E0 — Predicate ratio

Gates: BC1. Run before writing implementation code.

Question

Of the claims an agent actually generates during real work, what fraction can be expressed as machine-checkable predicates rather than prose annotations?

Annotations cannot gate. If the ratio is low, elench is a search index over agent regret with a signing layer attached — which may still have some value, but it is a different and much smaller project, and it should be built differently.

Method

Take 20–30 completed agent sessions on a real repository. Brownfield, not greenfield: the unevaluated fraction is the point.
For each session, extract candidate claims by hand or with a second model. Extraction method must be recorded; if a model does it, spot-check 20% by hand and report disagreement rate. A model extracting claims will invent structure that was never in the session, and that inflation is the main threat to this measurement.
Classify each: predicate (an executable expression can be written now, by the person doing the classification, in under five minutes) or annotation.
Report: ratio, distribution by session, and the predicate expressions themselves — the expressions are as informative as the ratio.

Pre-registered thresholds

State these before running. Do not adjust after seeing results.

Ratio	Reading
≥ 0.30	Proceed as designed.
0.15 – 0.30	Proceed with reduced scope: gating over a small predicate core, annotations demoted to a search index only. Re-derive requirements before continuing.
< 0.15	Stop. Build the search index if it is worth building; do not build the gate layer.

These numbers are a judgement call, not a derivation. The reasoning: below roughly one in six, the predicate set is too sparse to cover a change’s surface, so a passing gate carries almost no information and the human is back to reading everything. If someone has a better-grounded threshold, replace these — but replace them before running.

Secondary measurements

Cheap to collect, and each answers something the design currently guesses at.

dependsOn density. Mean premises per claim. If near zero, revocation cannot propagate and the central capability is unreachable regardless of the predicate ratio. Arguably a second stop condition.
Falsification rate. Fraction of claims falsified later in the same session. If near zero, revocability is solving a problem that does not occur at session scale, and the case rests entirely on cross-session and post-release revocation — a weaker and slower-to-validate claim.
Harness-derivable fraction. Of the predicate claims, how many could have been emitted by the harness without asking the agent? High is good: it means most of the load-bearing record does not depend on the audited party.

Threat to validity

The sessions were produced by a harness that was not asked to emit claims. Claims extracted retrospectively may not resemble claims emitted natively — plausibly they are cleaner, since hindsight removes the dead ends. This biases the ratio upward. Treat E0’s result as an optimistic bound and say so in the writeup.