ahepi/deepreason/expressibility-probe

Expressibility Probe (new in 0.5, from FR-23)

<!-- PROMPT-CORE-BEGIN --> "X is not expressible" is two claims and a conclusion, and each part is checked separately - by construction and execution, never by reading the model (FR-25).

1. THE DATA QUESTION: does the referent exist in ANY record reachable

from the anchor - including via joins the vocabulary cannot follow? Walk the record graph, not one record's fields. (The source cycle declared "no atom referent exists" while a two-hop path discharge -> challenge -> atom sat in the model.)

1. THE CAPABILITY QUESTION: can the vocabulary follow the path?

Construct the actual attempt and run it through the real validator; record the refusal code. A dataclass accepting a shape that validation refuses is exactly why reading is not enough.

1. THE MANDATED FRAMING: the two answers force the conclusion -

• referent absent everywhere: the choice is genuinely narrowed; say

to what;

• referent present, capability absent: this is evidence of a

MISSING CAPABILITY, an argument for amending the model, NOT for accepting the approximation. The decider chooses between an approximation and an amendment, and must be told that is the choice. Writing "narrowed" where the second case holds is the FR-23 defect.

1. Distinct inexpressibilities are distinct: name each one's fact and

the test that measures it (see decision-mapping rule 4).

1. Each probe outcome enters the record with an executed route

(discharge-typing rule 3): the construction script and its actual refusal or acceptance output. <!-- PROMPT-CORE-END -->