v0.26.0 - Physical output-margin sensors and acquisition headroom

v0.26.0 - Physical output-margin sensors and selective-acquisition headroom

Second of six authorized bounded research releases. A 2 minute 25 second
screen establishes useful page-level acquisition headroom and a measured
decision-sensitive observation. It remains a controlled, oracle-assisted research
result—not an online loader, speedup or larger-model capacity claim.

Result

• Final FFN: 35 physical neuron pages, deliberately two-bit resident base,
  three refinement increments per page. Other FFNs use six-bit decode.
• Of 32 diagnostic positions, base differs from the high-eight draft at nine.
  A privileged 17-page oracle repairs 9/9; fixed first-17 repairs 7/9.
  Both introduce two new draft-reference disagreements. Haction passes.
• Full diagnostic target agreement: base22/32, fixed29/32, oracle31/32,
  high-eight31/32. These are teacher-forced decisions, not accepted draft tokens.
• All 4,480 page-margin projections pass tolerance; maximum diagnostic
  absolute error 4.828e-6. All 28 FFN numerical controls and eight full-model
  readout/KV replay checks pass. Hobservation passes.
• Mean paid page observation 3.532 ms; full-model high-eight replay
  374.273 ms. Different work amounts: their ratio is not a speedup claim.
  Three diagnostic high-eight winners lie outside the base top-two; pairwise
  ordering does not certify full argmax.
• Research collection pays 65,379,336,192 H2D page bytes, plus construction
  and other charged work. It observes every page before oracle selection and
  physically reloads both partial controls. No online traffic saving is claimed.
• Peak extra CUDA 907.302 MiB under 1024 MiB. The final two-bit base saves
  10,321,920 resident bytes locally, but total sampled GPU peak is unchanged
  from v0.25. No larger-model capacity frontier was tested.

Worker 136.362 s + independent CPU analysis 8.711 s. A second model-free
audit reproduces the full summary exactly in 8.697 s.

Source/protocol freeze:
aa8f0d9071b87aac290320dd176432b81ea657e9.
Protocol,
full results and limits,
compact receipts,
inspirations.
The RMSNorm readout identity is an application of
Zhang and Sennrich's formula, not invention
of normalization or a transferred performance claim.

Validation and artifacts

Independent core/integration review added linked KV histories, exact pre-FFN
residual readouts, fixed-parent bitplane binding and separate reload timings.
All 456 CPU tests pass on the reviewed source candidate. Publication requires
the attached full clean-final-tip test XML and commit-bound validation receipt;
those do not rerun checkpoint inference.

Raw asset output-sensors-v026-raw.zip: 496,347,675 bytes, 212 members.
SHA256: 0f17b52828420b04e93145817ff381a058d22c0e000f3b5c90932cfdb711d32f.
Every member is decompressed and SHA256 checked; raw sources are rechecked.
Includes complete observations/logits, integer planes, source/protocol, physical
and state receipts, fixed inputs and licensing. Unchanged full embedded weights
remain the explicit pinned v0.23 dependency.

Next

This measured action headroom admits a separately frozen, minutes-scale spatial
page-field estimation screen: can a few paid observations identify useful
unobserved pages? Compare independent and shuffled-geometry controls; subsequent
recursive and value-of-information policies still need their own tests.

Complete bounded #41.
Milestone10
stays open. Stock b10919 is unchanged. No long matrix or native worktree follows.