The method, including the limits.

Data we read

We read public primary sources and public market data:

• Annual and quarterly reports, current reports and other public filings.
• Company announcements and investor-relations disclosures.
• Historical price and volume series.
• Public sector and industry classifications.

We do not use paywalled research, non-public information or purchased data sets that cannot be cited. If a source cannot be identified and dated, the number does not go on the page.

What the model does

The pipeline is deliberately simple. It ingests a source, normalises the sections it can identify, compares the document with the prior readout, and extracts the passages that changed. A language model then writes a plain-language digest of those passages. The source reference travels with the digest.

The output is a dated readout: what changed, where it changed, and what the change might mean in the context of the document. The model does not score the company, rank the security or set a target.

We do not publish an accuracy rate. The extraction can miss context, and the summary can be wrong. The validation path below exists because that is true.

What it does not do

The boundary is part of the product, not a disclaimer added at the end:

• It does not give price objectives, ratings or buy, sell or hold calls.
• It does not give personalised advice or manage money.
• It does not make market-direction calls or performance projections.
• It does not claim to be real-time or to replace the underlying filing.
• It does not treat a price or volume pattern as a signal.

Validation

A sample of readouts is reviewed by a person against the cited pages. The reviewer checks that the quoted passage exists, that the date and accession number are correct, and that the digest does not add a conclusion the source does not support.

If you find an error, email contact@heritagebistro.net with the ticker, the readout date and the source page. We log the report, correct the readout where the source supports a correction, and keep the corrected version in the dated archive. A correction is not a promise that every other output is free of error.

Known limitations

• Changing definitions. A company can change how it defines a segment, a metric or a fiscal period. A like-for-like comparison may not be possible.
• One-time items. Restructuring charges, impairments and other one-off items can make a reported change look larger or smaller than the underlying trend.
• Halts and low liquidity. A halted or thinly traded name may have a price series that does not describe a continuously available market.
• Source delay and restatement. A filing can be amended, a data feed can arrive late, and a later source can contradict an earlier one.
• Document structure. Tables, footnotes and scanned pages can be misread. The source link is the check; the digest is not a substitute for it.
• Peer sets. A peer group is a comparison choice, not a fact about the company. A different set can change the context.

No performance claims

We do not publish performance figures, a track record or a comparison against any benchmark. We do not claim that a readout would have improved an investment result. The sample on this site is a format demonstration, not a historical recommendation.

Past performance does not guarantee future results, and a historical price or volume pattern does not tell you what will happen next.