Date: 2026-09-26 Result: QUALIFIED READER COMPONENT FOR THE TESTED A → B → CLASS TASK FAMILY Foundation direct-fusion status: NOT QUALIFIED

Method revision

The first all-at-once raw source bank reached 1,063/1,536 (69.21%) exact answers and 120/512 (23.44%) complete counterfactual groups. Local lane permutation was perfect, but substituting a wrong G.START did not change outputs. Directly injecting the answer into G changed 0/384 outputs; putting it in C1 changed only 33/384. This isolated two failures in that particular candidate: query-conditioned path composition and a weak frozen-Foundation evidence/output bridge.

The generic edge-token relation model was stopped at step 600 because exact development accuracy stayed at 0% while training loss fell from 7.14 to 2.22. Its remaining 1,448 steps were cancelled. The third fusion arm was not run. See the diagnostic stop report.

The replacement extended EMMA’s already successful two-edge learned ContextPathReader into a three-stage learned pointer chain. It reused the independently hashed path-reader entity encoder and pair scorer as initialization, then trained one candidate on 32 groups. Training stopped at the planned 200 updates; no full-data fine-tuning was needed because the small candidate already solved all 64 unseen development groups.

Results

The exact artifact is [retained internal evidence], SHA-256 [checksum retained in the private evidence record].

• Micro-overfit: 96/96 exact on 32 training groups.
• Unseen development check: 192/192 exact and 64/64 complete groups.
• Full development: 1,536/1,536 exact; 512/512 complete groups.
• Learned path-edge selection, structural validity, and three-local-source provenance: 100% on full development.
• Query-swapped G.START: 384/384 exact; all 128 groups complete.
• Local-lane permutation: 384/384 identical payloads.
• Removing G.START: the reader abstained on all 384 diagnostic rows.
• One local lane plus G: structural path validity was 0%; exact output was 15–20%, consistent with the path being unavailable.
• Removing a required local lane: structural path validity was 0%. The model still emitted some labels from remaining distractors, so missing-evidence abstention is an outstanding limitation.

Before opening sealed data, a fresh-process restore reproduced the complete development prediction hash exactly: [checksum retained in the private evidence record].

The one-shot sealed result was 3,072/3,072 exact across 1,024/1,024 complete groups. Learned path selection, selected-path structural validity, and three-local-source provenance were each 100%. Sealed input SHA-256: [checksum retained in the private evidence record]. Sealed label SHA-256: [checksum retained in the private evidence record].

The candidate then passed the V3 runtime/ContextPacket integration on 96 development reads: exact packet and RETURN_EXACT output were 96/96, selected evidence sources were correct on every row, all five sources were encoded simultaneously, and Foundation weights remained unchanged. This qualifies the reader-to-runtime boundary, not Foundation’s ability to perform the path reasoning by itself.

Detailed machine-readable records are in [retained internal evidence]: [private artifact], [private artifact], [private artifact], [private artifact], and [private artifact].

Integrity and interpretation

The active Context-native checkpoint remained at SHA-256 [checksum retained in the private evidence record]. The qualified Foundation parent remained at [checksum retained in the private evidence record]. Neither was modified or promoted.

This is evidence that a small, independently trained, all-lane structured reader can generalize a learned entity-matching path across unseen identifiers, source assignments, and G query swaps. It is not evidence that the direct Foundation cross-attention bank solved general multi-Drone reasoning. The sealed set was opened once only after full development gates and fresh-process restore passed.