An academic research note on whether company-specific fundamental knowledge retains value when machine-generated breadth enters the experiment.
The useful signal is rarely found in a single headline metric. It lives in the choices a company makes, the language it repeats and the small disclosures that explain how an industry actually works. This project treats that specificity as a research variable.
Two artifacts, one controlled comparison: does domain-specific fundamental research retain an advantage when the same question is approached with a general-purpose machine system?
Built without AI: supply-chain and holding-structure graph research, then trained into networks built for this one problem.
An ensemble of Qwen models adapted with supervised fine-tuning (SFT) and LoRA on twenty-five years of publicly sourced research material.
A · random pork oinks · without AI
Research is assembled as a chain of evidence, not a pile of excerpts.
Track B counts documents. The question here is what each class is read for, and when it arrives — the rhythm the filing calendar sets for anyone working through a company by hand.
Gather the company’s own vocabulary across filings, notes and reporting periods.
Place each disclosure inside the operating structure of its industry.
Turn connected observations into a view that can be inspected and challenged.
Recurring definitions, accounting policies and changes in presentation become the first map of what deserves attention.
Filter the hand-made graph by relationship type, then select a company to see the relationships it is party to and the filing each one was read from.
Choose a company below, then follow one finding from the filed sentence to its industry implication. All three are quoted verbatim from the filings named.
Effect disclosed / $3.9bn less depreciation, $3.0bn more net income
The same fleet, the same cash, and $3.0 billion more net income — $0.24 a share — because a service life was reassessed. The filing quantifies it, which means the earnings line cannot be read without the policy behind it.
When hyperscalers extend server lives within a few years of one another, the question stops being accounting and becomes whether the hardware genuinely lasts longer — answerable only by reading the same note across the peer group.
A · random pork oinks · without AI
Track A does not stop at the reading. What the framework produces becomes labelled data, and that data trains networks built for this one problem — no language model, no pretrained weights, and not one word of an example written rather than filed.
Filings and calls are what the company states about itself. Published analyst forecasts are what informed outsiders expect of it, on earnings and on the line items underneath. Corporate bond markets are what capital charges the same company to stay funded, moving continuously in a way no filing calendar does. They overlap less than the list suggests: four of the seven questions below have exactly one source that can answer them at all.
Not synthetic text. Each example begins at a real earnings event, matched to the point-in-time 8-K Item 2.02 or 6-K that carried it, then sanitized and identity-masked. Every sentence is ranked for guidance, financial metrics, comparisons and numeric evidence; ten are kept, held in source order and capped at thirty-two words each. Nothing is paraphrased and no language model is involved — nothing reaches an example that was not filed.
An extract on its own teaches nothing. Each is paired with a reconstructed earnings surprise — reported against the consensus as it stood before the release, not as later restated — and with the next-day abnormal return, the move left once the market’s own move is taken out.
One event, extracted and labelled, is a row; more than 200,000 of them make the corpus the networks are trained on. They do not arrive evenly. Filers cluster into four reporting seasons a few weeks after each quarter end and thin to a trickle between them, and a corpus built from real events inherits that shape rather than smoothing it away. Every mark is one example.
Which architecture, how much history each example sees, how hard to regularise — none of it is obvious in advance for a problem this small and this noisy. Each faint path is one configuration evaluated; the marked path is the one that survived.
Building the corpus out of raw earnings releases, and training the networks on it. Run on rented neocloud capacity rather than on one machine, and none of it spent on a general-purpose model.
B · pork with memories · fine-tuned model condition
Three stages: the corpus the models read, the supervised fine-tuning that adapts them, and the synthesis that brings their outputs into one comparison. The middle stage is where the work sits.
Twenty-five years of publicly sourced material across three document types.
Curated pairs are passed through a frozen base model. Only the low-rank adapters are trained, and the pass is repeated once per ensemble member — this stage, not the merge, is the cost of Track B.
The fine-tuned members’ outputs are combined into the single response that enters the comparison.
Both tracks read the same filings and answer the same question. Nothing on this page is a result — each method is set out so it can be inspected before either artifact is scored.