Date: 2026-09-27. Status: bounded preflight passed; primary qualification open.

Research question and outcome

Can a frozen Foundation receive five independently encoded Context sources in one pass, learn independent contribution gates from G, and produce a derived class that is absent from every source atom? Yes, for the locked binary-count preflight. This does not establish general multi-Drone semantic reasoning or a computational advantage over a small typed model.

The candidate kept Foundation 001.x frozen and attached a new evidence-access module at blocks 4, 6 and 8. C1–C4 each carried one independently read observation. G carried a policy naming two local sources. A learned gate predicted which sources should contribute, and each Foundation block queried all five source representations. The classifier used the protected Foundation's existing byte output head for the first character of GREEN, AMBER or RED. No source text or input packet contained the result label. This was a class-logit task, not free-form answer generation.

Protocol and measured results

The task used the V6 deterministic evidence reader and explicit packet provenance. Each counterfactual group kept identical C1–C4 observations and changed only G's selected pair, obtaining all three answer classes. Data were locked before training: 1,152 training, 288 development, and 576 sealed examples, with disjoint entity IDs and balanced labels. The sealed split was scored once for the selected learned-gate candidate.

The learned access module had 3,603,690 trainable parameters; the Foundation parent was frozen and retained its hash. The gate is trained with a direct source-selection auxiliary target derived from G's typed policy, while the task class is trained from verified labels. Gate scores are independent sigmoids, not a one-winner softmax. The module's source-local attention, gate and mixer participate at three frozen Foundation blocks. The result shows causal use of two required local facts and G on this preflight; it does not isolate how much of the class computation occurs in the frozen Foundation versus the new access/adaptation layers.

Diagnostic sequence

Three unmasked bridge variants stopped at their development micro-gates instead of receiving repeated unchanged training: source-local summation reached 33/96, a bounded dense variant 39/96, and a source-slot projection 32/96. A tiny typed MLP achieved 100%, demonstrating that the task and atomized inputs were learnable. A deliberately answer-leaking positive control reached 96/96 and showed that the attachment could steer the Foundation head, but was never eligible for qualification. A parsed-G selection diagnostic reached 96/96 without an answer label. This localized the weak boundary to aligning G's selected source IDs with source-local states.

We then trained a G-parser-directed two-fact path on the full development set; it reached 288/288 and passed its interventions. Finally, the learned-policy variant replaced that deterministic masking with independent neural source gates. It reached 288/288 development by step 100 and passed the one-time sealed evaluation. The production input construction was audited after a harness defect was found: source values and G selection now come from reader packets and G text, not the dataset's audit-only bits and pair fields. A test mutates those gold fields and confirms model inputs are unchanged.

Qualification limits

This is useful evidence that the runtime can keep C1–C4/G source states separate, expose all five in the same frozen-Foundation pass, train multiple contribution gates, and make the final class respond to required facts and G while ignoring distractors on a held-out task. The model survived serialization and a new-process load.

It is not a full F08/C08 qualification. The preflight has one binary fact per local source, exactly two active facts among four, a typed G pair-selection policy, and only a three-class count operation. It has no natural-language ambiguity, missing selected evidence, time/authority conflicts, multi-step proof, novel operators, source updates/cache invalidation, or general free-form Foundation reasoning. Source versions are verified in packets, but are not encoded as neural features. The 2,931-parameter typed MLP also solved development perfectly, so this benchmark cannot demonstrate a benefit from using Foundation or the 3.60M-parameter bridge. The learned gate's auxiliary supervision makes the source-selection problem much easier than purely outcome-learned contribution. No new app route was admitted; the five existing V4–V7 routes remain the application baseline.

This V8 run stops here because the compact typed control already saturates the preflight: additional training or another sealed score on the same task would not resolve the central Foundation-side claim. The next experiment should change the task and control, not repeat this count benchmark. Use independently authored, multi-atom source worlds with variable selected-set size, contradictory/time-stamped evidence, missing required claims, and two-step derived conclusions. Compare the learned Foundation path against a matched bridge-only readout and a compact typed model on identical locked splits. Test version changes and cache invalidation causally. Retain class-logit evaluation separately from language rendering. Qualify F08/C08 only if required-source and G interventions, generalization, parent-task retention and restore remain sound on that harder family.

Artifacts and implementation

• Experiment and locked split generator
• Source-preserving access module
• Contract tests
• Local ignored run directory: [retained internal evidence]; candidate tensor SHA-256 [checksum retained in the private evidence record].
• Dataset hashes: train [checksum retained in the private evidence record], development [checksum retained in the private evidence record], sealed [checksum retained in the private evidence record].
• Targeted tests: 13 passed across the new V8 tests, multi-fusion contract and integrated app regression.

Mechanism references remain CEPE, FiD, and OpenFlamingo. They motivate separate external representations and frozen-backbone access; none validates this EMMA result directly.