Skip to content

Give ring training 6,400 updates with bounded runtime limits - #440

Open
mikkel wants to merge 11 commits into
developfrom
task/adopt-ring6400-no-lrcap-20261016
Open

mikkel wants to merge 11 commits into
developfrom
task/adopt-ring6400-no-lrcap-20261016

Conversation

@mikkel

@mikkel mikkel commented Oct 16, 2026 •

Copy link
Copy Markdown
Contributor

Registers a runnable Tier 2 stretch task in the discriminator stability view. It scores raw draws from a conditional GAN on 65 continuous features. Both groups have mean zero and exchange their narrow and wide axes. The scorer checks within-group mean and SD, coverage of 80% bag-summary intervals, interval score and energy outside the two signal coordinates. The trained-task controls include a correct sampler, a pure group swap, contracted variance, padding damage and checkpoint continuation.

The included full-budget CPU baseline uses the released EMPF configuration, seed 0 and the public deterministic initializer. G, D and the prior each received 1,200 updates, with 1,200 observed regularizer applications. Across the final generated samples, variability coverage was 65.7% and combined coverage was 71.4%. The worst group/coordinate coverage error was 23.4 percentage points against a 12-point tolerance. Forge graded the sustained result FAIL. The compact receipt and 24-frame training GIF are in reports/forge/wide-conditional-spread/.

The adapter currently supports the clean EMPF setup with a learned MoG prior and continuous penalty on each step. The check asks whether trained samples recover the right distribution. Companion PR #556 checks the observer directly.

Closes #554.

Update (base: develop). Merged current develop. Develop's discriminator_stability revisions 17–19 are kept byte-for-byte; revision 19 is archived and this task is added as revision 20 (Tier 2 stretch, order 107). The unmerged 2026-11-17 fixed test list is replaced by technique-inventory-full-suite-2026-11-20.json (develop's 69-test 2026-11-18 list + this task = 70). Because the host binds clean EMPF only, the SigReg admission test records it as an explicit, checked exclusion (preflight must still report BLOCKED). The word-hold budget test keeps the historical 76,620 s total and declares this task's own 600 s reservation separately. Catalog and Forge reports regenerated with the repo's tooling.

Develop merge — 2026-11-20 (base: develop)

  • Merged origin/develop into this branch. No rebase or force-push was used. Develop's view revisions 18 (Theory-3 isolation stretch) and 19 (Theory 4A stretch) are unchanged byte for byte. The exact revision-19 bytes are archived at configs/forge/view-history/discriminator_stability-v19.json, and this PR's ring change is now revision 20. It swaps only ring16_acquisition for ring16_acquisition_6400, and the previous-revision fingerprint is checked by the tests.
  • I took develop's copy of each conflicting generated file, then rebuilt it with Forge tooling (forge history, forge compile --summaries-only, reports/forge/regenerate_technique_inventory.py --refresh-publication). This covers the catalog, memory, compilation, technique inventory, family pages, problems and EXPERIMENTS_BY_TIER. No training or regrading was performed.
  • reports/forge/adopt-20261016/results.json stays in Git. It holds 130 per-attempt final-gate records, each with its evaluator verdict, metric checks and receipt SHA-256s. It contains no per-update stream or trace. It is the compact receipt that research_tiers.adopt_rows validates and renders, so it is not bulk log data.
  • Relevance of the K3P/KA2 no-LR-cap cards: they are family-level selected-recipe revisions (recipe-prior-refactor-{k3p,ka2}-v2), and default_adoption: false. They do not touch the package default. Develop's DEFAULT_RECIPE is currently empf; feat: promote rhbi_optimized default recipe, bound large-table simplicial memory, and prepare 0.10.0 #549 had promoted rhbi_optimized, and develop has since moved back to empf. Whichever is the default, these cards still matter only for the K3P and KA2 rows on the 6,400-update ring question. They have not been measured on current-source revision 20 evidence (the status is still awaiting_measurement). The historical adopt diagnostics (16/29 and 15/29, seed 0) do not justify adopting them as the default. They should stay as declared family revisions until a new study measures them. No defaults were changed.

🤖 Generated with Claude Code

mikkel and others added 7 commits November 14, 2026 19:47
…nger wall-clock limits

User decisions of 2026-10-16 (dig2 round 2), as new revisions; old identities kept.

- ring16: port ring16_acquisition_6400 from #433 (same gate, same 5-check
  sustained rule, same cadence); discriminator_stability revision 10 requires
  it; revision 9 archived in view-history. Atlas/E22 scoped views untouched.
- K3P/KA2: recipe-prior-refactor-{k3p,ka2}-v2 = v1 with
  network_lr_horizon_cap None; family-revisions-20261016.json declares them as
  the families' next selected recipes (validated by
  trainer_families.load_declared_revisions; forge inventory runs them). Measured
  pins in family-current-v1.json are unchanged until the new rows are measured.
- Wall-clock: port #434 (NOT MEASURED (wall-clock) outcome, contention-aware
  limit, in-adapter allowance) and make it the default with raised operational
  limits from measured per-update times (word smoke 1800 s, word hold 600 s,
  grid/rotated/staggered100 7200 s, archived ring16 600 s), one extension up to
  3x, no extension without progress or when the run cannot finish. Applied in
  the queue worker and the selected attempt intake; policy-cloud keeps declared
  limits; FORGE_WALLCLOCK_POLICY=declared opts out. Declared timeouts and update
  budgets are unchanged. drain --no-publish skips the board compile hook.
- Campaigns adopt-20261016-<group>-v1 for per-family measurement.
- No training run; published leaderboard not regenerated.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Replace the standalone adopt campaign files with ready studies
adopt-20261016-<group>-study (discriminator_stability revision 10, Tier 3,
RTX A6000 CUDA cohort, campaign adopt-20261016-<group>-v1, 120,000 s). K3P and
KA2 use their v1 recipe as control. Verified by enqueueing all six into a
scratch queue (no drain).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…cap, wall-clock limits

Imports the verifier-checked cells of the six adopt studies (source 1e02f18,
seed 0, plan/enqueue/drain --no-publish) into reports/forge/adopt-20261016/
results.json via collect.py, renders them as source-bound diagnostic research
rows (research_tiers.adopt_rows), and regenerates the leaderboard with
--refresh-publication. Totals over the 29 measured tests: K3P v2 16, KA2 v2 15,
R1/R2 16, GAN v3 15 (2 Tier 3 blocked by tier veto each); BCAP and BISPF 7/8
measured (21 not run, campaigns paused). No wall-clock hit; nothing NOT
MEASURED. Selected pins and recorded qualification are unchanged; their
1,600-update ring16 verdicts now show as not run under revision 12.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…dence

Publish the ring-only change as view revision 18, retain interpolation and precision contracts, preserve fixed-suite results separately from the unmeasured new task, and restore all reviewed merge-only contributions. No scientific training or qualification regrading.
@mikkel
mikkel force-pushed the task/adopt-ring6400-no-lrcap-20261016 branch from 2e84eb0 to 1e1b4b3 Compare November 14, 2026 20:18
mikkel and others added 4 commits November 20, 2026 05:20
Keep develop's revision 18/19 history byte for byte, archive exact revision-19
bytes, and swap only ring16_acquisition for ring16_acquisition_6400 in
revision 20. Generated Forge files take develop's version and are rebuilt with
Forge tooling.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rebuilt with reports/forge/regenerate_technique_inventory.py --refresh-publication
and forge compile --summaries-only; no training or regrading. Problem text and
adopt test now name revision 20.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant