Upstream Zero

The public evidence layer

What we have tested, and what we have only observed.

This is Upstream Zero's public evidence layer, not a blog. Every experiment, observation, and hypothesis is preserved and dated, so the chronology of what we have learned about how AI decides who to recommend stays transparent.

We study how AI systems evaluate, compare, recommend, and eliminate vendors during buying decisions. The records below are sorted by evidence level so you can tell, at a glance, what has been tested, what has only been observed, and what remains under investigation.

How to read the evidence

Hypothesis
Proposed and still under test.
Experiment
A dated run with a recorded result and its conditions.
Observation
A pattern seen across runs, not yet established.
Finding
Accepted only after replication and scrutiny.

We separate what we observed from what we infer, hypothesize, or have replicated. Nothing here implies cause unless a controlled test supports it, and we print the zeros.

Start with your question

Which business question are you trying to answer?

Start from the decision you care about. Each question maps to the research components that study it, and from there to the evidence.

Research components

The parts of how AI recommends companies that we study.

The main way we organize the evidence. Every experiment reinforces one of these enduring components, so the work builds toward defined questions rather than a stream of one-off tests.

Hypothesis defined

Requirement interpretation

How an evaluator reads a buyer's stated need and turns it into the criteria it screens vendors against.

Core question

When a buyer states a requirement, how does the evaluator interpret it, and does that interpretation stay stable across runs?

Evidence available

Recommendation set formation

How the initial set of recommended vendors is assembled before any requirement is added.

Core question

Which vendors does an evaluator surface for an unqualified category question, and what governs who is in the opening set?

Evidence available

Vendor elimination

How and where a vendor is removed from the recommendation set as requirements are added.

Core question

At which requirement is a vendor eliminated, and what makes an evaluator drop it?

Evidence available

Frontrunner movement

How the leading vendor changes as requirements accumulate.

Core question

Does the vendor that leads the opening set keep its lead, and if not, at which requirement does it lose it?

Evidence available

Recommendation survivability

Whether a company remains in the recommendation set as a buyer adds requirements, requests validation, and moves toward selection.

Core question

As requirements are introduced in sequence, does a company remain in the recommendation set, and what removes it when it does not?

Active research

Recommendation stability

Whether the same inputs produce the same recommendation, and whether unstable outcomes trace to unstable requirement association.

Core question

Are recommendation outcomes stable across repeated runs, and when they are not, is the instability in the recommendation or upstream in how requirements are associated?

No published evidence yet

Competitor displacement

How one vendor replaces another in the recommendation set.

Core question

When a vendor is added or removed, which competitor takes its place, and what drives the substitution?

Evidence available

Validation and evidence

Which sources, proof types, and trust signals an evaluator appears to rely on when it validates a vendor against a requirement.

Core question

What evidence does an evaluator appear to use when deciding whether a vendor credibly meets a requirement?

Featured experiments

The runs that answer the biggest business questions.

E-034ClosedSupported

Clinical trial management software

When the recommendation set collapses

Question
Why does an evaluator sometimes reduce a category to one viable vendor?
Business problem
A company can disappear before a buyer ever sees a comparison.
Observed result
The observed recommendation field collapsed to a single vendor.
  • CEO
  • CFO
  • Sales
  • Marketing
E-040ClosedSupported

Project management software

When the recommendation set holds

Question
Does the recommendation field survive when a buyer profile and commercial requirements are introduced in sequence?
Business problem
Companies cannot tell whether they survive the moment a buyer stops being generic.
Observed result
The field narrowed but did not collapse. Teamwork persisted across all three draws, and the generic category leaders dropped out once buyer context was introduced.
  • Sales
  • Marketing
  • Product
E-042ClosedSupported

Project management software

The unexplained recommendation transition

Question
Does the evaluator's narrated interpretation stay stable while its recommendation changes?
Business problem
An evaluator's stated explanation may not reveal what actually drove its decision.
Observed result
The narrated interpretation stayed stable across all three draws while the recommendation moved, and the narration accounted for the opening recommendation but not the transition.
  • Marketing
  • Product
  • Procurement

Browse all experiments

The complete library, generated automatically.

Every experiment in the program, including the runs still being reviewed into the format above.

What Closed means. A completed experiment is evidence of what occurred under its recorded conditions. Closed means the run is finished, not that the result is a universal truth. Replication, causal support, cross-evaluator agreement, and real-world corroboration are what let a conclusion travel further.

Observations

Patterns seen across runs, not yet established.

Candidate observations drawn from more than one experiment. Each is held at its honest level, names the runs it derives from, and is not a finding.

Hypotheses under test

What we are actively testing.

Open questions

What remains unknown.

Findings

0
Accepted findings

Zero, and we print it. A finding is accepted only after replication and scrutiny we have not yet accumulated. 3 founding claims are published at their tier in the claims ledger.

2
Methods

How anything here earns its evidence level. How we work.