The measurement
- Reproduces published answers
- yes
- Schema
- canli.alphac-reachability-harness.v1
Families
Nested beyond what this page renders. Read it in the artifact.
How to add a family. Write a probe returning a Reachability with its own ceiling_rate and ceiling_basis, register it, and add its expected verdict to PUBLISHED_VERDICTS if it has already been worked by hand. The harness refuses to run if any registered family stops reproducing its published answer.
Purpose. Run BEFORE a protocol is written for any family. A near-miss gate invites exactly one move (widen the detector until the number clears) which is tuning a measurement to agree with a target. This asks the opposite question first.
Verdict meanings
- GATE ALREADY CLEARS
- The gate already passes as measured; nothing to reach.
GATE BLENDS TWO POPULATIONS. One threshold is applied across populations that do not share the obligation it tests. The blended rate describes none of them.
GATE UNREACHABLE BY DETECTOR REPAIR. A perfect detector reaches less than the gate requires. The shortfall is what the documents contain, not what the parser reads, and no extraction work will close it. The identity needs a redesign that names the document carrying the evidence before it names a threshold.
REACHABLE BY DETECTOR REPAIR. A perfect detector would clear the gate, so the shortfall IS extraction and the work is worth doing. This is the only verdict that justifies parser effort.
What does not generalise. The CEILING PROBE. What a perfect detector would find is a property of the mechanism, so every family must supply its own; the harness makes that a required argument rather than inventing one. A harness that guessed the ceiling would be a rubber stamp.
What generalises
- the decision logic: measured rate vs gate vs ceiling, and the verdict that follows
- the heterogeneity test: whether a blended rate hides a subgroup that clears alone
Check it yourself
Every figure above is read from /glassbox/reachability_harness.json, the
artifact the engine wrote.