02 Populations

A population is a domain someone finally wrote down.

Pods are the unit; a population is what you get when a group of them covers a field properly. Below are the ones people have already asked us about — none of them are Trust & Safety, and that is the point.

food safety
Undeclared allergen
  • "may contain traces of tree nuts"
  • "produced on shared equipment"
  • "contains milk solids and casein"
  • "packed in a facility handling sesame"
0.78
zoology
Raptor call description
  • "a thin descending whistle at dusk"
  • "repeated kek-kek near the nest"
  • "harsh scream over open ground"
0.64
manufacturing
Weld porosity report
  • "scattered pinholes along the toe"
  • "gas entrapment near the root pass"
  • "surface pitting after grinding"
0.71
field survey
Invasive spread note
  • "dense monoculture along the bank"
  • "crowding out native seedlings"
  • "rhizomes visible under leaf litter"
0.58
compliance
Data-retention clause
  • "records held for seven years"
  • "deleted upon written request"
  • "retained only as long as necessary"
0.83
clinical intake
Medication non-adherence
  • "stopped taking it after a week"
  • "skips doses when travelling"
  • "halved the tablets to stretch them"
0.69

Illustrative pods. Scores shown are examples, not benchmarks.

03 A worked example

Aldous is one population, taken all the way

Foreshock's Aldous model is a full specification of affect — dozens of graduated dimensions, calibrated baselines, erasure behaviour, receipt rules. It exists because someone spent years writing down how a domain actually sounds.

We show it here only as a demonstration of depth. SemPods carries none of it. No affective observation, no Trust & Safety framing, no opinions about people. Your population starts empty and stays yours.

What Aldous proves is the ceiling: a population built on this engine can become a serious instrument, and it can be trained transparently on hardware you already own.

What a mature population has
  • A scope statement — what is in, what is deliberately out.
  • Phrase sets per concept — three to five, complete sentences that carry their own context.
  • Calibration notes — the thresholds, and why they sit there.
  • Receipt rules — what an affected party is shown, and what they are not.
  • A revision history — because a concept that never changed was probably never checked.
In development

Vision pods put the same geometry on photographs

Image embeddings behave the same way under a centroid classifier. That makes two populations we are especially interested in: line-side quality assurance, and field identification where the specimen is in front of you and the reference is not.

Tell us your population →