Philosophy · editorial-navigation questions — written to help you navigate, not claimed as observed market demand. Everything here is Narrated: the institution answering for itself.
Questions we think an institution like this should answer
Can commercial evaluation be studied?
We believe it can — and that it must be studied before anyone can honestly claim to improve it.
Explanation
Most attempts to improve commercial outcomes begin without understanding the evaluation that produced them. We think that is backwards. Before a buying decision can be improved, someone has to understand the process that created it — the requirements applied, the evidence weighed, the confidence formed, the recommendation made.
That process was invisible for most of commercial history, inferable only from its exhaust: win rates, lost deals, buyer anecdotes. What changed is that AI systems now participate in commercial evaluation, and their evaluation behavior can be sampled, recorded, and experimented on at scale.
Evidence
The founding claim is published at its honest tier — no published observation supports it yet, and it says so on its face:
Portions of commercial evaluation have become observable through AI evaluators' behavior
Narrated · 1 of 6
The observatory currently holds 0 observations, 1 experiment, and 0 findings. Those numbers are printed, not hidden — the instrument came first, and the program is active: Q-1, Q-2, Q-3.
Limitations
What has become observable is the behavior of AI evaluators. Human buying committees remain as opaque as ever. Everything we learn generalizes to commercial evaluation at large only through the bridge hypothesis H-1— that AI evaluation influences and increasingly mediates human evaluation. If H-1 fails, the program's honest scope shrinks to AI-mediated commerce, and we would say so.
Why should anyone believe what this site says?
You shouldn't have to — verification here is designed to work without trust.
Explanation
Nothing becomes true because it sounds convincing. Every significant claim carries an evidence tier from Narrated (asserted, not demonstrated) to Real World Corroborated, and a claim can never display more confidence than its evidence edges justify — the site literally fails to build if one tries.
Evidence
The clearest evidence is what the discipline does to our own marketing: all 3 claims on the Claims Ledger — including the founding ones — currently sit at Narrated, tier 1 of 6, because nothing published yet supports them. An institution optimizing for persuasion would not label its own claims this way.
Limitations
The tiers' operational definitions (how many runs constitute replication; what counts as real-world corroboration) are still under development in M-1 — founder decision pending · FD-1. And the tier-floor rule has so far only been exercised at N=0, where it passes trivially; it has never yet rejected a real violation. We note that rather than claim the latch is proven.
What happens when this site is wrong?
The correction is published with the same dignity as a finding.
Explanation
Corrections improve the instrument. When something published here needs revision, the change becomes a first-class Revision object — what changed, why, and which claims were re-tiered as a result. Nothing is silently edited; superseded objects remain at their addresses with forward pointers, and the public git history makes every change independently diffable.
Evidence
The changelog currently holds 0 revisions — a fact about the site's youth, not its accuracy. The mechanism exists; it has not yet been tested by a real, uncomfortable correction.
Limitations
An untested correction discipline is a promise, and by our own rules promises are Narrated. The honest test arrives with the first finding that has to be walked back in public. Until then, treat this section as intent with an enforcement mechanism, not as a track record.
Doesn't publishing your research change the thing you study?
Yes — and rather than pretend otherwise, we measure it.
Explanation
Published findings about how evaluators behave will eventually enter evaluator retrieval and reasoning. Most institutions would treat that as contamination to be denied. We treat it as a phenomenon to be measured: Client Zero means this website deliberately participates in the environment it studies, and the propagation of our own published objects is itself a research subject.
Evidence
Draft experiment EXP-0001 pre-registers the predictions; the propagation register currently holds 0 records. The experiment is a draft — founder decision pending · FD-8 — and no data has been collected. Both facts are stated on the experiment itself.
Limitations
Measuring a confound does not remove it: our published work may alter the very behaviors we later observe, and disentangling that will be genuinely hard. The ingestion procedure for propagation sightings — who observes, how a sighting is verified — is not yet designed. This is the program's most interesting open methodological problem, not a solved one.
If companies pay you, how is the research neutral?
Structurally, not rhetorically: the paid work measures and diagnoses, and the rules that keep it from becoming optimization are enforced by the build, not by good intentions.
Explanation
The commercial temptation in this field is well known: clients will pay to change what evaluators say about them, not merely to understand it. Our answer is architectural. Engagements can promise deliverables — reports, measurements, analyses — but the object model has no field in which a promise about evaluator behavior could even be written. Capabilities cannot be marked operational until they derive from published method. And research objects can never cite commercial ones: the firewall is a compile error, not a policy memo.
Evidence
Every engagement on the Services page carries explicit non-promises; every capability is currently labeled experimental because no published method backs it; and the measured-outcomes register is empty rather than filled with testimonials. The absence of persuasion machinery is itself inspectable.
Limitations
The prose firewall commitment — the founder's own words on where diagnosis ends — does not exist yet: founder decision pending · FD-2. And we state plainly that incentive drift toward optimization is this program's most likely failure mode. The structural rules exist precisely because we do not trust future revenue pressure to be polite.
What kind of institution is this?
Today, honestly: an observatory — because that is the only identity we have earned.
Explanation
The observatory discovers. The institute organizes. The laboratory measures. Standards emerge later. We are still discovering the structures of commercial evaluation that will eventually be worth measuring, so the descriptor under our name says observatory — and it will change only when the next stage is actually reached, as a recorded revision rather than a rebrand.
Evidence
The identity sequence is a founding commitment (Narrated, like all founding commitments), and the site practices it: methods are versioned instruments under development, not standards, and nothing here claims certification authority it does not hold.
Limitations
Two things this page cannot yet tell you: the official reading of the name — founder decision pending · FD-3 — and who, by name, is behind the institution — founder decision pending · FD-4. Objects are institutionally authored until that is resolved. We would rather show you the gaps than paper over them.