modestly" is not wrong, and neither is "lagged slightly". The rating reads fine wherever it sits. Together they produce a different note: pressure flattened, causality removed, and the reader only reaches the stance at the end.

Research summaries lead with the conclusion by convention. Move the rating to the final line and the summary does half its job.

There is one more trap specific to this industry. Rating scales are not interchangeable. Buy, Overweight, Outperform and Accumulate sit on different ladders at different houses. A rewrite chasing consistency will happily substitute one for another, and those words are not substitutes in anything that leaves the building.

How to check. Two things first: is the conclusion in the first sentence, and is the rating word identical to the source. Both are visible at a glance, no reading required.

Then convert the adjectives back into numbers. Every "modest", "slight", "notable" or "significant" has to map onto a figure in the original. If it does not, the magnitude has been converted into mood.

Self-regulation has already reached this point. In October 2026 the Securities Association of China circulated a report on the research business. It asked firms to build an internal review mechanism for AI-generated content and barred analysts from relying directly on unverified output (per Securities Times, October 2026).

The requirement binds the signing analyst and the firm; it says nothing about the tool. This piece describes how to review, and is not compliance advice.

Three steps: freeze the fields, then rewrite, then check only the differences

The four categories share one origin. They all happen at the moment a field leaves its sentence. So invert the order: take the fields out first, then let the rewrite work on what remains.

Step one: pull the fields that cannot move into a lock table. Six fields, fixed: value, unit, period, action, prior value, rating. Filling it in is mechanical and takes two minutes. Once the table exists, those six items are out of scope, kept only as a reference for the check.

The value lies in what the table replaces. "Don't change the numbers" becomes six empty cells. A prohibition asks the model to draw its own boundary. An empty cell does not.

Step two: hand over the movable part and the fixed part separately. Do not write "don't touch the numbers". A negative instruction hands the model a boundary to interpret. Draw the boundary positively instead: four categories may change, everything else is copied.

The template below can be used as is. Replace what is inside the braces.

Step three: check differences only, do not re-read the whole thing. Compare the rewrite with the original sentence by sentence, and work only on the sentences that differ. Differences come in two kinds. Wording changed and the fields did not: let it pass. A field changed: put the original back.

Do not send the corrected draft back for "one more pass". A second pass introduces new edits, and the version you just checked is no longer the same document. This is the one step of the three that cannot be handed off.

If you have the conclusion and the corresponding financial summary table, AutoClaw can take the first two steps. It reads both together, compresses to your template, and tags every figure back to its position in the source. What comes back is a summary draft and a six-field cross-check table; the difference check stays with you. Confirm the rating and the forecast against the latest version, and confirm that the causal wording is the note's own judgement.

Then run five checks.

1. Is every figure in the summary written to the same precision as the source?
2. Does every figure still have its unit, period and denominator year beside it?
3. Does every cut, raise or maintain carry both the new value and the prior value?
4. Is the conclusion in the first sentence, and is the rating wording identical?
5. Does every "modest", "slight" or "notable" map onto a figure in the source?

All five pass, the summary can go into the workflow. One fails, hold it: go back to the source, restore that one field, and run the pass again.

Where this process should not run

When the source has no conclusion, do not ask it to produce one. Leading with a conclusion requires a conclusion to exist. Some notes are pure data and inference, with the view scattered through the middle or left unwritten. Asking for a conclusion in that case means asking for a judgement the note never made.

Rating changes, target price changes and initiations: copy, do not compress. These three go through compliance and sign-off before they leave, and a single word can change their meaning. The summary still gets written, but these sentences are lifted verbatim.

When more than one version of the material exists, freeze the version first. Quarterly notes get corrections and financials get restated. The lock table belongs to one version, and a new version means filling it in again. Without a fixed version the check in step three degrades into a single question: which version do you believe.

AI can restate what has already been written. It cannot confirm that the statement still holds. What goes out carries your name, and whether a field moved is known only to the person who checked.