each agent answers exactly one question and holds only the tools that question requires. the badge says how the job is done today, not how important it is.
- model
- a model call runs, and its answer is checked before anything is stored.
- beta
- the same, recently added and still being watched.
- deterministic
- no model at all — an engine under app/services does it.
the last two are deliberate, not unfinished. a hypothesis is written from a template so that nothing reads the data before deciding what to claim about it, and the statistics have one right answer. a model in either place would trade a guarantee for a sentence.
the two that do run are constrained rather than trusted: the reviewer and the critic can only make a verdict harsher, never lighter, and the observer describes an anomaly a detector already found rather than finding one. every model call is recorded with its tokens and its cost, and refused before it happens if the day’s budget cannot account for it.
critic
beta testingWhy might this be wrong?
role
Attacks the result before anyone believes it
inputs
Experiment, ExperimentResult, dataset metadata
outputs
critic verdict PASS | NEEDS_MORE_DATA | FAIL (stricter only)
tools
db.read, db.write:experiment_results.critic_*
data_scientist
deterministicDoes the data support it?
role
Builds the dataset and runs the test
inputs
Hypothesis, token snapshots, social posts
outputs
Experiment, ExperimentResult
tools
db.read, db.write:experiments, compute
model role
MODEL_REASONING
model behind it
no — a deterministic engine does this job
observer
beta testingWhat does this measurement look like?
role
Reads an anomaly the detectors already found
inputs
Anomaly, Observation payload, detector thresholds
outputs
observation.payload.observer_reading
tools
db.read, db.write:observations.payload
researcher
deterministicWhat could explain this?
role
Turns anomalies plus memory into falsifiable hypotheses
inputs
Observation, Anomaly, memory search results, past experiments
tools
memory.search, db.write:hypotheses
model role
MODEL_REASONING
model behind it
no — a deterministic engine does this job
reviewer
modelIs every claim in this supported by a stored result?
role
Final gate before anything becomes public
inputs
ContentDraft, referenced source row
outputs
reviewer verdict PASS | FAIL
tools
db.read, db.write:content_drafts.reviewer_*
writer
modelIs there something worth communicating?
role
Drafts public messages from real results only
inputs
Observation, Hypothesis, ExperimentResult, Memory
tools
db.read, db.write:content_drafts