Under the hood · Reports
How a report is made
A Stynch report is produced in two strictly separate stages. The analysis engine computes. A language model writes, from these computations and nothing else.
The language model has no access to LinkedIn, and its general knowledge is not an authorized source: it only works on the data Stynch exposes to it.
Corpus
A bounded scope
A report covers one author and a precise window: a calendar year, or the last 100 posts. The window is delimited by an identified first and last post. Two reports on the same window cover exactly the same corpus.
Computation
Measures come from the engine
Topics, syntactic vector, rhetorical structure, archetypes, intents, funnel, mindset, citations, performance: all measures come from the Stynch analysis chain, the same ones as in the extension.
Each position is a percentile in a real population: profiles of the same topic by default, follower bracket when the topic does not match the author's field.
Data access
Tools, not memory
The model queries Stynch through the Model Context Protocol, using typed tools: profile analysis, metrics, positions, populations, comparisons, most representative posts, evidence. It also reads graph data, extracting only the fields needed for the current section.
No figure can appear in the report unless it appears in these sources.
Writing
A generation contract
The report is written section by section, under an explicit contract:
- what is not in the sources does not exist: missing data removes the related angle, it is never assumed;
- no outside reference: no "industry average", no "LinkedIn standard", no invented threshold or numeric target;
- same conventions across all sections (default population, rounding), so that figures match from one section to the next;
- defined vocabulary: Stynch labels only carry the definitions of the supplied glossary;
- each lever combines the numeric signal that justifies it, an observable action on upcoming posts, and the Stynch indicator that will measure its effect.
Intermediate text is kept: an interrupted rendering resumes where it stopped, and the layout can be rebuilt without calling the model again.
Control and delivery
Reviewed, encrypted, signed
- A new report is reviewed and approved in a review console before being made available.
- The PDF is password-encrypted and issued in the name of its licensee.
- It carries an Ed25519 signature, verifiable with a public key. The signed payload (licensee, licence, SHA-256 fingerprints of the source, visuals and PDF) is replicated in several places in the document, and visuals carry a watermark.
As in the extension, every observation can be checked in the posts: performance, mindset, copywriting.