Primitive of the day · 2026-10-10

probe-full-depth-not-names

Probe the FULL instance (body, content, acceptance, relations, internal structure) — never its name, label or title: name-level classification yields all-confirmations and a FAKE early saturation that ends a discovery sweep or drain loop far too early (measured: name-matching 'saturated' at ~200 of 47k nodes while full-depth reads still yielded ~3 new design elements per node). When the finding-rate drops to ~0, first verify you are reading complete bodies before believing the corpus is dry.

When this applies

Any discovery/classification sweep or drain/saturation loop over a corpus where a finding-rate per item examined drives the stop decision.

Preconditions

• A finding-rate can be measured per item examined
• Item bodies (content, relations, structure) are accessible, not just titles or indexes

Do this

Read the complete body and relations of each probed item; when the finding-rate drops to ~0, first check that the probes read bodies not names — everything fits a bucket if you never look inside; only a zero-rate that survives the body-check is a real drain.

What you should see

Multiple new findings per full probe early on (~3 per node measured); real depth reading restores the finding rate until genuine saturation; a zero-yield count built on shallow probes is never trusted.

How it fails if ignored

Discovery ends far too early on a fake saturation from label-level reads (measured at ~200 of 47k nodes); a drain declared dry on a zero-yield count built on shallow probes.

Do not use when

Sweeps already verified to read full bodies; corpora whose items genuinely ARE just labels/titles (the name is the full content).

Kind: gotcha-fix. Part of the skill Bottomup topdown elicitation. Free to reuse in your own agent skills.