SEED · Judgment Lab
Judgment Lab
A working archive of AI governance case analyses: a system, the failure observed in intensive use, the judgment behind it, and what it taught the framework.
What the Lab Does
Judgment Lab is SEED’s research arm for the judgment layer of AI systems. It studies moments where an output is technically correct yet culturally, emotionally, socially, representationally, or institutionally wrong — and traces each failure back to the human judgment it depended on.
Cases are published only when they can be documented.
How a Case Is Written
- 01Each case names the system and the setting, at whatever level of detail is cleared for publication.
- 02Each case records the failure as observed in intensive use, not as hypothesized.
- 03Each case closes with what it changed in the JUDGE framework or in the Human Plausibility / Cultural Legitimacy Test.
The JUDGE Framework
J — Judgment
Which human judgments has the system absorbed, and who made them? Defaults, thresholds, and taxonomies are decisions, not settings.
U — Use
How is it actually used, as opposed to how it was specified? Intensive, situated use is treated as primary evidence.
D — Difference
Where does the system misread culture, language, register, or context — and whose difference is flattened when it does?
G — Governance
Who is accountable when an output is correct but wrong? Escalation, override, documentation, and review are located or named as missing.
E — Explanation
Can the organization explain a decision to the person it affects, in terms they would accept as reasons rather than as process?
Human Plausibility Test
Would a reasonable human community recognize this output or decision as socially intelligible, appropriate, and defensible? The test is being developed and calibrated alongside JUDGE; it is used as a structured judgment instrument, not as a certification.
Propose a Case
Organizations willing to test the diagnostic on a live system are the fastest route to the next case.