Stage 1 preregistration — signposting at scale, plus a payload arm

Filed: 2026-09-05, before the first assignment ran. Envelope: 200 assignments (main matrix) + 40 (payload arm), authorised on aegis-vkirh and scoped on aegis-3npm7w. Gate cleared: Stage 0 certified delivery coverage at 100% against the 25% gate (2026-08-29-stage0-delivery-coverage.md).

This document is committed before the campaign so the analysis cannot be chosen after seeing the numbers. Everything below is fixed at filing time.

Why there are two matrices and not one

The two pilots agree on one thing: agents do not take pointers. Across both, zero signposts were adopted, and the second one established that the instrument — not the intervention — had been the problem.

So this campaign asks two different questions, and keeps them apart:

They are planned separately and scored separately, and their adoption rates must never be pooled, because adoption is not the same act in the two:

• a pointer arm’s signpost names a command, so adoption = the agent ran it;
• the payload arm has already run it, so adoption = the agent used a location the payload handed over.

Both facts are recorded on every row (ran_stack_command, used_injected_location) and adoption_kind says which one adopted holds. dp eval score refuses to aggregate a cell whose rows disagree about the kind. The metric that IS directly comparable across every arm is correctness, and it is the one to read first when the two adoption numbers differ.

Fixed design

• Tasks: 20, each with a blinded acceptance set adjudicated and committed before the campaign (eval/tasks.jsonl, eval/acceptance-sets.jsonl). eval/check-acceptance.sh asserts, at the pinned revisions, that every ground truth exists and falls inside its own accepted range — the defect that made the 2026-08-27 pilot’s correctness numbers unpublishable was an acceptance set that excluded valid nearby locations, so this is now a gate and not a habit.
• Families: codex and claude. Two families is a stated generality limit, not a claim about models in general.
• Conditions: the five above; 20 x 2 x 5 = 200. Payload arm: 20 x 2 x 1 = 40.
• Replicates: 1. The blocking is over tasks, not repeats.
• Assignment: deterministic blocked matrix, seed stage1, shuffled by hash so no arm runs in a favourable block.
• Budget: the 150 ms PostToolUse contract, unchanged. The PreToolUse prefetch runs, as it does in production.
• Index: the local resident index over the five pinned corpora (eval/build-local-index.sh). Certification runs against this, not against a shared backend that does not hold every corpus.

Endpoints, fixed now

• Primary: adoption rate among model-visible signposts, per family x condition, with adoption_kind attached.
• Secondary: correctness rate, turns-to-locate, tokens-to-resolution, all per family x condition.
• Payload-arm comparison: payload-signpost against always-signpost — they apply the same gate and differ only in what is injected. It is not compared against gated-signpost.
• Intent families are recorded per row (intent_families) and reported, but the campaign is not powered to compare families against each other; that is delivery-harness work, below.

The failure reading, stated in advance

An adoption rate indistinguishable from zero at n >= 60 model-visible signposts is a publishable negative result about signposting as an intervention. It closes the question; it does not motivate a third campaign.

If the payload arm is adopted where the pointer arm is not, the finding is about pointer vs payload, not about signposting being fixed — the payload arm spends stack work on every eligible search whether or not it is wanted, and that cost is the next thing to measure, not this campaign’s result.

No post-hoc arm selection

The arms, the endpoints, the comparator, and the failure reading are the ones in this file. Anything discovered afterwards is reported as exploratory and labelled as such.

What is certified on the delivery harness instead of here

Intent widening (find -name, git log -S|--grep, and symbol promotion) is a delivery property and is measured with no model in the loop, on the replay harness, paired and interleaved against the pre-change build. Model quota is not spent on it. See the Stage 1 delivery note.