What can a model reliably read?
Structure. A contents page. A close date printed as a date. A list of files in the pack. A table that is actually a table, with headers that mean what they say. Used as a junior reader, that saves the hour you would have spent making a map. It does not save the gate.
Clearly labelled mandatories in an evaluation table are the best case. Even then, tick them against the instrument, not against the chat. Models sound sure. Evaluators void you anyway.
What can they not reliably read?
Mandatories that hide. A “must” in schedule 4. A prequalification on another portal. A filename rule in the lodgement guidance. The hunt is still a hunt. A confident summary that missed one is worse than an empty page, because the room will stop looking.
Indemnity, liability caps, consequential loss in definitions, tables that are images, flow-down in a head agreement the model was not given. Cross-references are a system. Nearby sentences are not. The fifteen flags exist so a human starts that hunt.
Why does consistency matter?
A gate has to be reconstructable. The same pack, the same firm, the same clauses named. If a refresh produces a different list, the working is not working. It is a draw from a distribution. Partners cannot sign a distribution.
Citations are the test. “Somewhere in the conditions” is not a clause. A page number, a clause number, a quote. If the model cannot point, the call cannot travel to a person who was not in the room.
Where does this sit in the landscape?
Response tools that draft from a library are a different job. Discovery tools that find listings are a different job. Qualification is citing and repeating against your firm. A general LLM is a junior reader at best. selrob is built for the gate. Do not buy a chat window to do it, and do not pretend a fluent extract is a cited one.
Confidentiality is the next question: where the pack lives, who trains, whether a sensitive ATM should ever leave the building. That is security when putting tenders through AI.