Research question
Can the implicit requirements an AI evaluator applies be recovered by asking it directly?
Why it mattered
First contact. It established whether a recommendation set can be decomposed into the unstated requirements that produced it, and whether those requirements map onto identifiable buying-committee owners. This run produced the four-turn extraction protocol the rest of the program depends on.
Method
Evaluator: Google AI Mode, signed in, retrieval on. Two categories: ERP for law firms and corporate travel for law firms. Manual and exploratory. The extraction question was invented during this run and frozen only afterward. Sample: one draw per category. No controls. Ads and location were not logged; those columns did not exist yet.
Direct observations
- The evaluator produced a structured account of unstated requirements when asked directly.
- Extracted gates mapped to identifiable commercial uncertainties without forcing.
- VIP and partner servicing did not appear in the turn-one landscape.
Interpretation (not established)
The account may reflect real evaluation structure, or a culturally plausible reconstruction assembled from buying-guide and RFP text. Both produce identical-looking output. This ambiguity was named at the outset and remained unresolved through E-019. The VIP absence is consistent with an evidence-accumulation boundary, where relational requirements do not fossilize in public text.
Outcome and remaining uncertainty
Inconclusive. This was an exploratory, pre-protocol run that scored roughly two of four on a pre-registry card. Its lasting value is the method it created rather than a settled answer. Whether the extraction reads real structure or reconstructs a plausible one was left open.
Evidence and methods limitations
Self-narrated extraction, single evaluator, one draw per category, a single coder with a strong prior toward the frameworks under test, and no codebook. The protocol did not exist at the time; this run produced it.
Contribution to the research program
Established extraction as a method and named ghost requirements as a class. Both were later refined.
Follow-on
E-006 decomposed the same category thirteen questions deep.
Created from the canonical Upstream Zero research archive (July 2026). Raw outputs are retained privately and are not published.
Predictions, registered before results
- Matter-code integration will appear as a gate (HIT)
- Aderant and Elite-class incumbents will anchor the ERP field (HIT)
- Expense-first elimination will drive the kill order (MISS)
- VIP and partner servicing will appear as an explicit requirement (MISS; the ghost held)