issue #76 (inferswarm): qualification harness + Phase 0 stop — PHASE0_REFERENCE_COMPLETE / STRESS_SELECTION_BLOCKED (v2 methodology: inferswarm#77/#78) - #28
Open
ezutfen wants to merge 7 commits into
Conversation
…terogeneous numerical qualification Implements the minimum harness required to execute the frozen issue FlashML-org#74 methodology (inferswarm@f394dc9): case-driven single-arm reference runner (RTX 3090), three-stage chain runner (2x RTX 3060 node-01 + remote last stage on node-03 via a per-case R4 wire service), RowPruningSink capture (FlashML-org#71-compatible, host-side final-row pruning of the full BF16 logits matrix), o_proj checkpoint wrappers (layer-0 in/out, global-layer-15 attention projection), pure host-float64 reducer matching the frozen REDUCER.md identity, and torch-free unit/source-contract tests. No execution/model math is added or changed: all model execution flows through the accepted R6 GemmaDenseStage replay-prefill semantics.
…the full corpus sentinel-subset.json carries identity hashes without prompt text; runners now cross-check subset rows against the frozen calibration corpus before execution.
…stage 1's bundle)
…ock frozen stress selection The pre-registered margin definition (min over 8 greedy steps, producer 29e04d0, committed before the reference run) yields exactly-zero margins in 5 of the 48 frozen pool cases; the frozen selector rejects nonpositive margins. Stopped before candidate execution per the issue's stop condition. Full per-case margin table and run index retained; evidence bundle (3.6 GiB, tar sha 3b2eb3e1...) retained on inferswarm04 and the orchestrator.
ezutfen
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Execution campaign for Zutfen-LLC/inferswarm#76 (Gemma heterogeneous numerical qualification, frozen methodology inferswarm@f394dc9).
Status:
No candidate (3060) execution occurred. No threshold derivation. The holdout remains sealed and verified (custody FOUND_VERIFIED — see below). No R6 PASS or qualification claim. This PR is preserved as the physical harness + stopped-v1 evidence line; it must not be merged in a way that erases the stopped state.
Maintainer adjudication (2026-09-03) has been recorded on FlashML-org#76; the successor methodology v2 lives in Zutfen-LLC/inferswarm#77 / PR FlashML-org#78. The harness code in this PR may later be reused for v2 if code-identical and revalidated, but v2 requires its own prospective producer chronology.
Completed before the stop
29e04d0(this branch, off the required based4d1608), clean trees on inferswarm01/03/04, native extensions rebuilt, torch-free harness tests (17) green.3b2eb3e1d3de4bcef67e12b202be4b9da759baec786b0e5b5a3e25887b79a869; margin-table sha2567016136b042af4e420cdb8a8b6483f2d331d8a260880a70091c9317ca66f7bb0).Stop: zero top-1 margin ties
The producer pre-registered (before the reference run): margin = min over the 8 greedy steps of fp32(top1−top2). Under it, 5 of 48 pool cases have EXACTLY zero margins (genuine bit-identical fp32 logit ties for distinct tokens; steps 2/3/5). The frozen v1 selector rejects nonpositive margins pool-wide → selection cannot complete validly → stop per FlashML-org#76's own barrier.
Maintainer decisions: the margin definition is RETAINED verbatim (the five zero cases are legitimate evidence; alternative definitions were rejected as results-informed selection). The defect is the v1 selector's fatality semantics, fixed prospectively in methodology v2 (FlashML-org#77/FlashML-org#78): fresh pool, positive eligibility, ≥8 eligible, four smallest + four largest positive margins, same min-over-8 margin.
Full analysis, per-case table, positivity table:
docs/inferswarm_76/PHASE0-STOP.md.Holdout custody: FOUND_VERIFIED
The FlashML-org#74 recipient private key + secret seed were recovered by the maintainer from the original sealing laptop and verified against the committed certificate WITHOUT unsealing (public-key DER sha256 match
f0a89fea…f8693; secret-seed sha256 matches the commitment). The v1 holdout is preserved exactly — ciphertext sha25623311c5514b2561c66a2ecd0c9cfa25c3f4f91b83b67353aada8355f48e25c59re-verified unchanged, stillSEALED_NOT_CONSUMED— and remains applicable to v2 (only stress-case selection changes). Redundant custodian copies recorded; no key regeneration; no resealing; no private material committed.Non-decisions
Exact top-two ties are not a FreeToken bug, not an InferSwarm bug, not a qualification failure, not GPU corruption. No epsilon, no token-id tie-breaking, no execution-math changes. A tie is a zero margin.
No self-merge; awaiting maintainer review of FlashML-org#77/FlashML-org#78 before any v2 execution.