AI portfolio reviews — Bitnora docs

What a review covers

A review reads your whole book at once and returns a written report, not a score card. It is assembled in three passes so that each one reasons over the previous one’s conclusions instead of competing with them: first every holding is assessed on its own, then the portfolio is described as a whole, then what to do about it.

Every holding, classified

Each position is placed in one of nine categories — core index, compounder, dividend income, speculative, redundant, thesis weakening, hedge, cash-like, or unclassified — with the reasoning tied to your book’s own numbers. A reason that would read the same for anybody holding that ticker is not accepted as a reason.

Six scores, each explained

Overall, diversification, quality, risk management, income reliability, and alignment to your targets, each from 0 to 10 with an explanation naming what moved it and a confidence level. Where you have set no targets, alignment says so and scores at low confidence rather than inventing a standard to judge you against.

Allocation, concentration, and overlap

Findings about how the book is distributed, what is quietly running it, and where two funds duplicate each other. Overlap is computed deterministically before the model sees it, and it is labelled with its fidelity: company-level overlap means the funds genuinely hold the same names, while sector-level overlap means only that their sector mixes coincide — two funds can be identical by sector and share no company at all. The report is required to say which one it is looking at.

Risks, opportunities, and ranked recommendations

Each recommendation carries an action, why it applies, what it is expected to do, what it risks, a confidence level, and its invalidation triggers — what would have to be observed for it to stop applying. That last field is what makes a recommendation checkable months later rather than merely persuasive.

What it is reviewing for

The reviewer works for you and for nobody selling you anything, and it reviews against long-term compounding: own good assets for years, keep turnover and costs low, stay diversified enough that no single mistake is fatal.

It does not time markets and does not predict prices. An argument that depends on knowing where a security goes next is a guess wearing a review’s clothes, and it is instructed to say so rather than produce one.

The same discipline applies to length: a handful of high-conviction points beats a long list of things technically worth mentioning, and a long list of recommendations is treated as a way of avoiding a judgement.

Running one

Step 01
Open Reviews and choose a scope
A single portfolio, or all of them together. Reviewing everything at once is what surfaces duplication between sleeves you think of separately.
Step 02
Start the review and leave it
A progress checklist shows which of the three passes has landed. You do not have to watch; the report is waiting when you come back.
Step 03
Read the caveats before the recommendations
They tell you how much of the report is standing on solid ground. The caveats footer is always visible on a finished report, never behind a disclosure.

How long: Two to three minutes for a typical book.
How many: Three per day by default, resetting at midnight UTC. Each review is a real model run over your whole book, so the limit is a cap rather than a meter — you are not billed per review.
One at a time: A second review is refused while one is in flight. Open the running one instead.
If it fails: A failed review can be retried, and a retry keeps the passes that already landed rather than starting over.

What “data unavailable” means

Every number in a review has to come from the data pack assembled for it. The reviewer is not allowed to estimate a figure to fill a gap, to bring in something it happens to know about a ticker, or to restate one of the pack’s numbers in units the pack did not use. Where something is missing, it has to say the data is unavailable and reason around the hole.

The pack carries an explicit coverage block naming its own gaps, and every gap in it must reach the report’s caveats list in plain language. Any judgement leaning on missing data has to carry lower confidence because of it. Typical caveats:

Fundamentals were unavailable for some holdings, so quality judgements about them lean on less.
Overlap could only be measured at sector level for a fund, so its figures describe overlapping exposure rather than shared holdings.
A position had no current price, so it is absent from the weight calculations.
No targets exist, so alignment could not be scored against anything.

A review that reads as complete while sitting on an incomplete pack is the failure that matters most here — which is why confidence is a statement about evidence rather than about conviction. High confidence means the pack alone would persuade a sceptical reader.