Hallucination and the Audit Trail — Verify Everything That Moves Money

Working with people, confidence is genuine information. An analyst who knows a figure states it flatly. One who is guessing slows down, hedges, or offers to go and check. That signal lets you triage — you verify the shaky claims and take the rest on reasonable trust, which is the only reason working with other people is efficient at all.

None of that machinery exists here. The model has no internal quantity measuring whether a statement matches reality, so it has nothing to hedge with and no reason to slow down. What you get is a uniform surface: the same assurance, the same supporting detail, the same willingness to elaborate when you ask a follow-up, whether the number was lifted from the page you pasted or assembled out of how such numbers usually read.

Asking directly makes it worse rather than better. Put the question "how confident are you in that figure" and you receive another generated sentence — the text that plausibly follows a question about confidence. It is not a reading from a gauge. A model that says it is highly confident and a model that says it is uncertain are both producing prose, and neither statement has been checked against anything.

There is a specific reason the fabrications read so well, and it is worth sitting with. Invention is unconstrained. When the model is working from a real page it is bound by awkward, specific, badly-worded source material. When it is constructing, nothing pushes back, so the output comes out clean, well-proportioned and exactly in the register of good analysis. The smoothest paragraph in a brief is a fair place to start checking, not a fair place to stop.

Invention is not evenly distributed. It clusters in predictable places, and knowing them turns an impossible instruction — check everything — into a short list you can actually run. The pattern is consistent: the risk is highest wherever the answer has a precise, conventional shape that the model can produce from form alone, and where the underlying fact was not in front of it.

A ticker is a good illustration. It has a shape. Four to twelve capital letters, often an abbreviation of the company name, often with a familiar suffix. A model can generate something with exactly that shape for a company it has no listing information about, and the result is indistinguishable from a real NSE symbol until you look it up. The same is true of an ISIN, a circular number, a section reference, or a page citation. Structure is easy to imitate; correspondence to reality is the part that is missing.

Numbers behave the same way but with an extra trap. Arithmetic on figures in front of the model is generally reliable — that is not where the failure is. The failure is a ratio computed from one number you pasted and one recalled from training. The maths is perfect, the format is correct, and the answer is wrong in a way no amount of re-reading will reveal. This is why every computed figure has to arrive with both inputs and both references shown, so you check the ingredients rather than the sum.

The regulatory zone deserves separate mention because the consequences differ. Ask about a SEBI requirement, a margin rule, a disclosure threshold, a settlement timeline or an exchange circular and you will get a specific, confident, plausible answer. Regulation changes, phases in, and applies differently by segment, and a model has no idea what is currently in force. For anything regulatory, the correct source is the SEBI or exchange website, and the model's role is at most to help you phrase the question you go and look up.

A verification procedure that takes an hour will not survive a busy quarter. It will be done properly twice, skipped on a Thursday evening, and then quietly abandoned. So the design constraint is not thoroughness. It is that the procedure must be short enough to run every single time, including on the day you are tired and the market has been unkind.

The trick that makes it short is scope. You are not verifying the brief. You are verifying the handful of lines in it that could move a decision — typically four or five in a page of output, and often fewer. Everything else is structure, framing and connective prose, which cannot be wrong in a way that costs you money because you were never going to act on it.

Run the five steps below in order. The order matters: the cheapest and most decisive check comes first, because a failed quote search invalidates everything around it and saves you the other four. Most of the time the whole thing genuinely takes under a minute, and on the occasions it takes longer, that is because it found something.

One rule holds the procedure together, and it is the one people abandon first. A claim that fails verification gets deleted, not softened. The temptation is to keep it with a note that it needs checking, and three weeks later that hedge has quietly evaporated and the claim is sitting in your notes reading like everything else. Unverified material does not get to live in the same file as verified material.

You cannot prompt a model into being reliable. What you can do is prompt it into producing output whose reliability is cheap to test, and that is a different and entirely achievable goal. The difference between a brief that takes forty minutes to check and one that takes forty seconds is decided before the model answers, by what you asked it to include.

Three requirements do most of the work. Every claim carries a verbatim quote from the source, which converts verification into a text search. Every claim carries a page or note reference, which tells you where to look when the quote needs context. And an explicit instruction that "not stated in the source" is a correct, expected answer — because without it the model treats a gap as something to fill, and gaps are precisely where invention lives.

That last one is the least intuitive and the most valuable. A model asked seven questions about a section that only supports three of them will, left to itself, produce seven answers. Told plainly that absence is a valid answer, it will more often report the absence. You are not making it more truthful; you are removing the pressure to complete a pattern.

The prompt below is built for a claim you are about to rely on. Run it as a separate pass over output you already have, in a fresh thread, with the source text pasted again. A fresh thread matters: asked to re-check its own answer inside the same conversation, a model has the earlier answer in front of it as context, and consistency with that context is easier to produce than a genuine re-reading.

Verification is what you do in the moment. The audit trail is what survives it, and it is the part almost everyone skips, because in the moment the verification feels sufficient. It is not. Six months later the memory of having checked something is indistinguishable from the memory of having read it, and you cannot tell which lines in an old note were confirmed and which were absorbed.

The record is three fields per line. The claim, in one sentence, with its unit and period inside it. The source, precisely enough to reopen — document, period, page or note. And the date you checked it, which is not the same as the date the document was published. That is the whole format. It fits in a plain text file, a spreadsheet, or a note in whatever you already use, and it takes a few seconds a line.

The date-checked field earns its place more often than the other two. Figures get restated in later filings. Interpretations that were correct against one quarter's disclosure stop being correct against the next. Knowing that you last confirmed a line in a particular month tells you whether it needs revisiting — and a claim you cannot date is a claim you have to re-verify from scratch or drop entirely.

The figure below is the point in one image: the chain runs from the claim to the source to the date checked, and it only holds if every link is present. Break the source link and the chain does not weaken, it collapses — the claim reads exactly the same, and it can no longer be defended by anybody, including you.

And that last clause is the real argument. This is not paperwork for a regulator. It is the mechanism by which you can update. A claim with a live source can be reopened when new information arrives and either confirmed or dropped. A claim without one has no route back, so it cannot be tested, only defended — and beliefs that cannot be tested are the ones that quietly do the damage in a portfolio.

Two things make this sharper here than in most other places you might use these tools. The first is that the claims cost money. A misread margin, a wrong basis, a segment figure taken as company-wide — each of these can support a position you would not otherwise have taken, and the error only surfaces when the market has already priced the truth. Almost no other everyday use of AI has that property.

The second is regulatory. Research published or circulated in India by a SEBI Registered Research Analyst carries record-keeping and disclosure obligations, and the underlying expectation is that a recommendation has a documented basis. An AI-drafted paragraph with no traceable source does not meet that standard, and cannot be made to meet it retrospectively. If you publish research, run a group, or advise anyone in any capacity, the audit trail is not optional good practice.

It is worth being blunt about where responsibility sits. There is no arrangement in which the tool is accountable for the output. A model is a drafting instrument, and the analyst whose name is on the note owns every figure in it, exactly as they would if a junior had typed it. The article on guardrails and the human in the loop in this module goes further into what that means in practice.

None of this is an argument against using AI for research. Compression, structuring and consistent checklists are genuine gains, and they hold up. The argument is narrower: the verification cannot be delegated to the same system that produced the draft, and it cannot be delegated to your impression of how the draft reads. It has to be a procedure, run against the source, recorded with a date. That is the whole discipline, and it is short enough to keep.