LittlePlyModel + harness

CM-16 / outcome recorded 9 November 2026

A multi-day experiment. An unsuccessful replacement.

A record of what ran, what changed and what the evidence can support. CM-14 remains the learned model inside the playing agent.

CM-16 / multi-day attempt / not promoted

More objectives did not produce a better replacement.

CM-16 explored whether next-move training could benefit from extra supervision: win/draw/loss, a sampled future move, and board occupancy. After the original November 4 run and token-interface adaptation, the selected continuation ran from November 5, 09:42 UTC to November 9, 17:06 UTC. That is 103.4 elapsed wall-clock hours, including interruptions, not GPU-hours.

The retained trajectory reached update 665,007 and 5.481 billion ordinary target visits. Visits include repeated examples and are not a count of unique positions or games. The experiment was stopped for assessment and did not replace CM-14.

What the playing evidence says

The early periodic studies below consistently lost their direct comparisons with CM-14. A later matched CPU full-system comparison at CM-16 update 369,403 scored 33W / 2D / 15L against Sunfish (68% over 50 games), versus CM-14's 49W / 1D / 0L (99%) on the same 25 opening pairs. Its separate direct match against CM-14 was 8W / 0D / 42L (16% over 50 games). These are results for that intermediate checkpoint, not final update 665,007.

Final assessment limit. Later latest-checkpoint GPU cohorts failed their audit because required GPU telemetry was missing. The assessment was then stopped. Those results are excluded from the headline; no clean final-checkpoint strength result is claimed. “Unsuccessful” means the training experiment did not establish a better replacement for CM-14.

A compute-budget hypothesis, not a causal finding

My working hypothesis is that, with the compute available here, the extra objectives made learning the main chess task harder: more work per useful next-move update, more targets competing for the same model capacity, or both. This remains a hunch informed by the experiment. The run does not isolate those mechanisms or show that auxiliary objectives are generally harmful.

A useful test would hold data, architecture, optimizer, serving harness and evaluation fixed, then compare next-move/WDL training with and without each added objective under matched wall-clock and target budgets. Multiple seeds and fresh openings would help distinguish a repeatable effect from this development history.

What changed during the run

Data exposure, objective weights, output semantics, training duration and evaluation protocols all changed. The later two-task loss normalization is not directly comparable with the earlier four-task aggregate. Continuing the same weights after removing objectives is not a from-scratch ablation.

Eight earlier periodic evaluations, with their original sample sizes
CM-16 periodic results. W / D / L; score counts draws as half a point.
PeriodReported UTCSunfish gamesSunfish W / D / LScoreAgainst CM-14, 25 games
12026-11-0510031 / 7 / 6234.5%4 / 2 / 19 (20%)
22026-11-0510032 / 8 / 6036.0%4 / 1 / 20 (18%)
32026-11-0610030 / 11 / 5935.5%2 / 2 / 21 (12%)
42026-11-065018 / 4 / 2840.0%1 / 2 / 22 (8%)
52026-11-065019 / 3 / 2841.0%6 / 4 / 15 (32%)
62026-11-075020 / 4 / 2644.0%3 / 4 / 18 (20%)
72026-11-075021 / 4 / 2546.0%2 / 5 / 18 (18%)
82026-11-075024 / 5 / 2153.0%6 / 4 / 15 (32%)

Periods 1–3 used 100-game Sunfish cohorts. Periods 4–8 changed to a 50-game clock-aware search protocol without helpers. Do not connect these rows as one controlled curve. The separate head-to-head protocol stayed at 25 games: 12 color-reversed opening pairs plus one predeclared CM-16-black game. These development openings had prior exposure. They are not final-checkpoint results.

Dated outcome, checkpoint identities and source hashes · Matched CPU full-system comparison · Historical adaptation recipe

The original recipe, preserved in context

This is the dated November 5 plan and earlier observations. It records the choices made then; the outcome and later amendments above supersede its prospective status wording.

Read the historical adaptation plan, launch record and original sources

CM-16 launch receipt · CM-15 results · Play LittlePly on Lichess

Recipe decision · 5 November 2026 · unpromoted

Adapt the interface.
Then train the whole network.

The chosen CM-16 recipe now starts with a three-hour embedding-only adaptation phase, followed by three hours of full-network training and a checkpoint comparison. This describes the selected plan, not a new execution or evaluation receipt.

CHOSEN RECIPE · BENEFIT UNMEASURED

A better starting interface is the hypothesis.

The chess vocabulary is new while the GPT-2 transformer body is pretrained. The hypothesis is that adapting the token interface first can align chess inputs and outputs with potentially useful pretrained computational structure, before the whole network adapts to chess.

This does not establish that GPT-2 already has strong chess knowledge, that the interface can access it, or that adaptation improves subsequent training.

A heuristic, not an optimum

Three hours is a pragmatic choice, not a scientifically established optimum. The duration and benefit remain hypotheses. Later studies can compare adaptation lengths and matched training budgets. No measured improvement is claimed here.

Phase 1 · three hours

Train the tied token table.

Train only the shared input/output chess-token embedding table. Input embeddings and output readout stay tied: they are one shared parameter table, not two independently trained interfaces.

Freeze the pretrained transformer body and positional embeddings. This phase adapts the new chess-token interface while preserving the pretrained computation and positions.

Phase 2 · three more hours

Resume full-network training.

After the three-hour adaptation phase, resume training the whole network for another three hours, subject to normal health and numerical safeguards.

Frozen-stage playing strength does not gate unfreezing. This is a staged training recipe; the informative playing comparison is planned after the full-network phase.

Then · checkpoint and compare

Compare with the archived original.

Save a checkpoint after the three-hour full-network phase and compare it with the archived original CM-16 baseline. Report actual checkpoint identities, completed games, protocol and uncertainty when those results exist.

Further training choices and any promotion remain Andrei’s decision. CM-14 stays deployed; CM-16 remains unpromoted.

Three hours is not the 3h18m baseline.

PathTimeEvidence boundary
Archived original baseline3h17m38s, rounded to 3h18mStopped snapshot recorded 5 November at 02:10:01 UTC: update 18,170; 149,765,149 ordinary targets.
Embedding-only adaptation3h plannedTied input/output table trainable; transformer body and positions frozen.
Full-network phase after adaptation3h more plannedCheckpoint and comparison after this phase; approximately six planned training hours across the two phases.

Unequal budgets do not isolate causation. The proposed comparison is roughly six hours of staged training versus a 3h18m original baseline. A difference could reflect additional training time, the staged recipe or other differences. It would not by itself establish an isolated causal benefit of embedding adaptation. The durations are not matched compute budgets.

The original launch and early measurements below remain historical evidence. Their clocks, checkpoint schedule, optimization benchmarks and conditional runtime estimates are not measurements of this new adaptation path.

Baseline identity and recipe-decision record

Archived baseline checkpoint SHA-256: d8967de191f09fa9513d68e6fabb825ca1f3a5fa0d3fdfc1931ec04815258285

The stopped public snapshot reports 11,858.088 seconds of elapsed time. This is baseline provenance, not a playing-strength result.

Download the recipe decision and selected public baseline fields. This record does not assert an adaptation start time, completion, score or measured gain.

Historical baseline recipe and observations follow. The original full-network run’s stopped record is distinct from the newly chosen adaptation path above. Statements below describe their dated 4 November snapshot, not current process health or the new phase’s progress.
Status snapshot · 04 Nov 2026, 23:05 UTC In training · unpromoted

Full training started 04 Nov 2026 at 22:52:23.587 UTC from pristine GPT-2-body/chess-token initialization, not experiment weights. The independent audit verified 982 optimizer updates / 8,095,986 ordinary targets at 23:03:06 UTC, with finite states, the correct two ranks, no alerts. This is a dated snapshot, not a live counter.

982Updates · audited 23:03:06 UTC
8,095,986Ordinary targets · snapshot
132 hConditional central case
167 hSlow + 10% planning case
Historical full-run launch · audit snapshot 04 Nov, 23:03:06 UTC

The original baseline.
Keep its evidence separate.

Independent launch audit passed: finite execution, the correct two ranks, no alerts at 23:03:06 UTC. The latest audit at 23:03:06 UTC reports GPU cores at 49°C / 48°C and finite preclip norm 6.544; these are dated readings, not live telemetry. Sustained full-run throughput and MFU are not yet measured.

IN TRAINING · UNPROMOTED

8,095,986 of 4,527,358,926 ordinary targets.

About 0.179% of the two-epoch ordinary-target goal at 23:03:06 UTC. This is source coverage, not total task examples, labels or elapsed-time progress. The objective slider below remains an illustration.

Started 22:52:23.587 UTC from the unchanged pristine initialization. Disposable benchmark and optimization weights were not used.

Recovery and protection

Logical committed-trajectory recovery is selected: restore model, optimizers, ledgers, cursor and RNG together. Discarded computation may replay; retained two-pass ordinary coverage and independent per-task caps remain exact. No automatic promotion or live serving change.

Next checkpoint · planned

05 Nov, 00:52:28 UTC

Retain a model-only snapshot at every 2-hour checkpoint. Full restart generations are bounded separately: initial, latest 2, latest 2 daily and final, maximum 6.

No scheduled 2-hour checkpoint has completed in this snapshot. An initial restart generation or save/restore smoke is not a completed cadence checkpoint.

First evaluation · planned

05 Nov, around 04:52 UTC

Then every 6 hours: 100 Sunfish games + 25 games against a frozen live-CM14 baseline, followed by reported findings. Helper/conversion smoke passed and one evaluation scheduler is registered.

No full production evaluation has completed yet. Cadence is approximate: evaluation is queued after a checkpoint in each 6-hour bucket. Failure pauses evaluation and alerts while healthy training continues. Extended downtime can skip historical evaluation buckets; missed buckets must be explicitly reconciled and reported.

Operational safeguards and bounded recovery guarantees

Audited local safeguards: stop training below 60 GiB + 5 GB free space, above 65 GB run size, after repeated GPU temperatures of at least 88°C, after 15 minutes without an update, or on nonfinite state. At 23:03:06 UTC, disk was 151.52 GB free and the run occupied 2.60 GB. These are dated observations, not current counters.

Producer test: 48 plans, using a simulated 0.7-second consumer; total steady waiting was 0.636 seconds across 38 plans. Cold rebuild was exact. This supports functionality but is not measured week-scale producer throughput.

Cache bounds: depth 8 and strict 32 MiB per-plan checks keep prepared data below 288 MiB, including a temporary write. The independent audit at 23:03:06 UTC reports 7.64 MB cached. An independent monitor enforces a 2 GB cache ceiling, excluding both exposure ledgers; both ledgers still count within the 65 GB whole-run budget. No trainer restart or training-math change was made.

Bounded durability guarantee: an independent guard validates generation hashes and fsyncs a durable latest-generation pointer. Startup acknowledgement was confirmed. Durability applies after that acknowledgement, not at the instant a new generation is published. Missing, stale or invalid pointers, incomplete stages and corrupt-generation recovery passed tests. Stopped-job recovery reconciliation refuses to run while training is live. Independent audit confirmed the cache monitor and durable startup acknowledgement. Future cold restart must first reconcile the durable pointer; automatic restart is disabled. These bounded guarantees do not certify every arbitrary crash.

Evaluator scope: 1,990-token adapter, current helpers and 25-game summary repair passed smoke tests. The first full 125-game evaluation remains future. Six-hour cadence queues after a checkpoint in the corresponding bucket, so timing is approximate. Evaluation failure pauses evaluation and alerts while healthy training continues. Extended downtime can skip historical buckets; missed evaluations require explicit reconciliation and reporting.

Launch integrity receipt and scope

Independent audit passed. Actual start: 22:52:23.587 UTC. The immutable launch receipt records 67 updates / 552,336 ordinary targets. A separate early owner health snapshot recorded 104 / 857,217. The latest independently audited snapshot is 982 / 8,095,986 at 23:03:06 UTC. These are different observations, not conflicting counts. None is a sustained-throughput or playing-strength result.

Pristine initialization SHA256: 051e474f45711824f4120bb013d891f1df4252aa30e347860aaafb8e340ebfa3

Sealed recipe SHA256: 29d1a3b50a1553dcc6dc3ed3ecf4be049177ae2ce226a7a238505fd8e20e3415

Certified corpus SHA256: 8f871184d45438b8822528fd63df98cbf2c8a6c5af6bcdb3a92e2d9fbcbe502b

Canonical corpus digest: 26e9e867613e2a6ca3d8dbe939cc6de24f859898e9a2b70613b1eb02684268f3

Public launch-receipt SHA256: 3fb4127dd36f0a8f03972fbbe0641de45f85cbce56a0799ca67a97509238230f

Conditional runtime remains 132–167 hours, about 5.50–6.96 days. This is planning evidence from prior tests, not measured sustained full-run speed, a confidence interval or a guaranteed ceiling.

Completed isolated optimization comparison · 04 Nov

Prefetch helped.
Larger rows were not the winner.

Each passing case used 6 warmup + 30 measured updates. Initial logits, gradients and target counts matched for the selected comparison. Global examples, effective batch, masks, outcome weights and precision were held fixed.

SELECTED: 8 ROWS + BOUNDED PREFETCH

9,528 ordinary anchors/s · 12,513 task examples/s.

A 16.0% gain against the matched 8-row uncached control, at the same 5.06 GB peak rank-0 allocated memory. This is short-test evidence, not a guarantee of sustained week-scale throughput.

Earlier isolated tests

The isolated experiments finished at 22:31:04 UTC and released both GPUs then. Pristine initialization and v2 hashes were unchanged. The later full run started separately at 22:52:23.587 UTC from pristine state. These test rates are not full-run rates.

Execution caseOrdinary anchors/sTask examples/sPeak rank-0 allocatedOutcome
8 rows · uncached8,21010,7835.06 GBMatched control
16 rows · uncached8,35210,9698.45 GBPassed; modest gain, more memory
32 rowsNot measuredNot measuredNot reportedRejected by fixed numerical gates before any update; tolerances unchanged
8 rows · cached, blocking7,1679,412Not reportedSlower than matched control
8 rows · cached + prefetch9,52812,5135.06 GBSelected · +16.0%

The 16.0% comparison is against this matched short-test control, not against the earlier 30-minute pilot’s phase rates. Those measurements used different windows and are retained below as historical evidence. The 32-row case was rejected on numerical gates, not reported as an out-of-memory result.

Cache evidence and scale limit

Wait time fell from 7.45s to 1.30s.

The compact tested cache was 4.73 MB, with 36/36 hits. Building and verifying it took 29.12 seconds. A warm, fully hit cache does not establish that a rolling producer can continuously feed the complete dataset.

Continuous producer performance must now be observed during the run. Actual launch uses depth-8 rolling preparation with 8 rows/rank. The 2 GB cache ceiling is now monitored separately from both exposure ledgers, which remain within the 65 GB whole-run budget. This is a bound, not evidence of warm-cache performance across the full corpus.

Resources and evidence scope

Storage and monitoring are part of the run.

At experiment completion: 160.35 GB free disk; new artifacts 2.23 GB. A core temperature of 46°C was sampled. Hotspot/junction temperatures and throttle flags were not captured.

MFU remains unmeasured. GPU busy-time utilization is not model FLOP utilization. Neither the earlier 73% / 72% device-utilization sample nor a warm cache establishes MFU or compute saturation.

The native production two-update test ended at 22:10:07 UTC. All 48 Muon + 100 AdamW states stayed finite and synchronized. Atomic model/optimizers/ledger/cursor/RNG restore reproduced the exact next batch. Global accumulated-gradient parity: relative L2 6.55×10⁻⁸; cosine ≈1; maximum absolute difference 3.05×10⁻⁵, within unchanged tolerances. This is actual full-model evidence, distinct from tiny double-precision reference tests. It does not certify sustained throughput or arbitrary-crash recovery at every failure point.

Certified default · anchor-count calculator

Two epochs. Scale by unique ordinary-eligible anchors.

At the measured hardware/settings and representative distribution: approximately 5.83 training hours per 100 million unique anchors; slow + 10% planning case 7.38 hours. The default A is the certified frozen union: 2,263,679,463 anchors. Other inputs are illustrative scenarios.

million

Central training-time scale132.00 h5.50 days
Conservative planning case167.00 h6.96 days
Ordinary examples across 2 epochs4,527,358,926Exactly 2A; auxiliaries are additional
Conditional planning, not a confidence interval or guaranteed ceiling. This calculator scales the latest owner-reported full-corpus forecast: 132 hours central / 167 hours slow + 10%, at the certified A. It is not a direct 16% multiplier on the older estimate or a new benchmark generated by this page. The forecast assumes the selected pipeline and representative full-union source/history/task distribution. Continuous cache supply, production integration, final settings and real checkpoint/evaluation overhead can change the result.

This calculator does not infer games, total task examples, active labels or physical rows from A. Exact auxiliary counts need the sealed schedule. The 132 / 167-hour baseline is rounded; scaled scenarios are approximate, not extra timing precision. The objective slider below is an independent illustration: moving it does not change these unmeasured runtime assumptions or recompute a new benchmark.

00 Earlier corrected pilot · completed

Earlier phase benchmark, retained for comparison.

About 12,000 task examples/s on two RTX 4090s. This earlier corrected trial remains valid for its measured pilot scope. The newer matched optimization comparison above selects bounded prefetch. Do not compare the absolute rates as though the windows were identical.

CORRECTED TRIAL COMPLETE

27m 14.84s, within the 30-minute budget.

Started 20:35:11 UTC · training work ended 21:02:17 · completion receipt 21:02:20.682 · guardian exited 21:02:26.701 with code 0. Independent terminal observation confirmed exit and no benchmark/GPU compute process at 21:05:37, with matching reported rates. Pristine initialization is unchanged.

Scope of this result

Isolated disposable training only. No full CM16 run or model deployment was launched by these tests; the optional asynchronous board-readout monitor remains off. The first trial remains INVALID-FOR-ETA and contributes no rate or projection here.

early mixture11,941task examples / second

10,748 ordinary anchors/s

Actual example shares
Ordinary90.014%
WDL2.495%
FUTURE82.495%
Board4.996%

Median 12,033 · P10 11,871 task examples/s

middle mixture12,137task examples / second

9,407 ordinary anchors/s

Actual example shares
Ordinary77.511%
WDL8.747%
FUTURE88.747%
Board4.996%

Median 12,335 · P10 11,653 task examples/s

late mixture12,070task examples / second

7,846 ordinary anchors/s

Actual example shares
Ordinary65.007%
WDL14.998%
FUTURE814.998%
Board4.997%

Median 12,163 · P10 11,899 task examples/s

Warmup excluded from the phase rate comparisons. These are shortened representative mixture windows, not evidence that the complete two-epoch source schedule finished during this trial. Rounded actual shares may sum to 100.001%. Detailed tables scroll horizontally on phones.

Measured work and units

19,399,839 consumed task examples.

Ordinary examples / anchor uses
15,116,473
WDL examples
1,657,069
FUTURE8 examples
1,657,069
Board examples
969,228
Optimizer updates
1,834
Game visits
177,961
Packed physical rows
230,171
Non-padding tokens
≈90.51 million
Active labels
≈81.43 million
Physical slots, including dummy rows
≈118.14 million
Padding slots
≈27.63 million

A game contains many anchors. An anchor can support several task examples. A board example has 65 active labels but counts as one example. A packed row is a physical container, not a game or a learning task.

End-to-end: 11,867 task examples/s and 9,246 ordinary anchors/s, including the measured terminal elapsed interval. These are not the warmup-excluded phase rates.

Timing decomposition

1,610.59s in measured steps.

ComponentTimeStep share
Data300.07 s18.63%
Packing + transfer120.24 s7.47%
Forward354.50 s22.01%
Backward659.36 s40.94%
Optimizer115.62 s7.18%
Remaining step overhead60.80 s3.78%

24.25s outside measured steps brings elapsed time to 1,634.84s. Recorded initial validation: 0.578s; final validation: 0.053s; checkpoint save: 2.371s. These are not extra additions to the terminal total.

Validation used only 8 fixed games. These short validation timings are not a production validation-cost estimate. Full-run cadence and checkpoint/ledger scale need separate measured treatment.

Pilot and holdout evidence

9,343,433 unique training anchors.

The manifest contains 110,000 canonical-deduplicated games / 9,349,287 ordinary anchors. Reserving 64 local-validation games leaves 109,936 training games / 9,343,433 unique ordinary-eligible anchors. These are pilot figures, not a full-union count.

Elite training games
99,945
Acquired-human training games
8,991
Engine training games
1,000
Consumed Elite ordinary targets
13,768,458
Consumed human ordinary targets
1,188,178
Consumed engine ordinary targets
159,837

The pilot was globally shuffled; phases did not use separate source blocks. History-length distributions were similar across phases, mostly 50–149 plies.

Historical holdout audit passed against 8,282 frozen games: zero original-key, complete-history and exact first-30-ply overlaps. Additional family exclusions rely on sealed-source receipts. These pilot checks do not by themselves certify all transpositions. The later complete frozen-union audit is reported separately in the source section.

Ordinary cursor reconciliation passed; FUTURE8 destination maximum was 2; WDL/FUTURE8 quota skips were 0; board eligibility shortfall was 41. The timed trial covered about 1.618 ordinary passes, not two full pilot epochs; separate coverage/cap tests exercised exact two-epoch behavior. Board examples were about 4.996% of all examples, 763 below floor(5% of the total); the 41 reported eligibility shortfalls do not explain that entire gap. Per-update quota rounding/guard effects also apply; no exact decomposition is claimed.

Hardware and stability

Two RTX 4090s, 24 GB each.

TelemetryGPU 0GPU 1
Maximum core temperature53°C47°C
Median utilization73%72%
Peak driver memory6,361 MiB5,744 MiB
Peak power249 W231 W

Rank-0 allocator peak: 5.06 GB allocated / 5.30 GB reserved. No hardware thermal slowdown was reported. Hotspot, memory-junction, CPU-package and NVMe temperatures were unavailable.

torch 2.9.0+cu128 · Transformers 5.21.0. Finite synchronized optimizer states: 48 Muon and 100 AdamW.

Every steady update clipped at norm 1. Finite execution is not proof of optimal stability or quality. The target 8,192 ordinary anchors/update, whole-game rounding, 8 rows/rank and variable accumulation were not VRAM-forced in this earlier pilot. The later matched comparison selected 8 rows + prefetch. Clipping evidence remains visible; clipping alone is no longer a launch blocker.

Validation scope, terminal integrity and the excluded first trial

Repaired history path: the overflow-history bug was reproduced and corrected. Multi-row regression and packed-gradient equivalence checks used tiny double-precision reference models, including overflow. Later actual full-model evidence includes two-GPU BF16 finite execution, rank agreement, committed-checkpoint exact resume and global accumulated-gradient parity within unchanged numerical tolerances. The native two-update receipt is summarized below; tiny-reference tests remain a distinct scope.

Eight-game teacher-forced sanity panel: final CE ordinary 3.0443; WDL 1.0047; FUTURE8 5.6221; board 0.3948. These are pipeline sanity measurements, not whole-board free-generation accuracy, internal-state proof or playing-strength evidence.

Terminal integrity: guardian exit code 0; elapsed 27m 14.84s inside the 30-minute budget; no benchmark/GPU compute process at 21:05:37; pristine initialization unchanged. The disposable checkpoint is not the pristine initialization and has not been deployed.

Pristine initialization SHA256: 051e474f45711824f4120bb013d891f1df4252aa30e347860aaafb8e340ebfa3
Disposable corrected checkpoint SHA256: 2f585f10f3cec935cc32857a1ccd607f1e992607c4eb0a7ab568775bfe0d0b12
Corrected packer SHA256: 070d61ceb3e7baa95d5fd705dbec3c3d6ca4af5007eb991b398a99dc70fc0554

First trial: INVALID-FOR-ETA. It started at 20:23:35, saved update 485 at 20:29:56 and exited at 20:30:06 after the history bug. Its logs/checkpoint remain invalid for ETA and none of its rates are used here. The corrected run above started separately at 20:35:11.

Earlier pilot log and static-code audit

Clipping is substantial.
The cause is not isolated.

Every logged update in that pilot clipped at norm 1. All states stayed finite and a tiny sanity panel improved, but this is not evidence of optimal convergence. Clipping alone is accepted as non-blocking. Clip 1 is retained; the completed execution/cache comparison above changed neither the clipping threshold nor the learning-rate plan.

Observed

Median raw-gradient scales: 15–28%.

Measured windowUpdatesMedian preclip normMedian raw-gradient factorClipped
Warmup2325.4600.1831100%
Early, after warmup5456.4810.1543100%
Middle5464.7660.2098100%
Late5113.5740.2798100%

Initial preclip spikes, updates 1–6: 708.336 → 499.543 → 236.104 → 118.277 → 77.036 → 55.773. Last 100 updates: median 3.601, maximum 5.689. Norms are logged before clipping.

At threshold 1, a finite norm g greater than 1 is rescaled by approximately 1/g. This modifies the raw gradient, not directly the learning rate or parameter update. PyTorch clipping reference.

Do not read 0.154 as “15.4% of the learning rate.” AdamW uses running first/second moments; Muon orthogonalizes matrix updates. Their update magnitudes are not simple multiples of the clipped raw norm. AdamW algorithm · Muon update analysis.

Meaning and limits

Accepted as non-blocking; no observed divergence.

The fixed eight-game ordinary CE improved from 13.885 to 3.044; all optimizer states stayed finite. This establishes a narrow pipeline sanity result, not good long-run convergence, free-board accuracy or playing strength.

No obvious scaling, ordering or duplicate-parameter defect was found in static review: one clip after accumulation/DDP synchronization; global-update denominators; world-size compensation for DDP averaging; no extra accumulation divisor; no sum of norms across replicas; tied parameters counted once; optimizer partitions disjoint and exhaustive. That static review alone is not full-model numerical gradient parity; a later native functional test separately passed global accumulated-gradient parity within fixed tolerances. DDP reference.

Why it happens is not proven. Fresh chess-token/readout calibration against a pretrained body is a plausible hypothesis, not a finding. Initial ordinary CE exceeded uniform 1,990-class CE (about 7.596), but initial logits, per-task/per-layer/group norms and update-to-weight ratios were not logged. Phase trends mix training progress with changing task shares.

For context only: 50 periodic late-CM15 records had median norm 0.705 and no clipping. This was not a matched fresh-model stage, task mixture or batch comparison.

Critical measured-setting distinction

180-second wall-clock LR warmup.

The actual trial ramped optimizer LR from 0.1× to 1× peak over 180 seconds, with no GradScaler. The same window was omitted from steady throughput comparisons. This is different from the actual full-run first-5%-of-source-progress warmup.

The after-warmup optimizer LR trajectory was not extracted in this audit. The short trial did not validate a week-scale source-progress LR trajectory. The auxiliary objective-share ramp is separate from both LR schedules.

What was tested, and what remains diagnostic work

Completed: matched 8/16/32-row execution tests and blocking-cache versus bounded-prefetch tests. The 32-row case failed unchanged numerical gates before an update; the passing selected 8-row prefetch case gained 16.0%. See the exact results above.

Still proposed diagnostic probes: group-gradient and per-layer/per-task attribution where useful, initial logit spread/entropy and parameter update-to-weight ratios at matched batches. These were not logged by the earlier clipping audit and are not silently claimed completed.

The earlier benchmark spent 420.31 seconds, about 26.1% of measured step time, on data plus packing/transfer. Input overlap helped in the subsequent matched tests, but this does not mean all such time is removable. PyTorch performance-tuning guide.

No clipping threshold or LR change was made. Clipping alone is accepted as non-blocking; continue reporting finite-state, norm and quality evidence without treating raw-gradient factors as parameter-update or LR factors. GPU busy-time utilization is not MFU, which remains unmeasured.

01 Complete preparation recipe

The original specification, with its gaps visible.

Everything below separates the requested plan, historical records and completed portable checks from actual CM16 integration. An open field is not silently filled with a historical default.

Last updated · 04 Nov 2026, 23:05 UTC

CM16 is in training and unpromoted.

Actual start 22:52:23.587 UTC. Independent verification at 23:03:06 UTC: 982 updates, 8,095,986 ordinary targets, finite execution, correct two ranks/no alerts. Fresh pristine initialization was used; no experiment weights.

ACTIVE · NO PROMOTION

Two retained ordinary epochs target 4,527,358,926 ordinary examples. Checkpoints every 2 hours and 100-Sunfish + 25 frozen-live-CM14 evaluations every 6 hours are scheduled plans; no scheduled 2-hour checkpoint/full evaluation has completed yet.

Actual pipeline receipts · distinct from portable reference tests

Access and pilot data

Actual source records are readable and the frozen corpus is now certified. The benchmark environment was torch 2.9.0+cu128 / Transformers 5.21.0 on two RTX 4090 GPUs; initialization and native restore checks are summarized alongside the data evidence.

Reported canonical-deduplicated pilot census: 110,000 games / 9,349,287 ordinary anchors. The final 64 pilot games are excluded from training. After excluding those 64 games, the training pilot is 109,936 games / 9,343,433 unique ordinary-eligible anchors. Neither census is a full-union count or total emitted task-example count.

  • Elite: 100,000 games / 8,519,885 anchors.
  • Acquired human: 9,000 / 731,597.
  • Engine: 1,000 / 97,805.

The frozen full union is now certified separately: 26,703,686 games / 2,263,679,463 ordinary anchors. The earlier pilot audit found zero original-key, complete-history and exact first-30-ply overlaps against 8,282 frozen historical games. Additional family exclusions rely on sealed-source receipts.

Actual initialization and smoke receipts

The owner reports that the Transformers 5.21.0 adapter checked every parameter and 12 causal buffers; actual source/target hidden states matched exactly. All 1,990 chess/board rows were freshly initialized and tied, the first 512 position rows were copied, and native cold-restore logits matched exactly.

The auxiliary overflow-history bug was corrected and multi-row regression checks passed. Packed gradient equivalence was tested on tiny double-precision reference models, including overflow. Actual full-model checks establish two-GPU BF16 finite execution, rank agreement, committed-checkpoint exact resume and global accumulated-gradient parity within unchanged tolerances. Separate coverage/cap tests exercised exact two-epoch behavior.

Pristine initialization SHA256:
051e474f45711824f4120bb013d891f1df4252aa30e347860aaafb8e340ebfa3

SPECIFIED requested contract   PROVISIONAL needs sealing   UNVERIFIED actual evidence missing
On phones, detailed tables scroll horizontally.

A Scope, epochs and independent budgetsSPECIFIED

Run scope: all eligible available engine, Elite-human and fresh Lichess games. Exactly two source epochs; no eight-hour ceiling. Completed disposable tests are distinct from a full-run launch. The authorized full run started on 04 Nov at 22:52:23.587 UTC from pristine initialization; its independently audited early status is reported above.

What one source epoch means

  • Traverse each unique eligible canonical game once per epoch. Supervise every eligible ordinary next-move anchor once on each pass, exactly twice overall, without outcome-biased selection. Preserve complete clean histories, provenance and heldout exclusions.
  • Multiple isolated task examples may be emitted from a source visit. Two traversals do not mean at most two examples total across the four tasks.
  • Use deterministic source-stratified ordering. Do not inherit equal-month exhaustion repeats or hidden engine oversampling from earlier runs.

What one task exposure means

  • Identity: canonical game + semantic source anchor + learning task.
  • Every eligible ordinary anchor uses its two next-move credits exactly, one per epoch. WDL, FUTURE8 and board each retain an independent maximum of two semantic uses. An auxiliary example spends no next-move credit.
  • Ordinary target destinations have their own two-use limit. FUTURE8 destination moves have a separate two-use limit within FUTURE8, including collisions across overlapping windows.
  • A whole board answer is one exposure. Repeated WDL class tokens are not global collision identities. Identical boards from different histories are not automatically one sample. Logical committed-trajectory rollback/replay is selected below; discarded physical computation may repeat without adding retained examples.

Must be sealed: actual scheduled per-task allocations and complete two-epoch ordering; the certified source and eligibility census is available, but scheduled auxiliaries are not its capacity counts; exact one-per-epoch ordinary coverage, auxiliary reuse, destination collisions, natural outcome counts and normalized loss-coefficient totals. Any additional prefix/state dedup policy needs its own declared identity.

B Exact coverage, weighted next-move loss and mixtureREVISED PROPOSAL

Latest change: assemble a deterministic two-source-epoch dataset with every eligible ordinary anchor exactly twice, once per epoch. No outcome-biased selection, winner oversampling or winner-favored repeat allocation. Preserve loser moves and recorded game-ending moves; a terminal state with no recorded next move has no ordinary target.

Selection stays unchanged; only ordinary loss changes

Use tentative centered raw next-move CE weights winner 4/3 / loser 2/3 / draw 1. Determine the class from the actual mover and the recorded game result, not fixed White perspective. A White mover in a White win gets 4/3; a Black mover in that game gets 2/3. A Black mover in a Black win gets 4/3. Draws get the centered neutral raw coefficient 1. These values preserve a decisive winner-to-loser ratio of 2:1 and are not claimed optimal.

WDL and FUTURE8 keep unit example losses. Board keeps the mean CE over its 64 cells + END, then unit example weight. This outcome weighting never applies to auxiliary labels, padding, ignored labels or history tokens.

Draw weight 1 sits midway between the decisive raw weights 4/3 and 2/3. Global ordinary normalization can still move its final factor above or below 1; “neutral” describes the centered raw convention, not a guaranteed normalized factor.

Normalize across the complete optimizer update

Let M be ordinary-example count, S the sum of their raw weights, and N all consumed examples in the complete global optimizer update, across every rank and accumulation microbatch.

ordinary factorᵢ = wᵢ × M / S
loss = (Σ ordinary factorᵢ × CEᵢ + Σ auxiliary example losses) / N

Ordinary factors sum to M, so each task retains its scheduled example-coefficient mass. Do not divide by the global sum of raw weights, and do not normalize separately per rank or microbatch; either silently changes the intended objective mixture.

For M = 0 there is no ordinary normalization; auxiliaries stay unit-weighted. An empty update has no valid loss denominator. This is a ratio estimator: a homogeneous ordinary subset, including a single ordinary example, normalizes every ordinary factor back to 1 and cancels outcome weighting. Update grouping therefore matters; this is not generally equivalent to a fixed dataset-level outcome-weighted risk.

Implementation and DDP: use unreduced CE, valid-answer masks and per-example reduction. The CE API’s class-weight argument is not sample outcome weighting. Define final ordinary coefficient = (w × M / S) / N and auxiliary coefficient = 1 / N. If DDP averages gradients, each rank backpropagates world_size × the sum of its final-coefficient-weighted example losses, summed over every accumulation microbatch. Do not add another accumulation divisor unless explicitly compensated. Obtain global M, S and N across all ranks and all microbatches for the complete update before applying its factors; test the actual reducer and trainer.

What 2:1 means: within the same update, comparable winner and loser ordinary examples have a 2:1 loss coefficient ratio. This is not a guaranteed gradient-norm ratio, aggregate class contribution or playing-strength improvement. The ratio is a selected tentative strength, not validated as optimal for training quality. The former winner-biased sampling approach is superseded.

Unchanged source-progress mixture

  • Ramp axis: global source-anchor-visit progress across both epochs, not elapsed time.
  • Linear combined WDL + FUTURE8 share: 5% → 30%, divided equally.
  • Separate board share: steady 5% from the first consumed batches.
  • Start: next 90%, WDL 2.5%, future 2.5%, board 5%. End: 65%, 15%, 15%, 5%.

The actual selected run uses 30% excluding board. These percentages are unchanged by outcome-based ordinary loss weights.

What the assembly and audit must prove

  • Each eligible ordinary anchor appears once in epoch 1 and once in epoch 2, irrespective of outcome.
  • Auxiliary task ledgers remain independent and capped at two semantic uses. FUTURE8 destination caps remain separate from ordinary destinations.
  • Record natural winner/loser/draw counts, raw-weight sums, normalized ordinary coefficient sums, auxiliary counts and total N per global update.
  • Verify data order and selection are outcome-unbiased; reject an accidental leftover winner sampler. Standard distributed-sampler tail padding can repeat examples, while drop_last omits them. Exact two-pass coverage requires supported uneven tails or masked dummy slots excluded from loss and all example/count denominators.

Accounting: consumed task examples, source visits and physical label tokens remain distinct. Source-progress ramp integration still requires an exact plan. The certified ordinary-eligible census is A = 2,263,679,463, so ordinary examples are exactly 2A = 4,527,358,926; total examples additionally include the independently allocated auxiliaries.

C Source union, eligibility, holdouts and provenanceCERTIFIED CENSUS

The frozen eligible processed-source union is certified: 26,703,686 unique games / 2,263,679,463 ordinary anchors. All 256 shards and independent audits completed. Exactly two ordinary passes imply 4,527,358,926 ordinary targets. This census supersedes the earlier unresolved-overlap status, not the remaining production-readiness gates.

  1. Inventory and containment: original Elite, engine, October 2026 and January/May/September 2026 reconcile to the certified union. The older 8,137,577 reference games are contained in Elite and add zero inventory games; containment does not establish previous training consumption. October aliases are counted once; incomplete March is excluded.
  2. Canonical identity: preserve complete normalized legal histories and root state, not occupancy alone. Certified duplicate-reference and conflicting-result group counts are both zero. Serialized-mask recount and global index/order reconciliation passed; source and provenance hashes must remain attached to the frozen build.
  3. Frozen exclusions: historical exact-history and first-30-ply exclusions passed. Additional game-family filtering relies on sealed-source receipts. All branches of a canonical game must stay in one split. This is not a claim of exhaustive transposition/position-disjointness.
  4. Source-appropriate eligibility: the observed lower-player rating bins cover 26,531,030 human games and imply both players are at least 2,300. These distribution receipts alone do not prove the precise admission-code implementation or an upper admission bound. Engine rating/clock distributions are unavailable. Keep exact-clock records and unknown sentinels distinct.
  5. Task eligibility: WDL capacity 2,263,679,463; FUTURE8 2,050,074,973; board 2,263,679,327 eligible anchors. These are capacities, not scheduled auxiliary example counts. Engine book positions (2,700,376) remain context-only. Near-terminal FUTURE8 ineligibility and overlong board skips must not remove valid ordinary targets.
  6. Independent replay: 512 games / 44,459 plies checked, with zero checkmate/result mismatches. This is sampled replay, not an exhaustive replay of every game. The source explorer contains exact inventory and carefully scoped game-level distributions.

What remains open: sustained production/cache delivery; actual consumed auxiliary counts; full-scale atomic checkpoint costs and retention behavior; logical-recovery fault coverage; production evaluation completion. Native atomic restore passed a two-update smoke and the full run has now launched. The earlier pilot remains a separate measured subset, not the full dataset.

D Exact examples, visibility, targets and lossPORTABLE CONTRACT

Let H = BOS + all actual prior whole-UCI move tokens from a complete standard-start history. Moves, including promotions, remain one token each. Targets come from recorded game data or deterministic legal replay, never from a cropped suffix.

TaskInput / queryActive targetsExample loss
Next moveH only; no query tokenOne recorded immediate next move, t+1One-token CE × ordinary outcome factor; raw 4/3, 2/3, 1; normalized to ordinary mean 1 over the global update
WDLH + WDL_QUERYOne recorded WIN / DRAW / LOSS, fixed White perspectiveOne-token CE, unit weight
FUTURE8H + FUTURE8_QUERYOne recorded move sampled from t+2…t+9; excludes t+1One-token CE, unit weight
BoardH + BOARD_QUERY, then gold cells within its own causal answer branch64 occupancy classes in a8…h8, rank 7…rank 1 order + BOARD_ENDMean CE over 65 labels, then unit example weight

Visibility must match semantics

  • Every query sees its authoritative clean prefix and itself. Earlier answer squares are visible only to later positions in that same board branch.
  • No gold result, sampled future move, board answer or query becomes ordinary move history or leaks into another objective branch.
  • Loss labels outside the intended answer are ignored. Verify exact one-position shifts, BOS, padding and branch boundaries.
  • FUTURE8 is not an eight-move rollout. Its sampled target can be illegal on the current board, so a current-board legal-move mask is inappropriate.

Normalize examples, not raw labels

  • First apply the globally normalized outcome factor only to ordinary CE; leave auxiliary per-example losses unchanged. Sum adjusted example losses and divide by the global example count N across the complete optimizer update, including all ranks and accumulation microbatches.
  • At 5% board examples, its 65 labels would be roughly 77.4% of all active labels if other tasks each have one. A naive shared token mean would change objective weights substantially.
  • Report example shares, active-label shares, normalized coefficient mass and measured weighted-loss contributions separately. Ordinary weights sum to ordinary count M, but weighted losses and gradients need not preserve their numerical proportions.
  • Full shared-vocabulary CE over 1,990 IDs is the continuity default proposal. Task-specific support masking would be a separately versioned change.

Actual integration status: the overflow-history bug is repaired and the corrected pilot benchmark is valid for its tested scope. Multi-row packed-gradient equivalence checks used tiny double-precision reference models. Actual full-model BF16 checks now cover finite execution, rank agreement, committed restore and global accumulated-gradient parity within unchanged tolerances. Source/index/cap tests and sampled independent replay passed; sustained source/cache delivery and durable run operations remain ongoing verification requirements after launch.

E Board vocabulary, architecture and contextPROPOSED CONFIG

Recorded architecture baseline

  • GPT2LMHeadModel / GPT2Config; 12 transformer blocks.
  • Hidden width 768; 12 attention heads; MLP width 3,072.
  • 512 learned absolute-position slots; recorded dropout 0.
  • Tied input/output vocabulary weights; recorded bias-free output.
  • No new transformer layers, dimensions or neural prediction head.

The owner now reports actual adapter parameter/buffer checks, exact source/target hidden-state parity and native restore/tie checks. The full BOARD1990 token export, native cold restore and two-GPU functional integration now passed. Detailed norm, activation and attention configuration values are not all enumerated in this snapshot; their omission is not a claim that integration remains prototype-only.

Append-only token meanings

  • Legacy move IDs 0–1967; White WIN 1968 / DRAW 1969 / LOSS 1970; BOS 1971; PAD 1972; WDL_QUERY 1973; FUTURE8_QUERY 1974.
  • BOARD_QUERY 1975; BOARD_END 1976.
  • P, N, B, R, Q, K, p, n, b, r, q, k, EMPTY (“.”) map to 1977–1989 respectively.
  • 15 additions → vocabulary 1,990. At tied width 768, +11,520 weights: 86,966,016 → 86,977,536 unique parameters.

Context arithmetic: with n prior plies, H contains BOS + n moves. Teacher-forced input H + QUERY + 64 cells is n + 66 tokens. END is the shifted final target, not an extra input. Thus n ≤446 at context 512. Preserve the verified shifted-layout contract in the actual run. Never crop history; skip and count an overlong board query.

Only occupancy: no turn, castling rights, en-passant state, halfmove/fullmove clocks or repetition history. These remain authoritative rules/harness metadata. Exact occupancy does not certify complete state, rule awareness or playing strength.

Actual owner receipt: a fresh complete 1,990-row chess/board table is tied and native cold-restore logits are exact. Retain tokenizer/config/provenance hashes and retain the verified full BOARD1990 export; do not mistake the earlier portable partial manifest for the actual transfer receipt. The superseded 35-token FEN proposal was never deployed and is not legacy vocabulary to preserve.

F GPT-2 initialization and controlled comparisonACTUAL OWNER RECEIPT

Actual owner receipt: the GPT-2 body transfer and trainer-native initialization now exist. Transformers 5.21.0 adapter checks covered every parameter and 12 causal buffers; actual hidden states and native cold-restore logits matched exactly. All 1,990 chess/board rows are fresh and tied, with the first 512 positions copied. The later auxiliary-history bug was separate from these initialization checks and has now been corrected and regression-tested. This is not a continuation of CM15’s trained state.

  1. Pin and verify input integrity: proposed source is openai-community/gpt2 at revision 607a30d783dfa663caf39e06633721c8d4cfcd7e. The cached revision recorded with prefix 11c5a3d… has exact configuration/weight matches to that proposed source; only the Hugging Face text-tokenizer configuration differs. Source model configuration and weight identity/hash checks passed. Retain the full source/target contracts and original artifact hashes in the actual receipt.
  2. Audit compatibility: require identical relevant math, not just matching dimensions. Validate all parameter names/shapes/dtypes, finite values, Q/K/V ordering, layer norms, activation, residual structure, attention flags and tied-head consistency. Reject unsupported differences.
  3. Copy the body: exact transformer-body tensors. Explicitly transpose HF Conv1D [in,out] to Linear [out,in] storage where the verified target requires it, including square projections. Do not infer head reordering.
  4. Copy positions: first 512 of GPT-2’s 1,024 learned position rows; no interpolation or invented positions.
  5. Fresh chess embeddings: discard all 50,257 GPT-2 text-token rows. Freshly initialize the entire 1,990-row chess/board embedding table, including existing chess token meanings, from a recorded independent seed. Tie input/output by shared storage. Preserving legacy IDs does not mean importing CM15’s learned embeddings.
  6. Controlled random arm, if run: random and pretrained arms share the exact same fresh chess/board embedding table, architecture, dropout, data order, splits, losses, optimizer/scheduler and token/update budget. Record independent body and common embedding seeds. Multiple paired seeds and a reinitialized-position ablation are proposals, not completed runs.
  7. Start fresh training state: create a new optimizer for any later authorized experiment; import no source optimizer, scheduler, scaler, RNG, global step or resume state. Save initialization separately from any existing run.
  8. Verify before use: reload hashes and real tie storage, evaluate layerwise hidden states with identical external embeddings, test the actual LittlePly forward path, packed masks, positions, dtype and native checkpoint restore. Different text/chess vocabularies make raw source text-token logit parity inappropriate.

Coverage, not a quality prediction: 85,056,000 non-embedding body parameters + 393,216 position parameters = 85,449,216 copied parameters, about 98.24% of the proposed target. This says nothing by itself about chess improvement.

Earlier portable stage, before actual framework integration: 39 NumPy reference tests passed and 1 actual-HF test was skipped. Those checks used small synthetic fixtures and did not establish native or GPU parity. The newer actual Transformers 5.21.0 and native-restore receipts above supersede the earlier environment blocker; they do not retroactively turn the NumPy tests into framework tests.
Board1990 prototype archive SHA256: 42e3e4218240282534d669b584e99605c05f8d6f57279d5b951e767069a2eccb
G Optimizer, batches, packing, validation and checkpointsACTIVE RUN

The actual full run uses 8 rows/rank with depth-8 rolling preparation, unchanged global batch/objectives/precision/clip 1, and the source-progress LR schedule below. Clipping alone is accepted as non-blocking. Retention and cadence are selected; scheduled production checkpoints/evaluations have not yet completed. Historical settings below remain labeled context.

FieldCurrent CM16 value / stateWhat is needed
Optimizer family / groupsACTIVE OPTIMIZER FAMILY
Muon for eligible hidden matrices; AdamW for the remainder.
Native functional integration verified 48 Muon / 100 AdamW states finite and synchronized. Actual training started from pristine initialization and fresh optimizer state, not disposable experiment checkpoints; retain the verified grouping.
Optimizer hyperparametersPROVISIONAL BASELINE
Historical Muon / AdamW values listed below; clipping 1.
The full run retains the selected optimizer settings and clip 1. The launch receipt explicitly confirms its source-progress LR schedule. Detailed optimizer constants below remain their recorded baseline provenance; no clipping/LR change was introduced by the execution/cache tests.
Measured benchmark LR warmupACTUAL TRIAL
180 wall-clock seconds, from 0.1× to 1× peak LR. The same window was excluded from steady timing. No GradScaler.
This is not source-progress warmup. The after-warmup optimizer LR trajectory was not extracted in this audit; the short trial did not validate the full-run trajectory.
Actual full-run LR scheduleACTIVE · SOURCE PROGRESS
First 5% warmup from 10% peak; plateau to 85%; final 15% cosine to 10%, over source progress.
The launch receipt confirms this source-progress trajectory. It is different from the earlier benchmark’s 180-second wall-clock warmup and from the auxiliary objective-share ramp. A short benchmark did not validate the entire long-run trajectory.
Global / device microbatchACTIVE RUN CHOICE
Target 8,192 ordinary anchors/update; whole-game updates may be 8,255 or 8,273. Eight physical rows/rank/forward across 2 GPUs, depth-8 rolling preparation, with accumulated global normalization.
This target is half CM15’s effective 16,384 ordinary anchors/update. The later matched comparison selected eight rows/rank + bounded prefetch; 16 rows gave only a modest gain with more memory, and 32 failed fixed numerical gates before an update. Global effective batch stays unchanged; do not equate anchors, task examples or physical rows.
Packing / attentionREPAIRED / PILOT TESTED
Overflow history corrected; multi-row regressions and tiny double-reference packed-gradient equivalence passed.
Actual full-model BF16 finite/rank/restore checks and global accumulated-gradient parity passed within unchanged tolerances. Preserve complete histories and tested branch isolation in the final run; this is not a proof for every arbitrary input or failure point.
Precision / kernels / hardwarePROVISIONAL BASELINE
BF16 autocast; FP32 weights and optimizer state; deterministic configuration. Actual trial: 2 × RTX 4090, 24 GB each; torch 2.9.0+cu128 and Transformers 5.21.0. Selected kernel details are not recorded here.
Actual trial had no GradScaler. Seal attention implementation, compilation, memory constraints, actual numerics and reproducibility; do not infer gradient scaling from BF16 alone.
Evaluation cadence / sizeREGISTERED · NOT YET COMPLETED
Every 6 hours: 100 Sunfish games + 25 against a frozen live-CM14 baseline, followed by a report. First expected around 05 Nov 04:52 UTC. No promotion.
Helper/conversion smoke passed and one evaluation scheduler is registered. Verify actual job completion, frozen baseline identity and comparable seeds/openings/colors/search budget/adjudication when reporting results. Frozen heldout/task validation and board generation remain separate checks.
Checkpoint cadence / retentionEVERY 2 HOURS
Preserve a model-only snapshot at every checkpoint. Full restart generations: initial, latest 2, latest 2 daily, final; maximum 6.
Native NCCL/atomic restore smoke passed. Next production checkpoint expected 05 Nov 00:52:28 UTC; no scheduled 2-hour checkpoint completed yet. Observe full-scale costs against the ~65 GB checkpoint/archive budget and 60 GiB free reserve. Depth-8 preparation and a monitored 2 GB cache ceiling are active; exposure ledgers are excluded from that cache ceiling but included in the whole-run budget. Durable generation-pointer acknowledgement is separate from initial publication.
Randomness / seedsSeparate body and embedding seeds required for paired comparisons. Exact run seeds unestablished.Record task sampling, source ordering, dropout/data-loader/DDP and initialization RNGs; prove deterministic resume.
Stopping condition / ETATwo sealed source epochs; no fixed eight-hour cap. Ordinary targets: 4,527,358,926. Exact auxiliary counts, steps and production runtime remain unestablished.Latest selected-profile forecast at certified A = 2,263,679,463: 132 central / 167 slow + 10% hours. This is not a direct 16% multiplier on the older estimate or a guarantee. Continuous cache supply, production changes and actual checkpoint/evaluation overhead may change runtime. The full run started 04 Nov at 22:52:23.587 UTC; sustained rate remains unmeasured.

Recorded CM15 values · source for the provisional baseline

These settings were supplied by the existing training owner as historical records. Optimizer, LR, BF16 and batch values are proposed carry-over candidates only; the other rows are historical context. None is a final CM16 setting, proof of actual current tensors or authorization to run.

Historical fieldOwner-recorded CM15 value
Muon48 hidden matrices, RMS-matched; peak LR 4.5×10⁻⁴, floor 4.5×10⁻⁵; momentum 0.95, Nesterov; 5 Newton–Schulz iterations; epsilon 10⁻⁷; weight decay 0.1.
AdamWRemaining 100 states; peak LR 3×10⁻⁴, floor 3×10⁻⁵; betas (0.9, 0.95); epsilon 10⁻⁸; matrix decay 0.1, vector decay 0.
LR schedule / clippingSource-progress schedule: first 5% linear warmup from 10% of peak; plateau until 85%; final 15% cosine to 10% of peak. Gradient clipping 1.
Numerics / attentionBF16 autocast with FP32 weights and optimizer state; deterministic configuration. Custom GPT-2 blocks using explicit-mask SDPA. Actual CUDA kernel, activation details, layer-norm epsilon and attention flags still unknown.
Batch / packing / loss8,192 source anchors per global microbatch; 2 microbatches per update across 2 GPUs. Supervised target count varies. 512-physical-token rows, no truncation; exact row counts and position tensors unavailable. Historical unit-target CE normalized by actual global supervised targets, not proof of the required CM16 board-per-example normalization.
SeedRecorded training seed 20261108. No CM16 run seed has been sealed.
CheckpointsStartup checkpoint after 10 updates; rolling roughly every 600 seconds; immutable hourly snapshots. Rank 0 broadcasts save decisions.
ValidationInitial / hourly / terminal when sufficient time remains. Ordinary panel 30,810; auxiliary panel 21,566 ordinary + 4,622 WDL + 4,622 FUTURE8. No board panel existed.
Resume / exposureModel, both optimizers, update, source cursor, ledgers, counters, per-rank RNG and journal sequence; conservative write-ahead exposure reservations. CM15’s old shared cap is not CM16’s independent-per-task contract.
Recent training rateReported medians: 32,622 examples/s at 17:00; 21,309/s at 18:00; last 19,058/s. Unequal mixtures and depletion; validation/checkpoint overhead excluded. None is a CM16 throughput benchmark or ETA.

Checkpoint state proposal: model/config/tokenizer and provenance hashes; optimizer/scheduler/scaler; source epoch and cursor; task and destination exposure ledgers; actual-consumption mixture cursor; all relevant RNG states; validation identity and metrics; software/environment revision. Production atomic model + ledger restore passed a two-update functional smoke. Retain the complete required state and test full-scale publication, bounded retention and fault handling; the smoke does not certify every crash point.

H Recovery accounting, evaluation and runtime evidenceACTUAL TESTS OPEN

Implement selected logical recovery semantics

  • Selected: logical committed-trajectory rollback/replay. Restore the committed model, optimizer, source cursor, RNG and ledgers together; replay can repeat discarded computation without adding retained training examples.
  • Not selected: strict conservative physical-attempt reservations. Counting uncertain/discarded attempts against the cap can leave holes in retained coverage. Do not claim it simultaneously guarantees exact two-pass retained coverage under arbitrary crashes.
  • Committed-checkpoint exact resume passed in the actual full model. Production NCCL/atomic model + ledger restore also passed a two-update smoke. Both are narrower than arbitrary-crash reconciliation; selected logical recovery is not silently certified for every failure point.
  • Mixture denominators count retained consumed examples, not discarded reservations or masked dummy slots. The per-update board quota rounds/guards at update boundaries; report actual shares and eligibility shortfalls rather than assuming exactly 5%.

Evaluation that answers the real question

Before launch: test generation/scoring plumbing, completion handling, branch isolation and evaluation reproducibility. Fresh untrained board rows are not expected to meet an accuracy threshold. High occupancy accuracy is a later training outcome, not a prerequisite requiring unauthorized training.

  • Teacher-forced loss plus raw free-generated whole-board exact match, per-square and occupied-square accuracy, piece/color errors and completion validity.
  • Long-history and special-move slices: castling, en passant, promotion and captures. Initial/short-prefix duplication and transposition effects must be reported separately.
  • No earlier gold squares in free generation. A history-only square-query or answer-visibility ablation can help distinguish readout from answer completion.
  • Frozen legacy move, WDL and FUTURE8 regressions, plus identical fixed-budget playing evaluations where the harness supports them. Occupancy exactness alone is not playing strength.

Corrected 30-minute-budget benchmark · complete

The corrected trial completed in 27m 14.84s with code 0. Early/middle/late rates were 11,941 / 12,137 / 12,070 task examples/s; full elapsed rate was 11,867/s. It consumed 15,116,473 ordinary examples, about 1.618 passes over the 9,343,433-anchor training pilot, not two complete pilot epochs. Exact two-epoch behavior was exercised in separate coverage/cap tests. The calculator extrapolates two source epochs conditionally.

Measured settings: two RTX 4090 24 GB GPUs; torch 2.9.0+cu128 / Transformers 5.21.0; target 8,192 ordinary anchors/update rounded to whole games; 8 physical rows/rank/forward with variable accumulation and global normalization. Every steady update clipped at norm 1; batch/row settings were neither optimized nor VRAM-forced in that earlier pilot. The later matched optimization comparison is reported separately above.

Historical holdout audit passed against 8,282 frozen games with zero checked key/history/first-30-ply overlaps. Family exclusions additionally rely on sealed-source receipts. The later frozen canonical union is certified at 26,703,686 games / 2,263,679,463 ordinary anchors. See phase shares, timing, telemetry and terminal evidence above.

The first trial remains INVALID-FOR-ETA and contributes no rate to this result. Pristine initialization is unchanged. No full run, deployment or asynchronous game monitor was launched by that pilot.

Representative throughput: measure rows/s, consumed examples/s, active labels/s, padded and unpadded tokens/s, memory, corpus replay/build cost, validation, checkpoints, startup and safe shutdown on the actual final source/length/objective mixture. Runtime is the sum of exact task counts divided by relevant measured rates, plus measured overhead; report central and conservative cases.

Latest conditional scale: two epochs require exactly 4,527,358,926 ordinary examples at the certified A. The selected-profile forecast is 132 hours central / 167 hours slow + 10%, about 5.50 / 6.96 days. The anchor calculator scales that forecast, approximately 5.83 / 7.38 hours per 100 million anchors. It is not a statistical interval, guaranteed ceiling or direct 16% discount on the earlier benchmark projection. Actual checkpoints/evaluations, continuous cache production and final integration can change runtime.

Readiness boundary: portable tests, actual initialization/restore receipts, repaired history tests, valid terminal pilot benchmark and matched optimization tests are completed evidence. The full corpus is certified. Native production NCCL and atomic model + ledger restore passed a two-update functional smoke. The local guard is active and an evaluation scheduler is registered. Sustained rolling delivery, full-scale checkpoint/retention costs, completed production evaluation still need operational receipts. Logical recovery is selected; arbitrary-crash coverage and durable recovery remain distinct scopes. Clipping alone is accepted as non-blocking. The authorized full run is now active and unpromoted; the independently audited early snapshot is dated above. The optional asynchronous board-readout game check remains disabled and separate from requested health/evaluation monitoring.

02 Representation

One square. One class.

Select a square to inspect its class and proposed token ID. Change the prefix to see how the same history feeds each learning task.

White at bottom · a8 → h8 … a1 → h1Click or use arrow keys
Only occupancy is supervised: 12 colored pieces + EMPTY. No side to move, castling rights, en-passant state, clocks or repetition count.
Squaree4
ClassP
Token ID1977
Complete authoritative inputBOS + actual prior plies
Teacher-forced answer64 squares + BOARD_END

Same prefix

Complete history is never cropped for the board branch.

Separate branch

Other tasks cannot see gold board, result or future answers.

Same decoder

Append vocabulary rows. No new neural head.

Minimal output contract

15 new tokens, not a new head

BOARD_QUERY = 1975 · BOARD_END = 1976. EMPTY is written as “.” in the portable schema. Preserve every legacy ID 0–1974. New vocabulary: 1,990. Owner receipts now cover the fresh tied 1,990-row table and native restore; the full BOARD1990 token export is verified.

64 classes + ENDmean CE1 unit-weight example
Capacity & meaning

Readout evidence, not mind reading

  • 65 active labels per board answer, including END. Average their cross-entropy first, then give the answer unit example weight.
  • At 512 context, the shifted input has n + 66 tokens and permits 446 previous plies. END is the final target, not an extra input. Verify the real loader’s shift convention; skip overlong queries, never crop history.
  • Tied width 768 adds 11,520 parameters: recorded 86,966,016 → proposed 86,977,536. Actual tying, native restore and full BOARD1990 export are verified in owner receipts.
  • A mismatch diagnoses reconstruction/readout failure. It does not prove the hidden internal state was wrong.
03 Allocation

Four tasks. Separate budgets.

Two source epochs are two source traversals, not a maximum of two total task examples at an anchor. Auxiliary tasks do not consume ordinary-move credits.

Explore the proposed schedule

Source progress, not elapsed time

WDL + FUTURE8 ramp together from 5% to 30%, divided equally. Board stays at 5%; next-move receives the remainder.

0%SOURCE PROGRESS
START / epoch 150% / epoch boundaryEND / epoch 2
90%Next move
2.5%WDL
2.5%FUTURE8
5%Board

Selected default: 5% board is additional to the combined WDL + FUTURE8 ramp. This slider is a review illustration only and never changes a run.

Synthetic global-update illustration · not training data

Same examples. Different ordinary loss coefficients.

Fixed illustrative update: 15 ordinary + 2 WDL + 2 FUTURE8 + 1 board = 20 examples. Change only the synthetic outcome composition to see the normalization. Task shares stay 75% / 10% / 10% / 5%; this is not a configured batch or a measured result.

ORDINARY COVERAGEEpoch 1: every anchor ×1Epoch 2: same anchors ×1Exactly 2A ordinary examples
No outcome-biased selection
6
6 winner6 loser3 draw5 auxiliary
Winner factor1.3333×(4/3) × M / S
Loser factor0.6667×(2/3) × M / S
Draw factor1.0000×1 × M / S
Ordinary coefficient mass15.0000Preserved at M = 15
Auxiliary coefficient mass5.0000Unchanged unit examples

M = 15 · S = 15 · N = 20 · ordinary normalization = 15 / 15

The tentative raw weights are winner 4/3, loser 2/3, draw 1. They are not claimed optimal. The displayed factors normalize ordinary weights to mean 1 over this whole synthetic update. Board’s 65 labels are averaged before its one example enters the auxiliary sum. These controls change this illustration only.

Portable checks: 24 outcome-weighting tests + 25 board tests = 49/49 pass. They cover mover/result mapping, centered weights, masks, board mean, task coefficient mass, unequal-rank/accumulation arithmetic and cancellation/ratio-estimator cases. These portable checks alone are not torch/DDP/trainer proof. New actual full-model smoke/committed-resume receipts and the repaired history tests are listed above. Actual full-model accumulated-gradient parity now has a passing numerical receipt. Arbitrary crashes and training-quality effects remain distinct scopes not established by these portable tests.

Independent semantic ledgers

Up to two uses, per learning task

Next move
12
WDL
12
FUTURE8
12
Board
12

Each canonical game + source anchor has an independent ≤2-use limit per task. Every eligible ordinary anchor appears exactly twice, once per source epoch; auxiliaries remain at most twice. One whole board answer is one board exposure, not 64 exposures.

  • FUTURE8 also caps each destination move at two uses inside FUTURE8 only. That is separate from ordinary next-move credits.
  • Use deterministic ordinary coverage with no winner oversampling. Only ordinary CE gets tentative centered raw weights winner 4/3 / loser 2/3 / draw 1, based on actual mover and recorded result.
  • Normalize those weights to mean 1 over all ordinary examples in each complete global optimizer update. This is a 2:1 comparable-example coefficient ratio, not a sample-count or gradient-norm ratio.
Synthetic evaluation demonstration

One wrong square breaks exactness

A generated board can be 63/64 correct and still fail whole-board exact match. This independent example always uses the after e4/e5 reference, regardless of the board selector above. It changes only e4 from P to EMPTY and is not a model measurement.

100%64 / 64 squares correct
PASSWhole-board exact match

Low teacher-forced loss may exploit earlier gold squares. Evaluate free-generated whole-board exactness, per-square and occupied-square accuracy, plus piece/color error slices.

Free generation starts from the clean history + BOARD_QUERY, with no gold answer prefix. Also check output length, valid classes and correct END behavior.

04 Certified frozen eligible union

The inventory now reconciles.

Certified 04 Nov at 21:43:50 UTC; independently verified at 21:48 UTC. 26,703,686 unique games / 2,263,679,463 ordinary anchors. Two source epochs require exactly 4,527,358,926 ordinary targets. This is data certification, not full-run readiness.

Disjoint certified inventory · Select a source class

Original Elite selection

June 2020 through November 2025. Contains the older 53-month reference pool.
89.0468% of ordinary anchors, not game share.

23,691,842Unique games
2,015,734,378Ordinary anchors

Engine selection

2,700,376 book positions remain context-only, not ordinary loss targets.
0.7445% of ordinary anchors, not game share.

172,656Unique games
16,853,989Ordinary anchors

October 2026

Alias paths reconciled and counted once.
2.6072% of ordinary anchors, not game share.

724,580Unique games
59,019,257Ordinary anchors

January / May / September 2026

January 728,232; May 698,300; September 688,076 games.
7.6014% of ordinary anchors, not game share.

2,114,608Unique games
172,071,839Ordinary anchors
Containment is resolved. The older 8,137,577 references are contained in original Elite and contribute zero additional inventory games. This does not prove earlier training consumption. October is counted once; incomplete March is excluded. Natural engine share is 0.7445% of ordinary anchors, not the historical 20% oversampled target share.
Eligibility, not allocation

Auxiliary capacity is not scheduled count.

WDL-eligible anchors
2,263,679,463
FUTURE8-eligible anchors
2,050,074,973
Board-eligible anchors
2,263,679,327
Engine book positions, context only
2,700,376

These are eligibility capacities in the frozen source union, not emitted auxiliary examples, labels or two-epoch quotas. The chosen schedule, rounding, collisions and independent task/destination caps determine actual allocation.

Completed audit evidence

256 of 256 shards reconciled.

Builder and audit exited with code 0. Checks include 210 fresh source-file hashes; separate engine-result sidecar verification; serialized-mask recount; global index/order reconciliation; zero duplicate-reference and conflicting-result groups.

Historical exact-history and first-30-ply exclusions passed. Additional family filtering is inherited from sealed-source receipts. Independent deterministic replay covered 512 games / 44,459 plies, with zero checkmate/result mismatches. Replay was sampled, not exhaustive.

Certified distributions: results, lengths, ratings and time controls

Denominators matter. Results and length bins below are percentages of games, not ordinary-anchor shares, history-prefix distributions or task-example shares. Rounded bins need not sum to exactly 100%. Exact length quantiles were not established.

Game results

Recorded resultAll-game share
White wins46.71%
Black wins41.77%
Draws11.52%

Game draw rates: engine 92.64%; Elite 11.27%; acquired months 8.60–8.77%. These do not prescribe outcome sampling or loss mass.

Coarse game length

Total game pliesGame share
0–4915.91%
50–9953.24%
100–14925.81%
150–1994.22%
200+0.81%

Human lower-player ratings

Lower-rated player binHuman-game share
2300–239928.40%
2400–249936.56%
2500–259923.12%
2600–26997.81%
2700+4.10%

26,531,030 human games have rating metadata; engine games are excluded. Observed bins imply both human players are at least 2,300. These receipts do not establish the exact admission-code implementation or an upper admission bound.

Time controls: keep unknowns separate

Classification uses base time plus 40 increments. Counts below have exact clocks; they are game counts, not percentages.

Human sourceBlitzRapidClassical
Elite19,031,4631,032,32636,529
January689,86137,966405
May663,78434,184332
September650,48137,315280
October687,93436,387259

Elite additionally has unknown exact-clock records: 3,339,393 known blitz-category, 239,130 known rapid-category and 13,001 unknown classification. These sentinel/unknown records are not classical. Engine clock data are unavailable.

Fresh AdamW

CM13

Move + fixed-White WDL. Initial Elite / engine sources with an 80/20 supervised-target exposure mix.

Fresh Muon / AdamW

CM14

Not initialized from CM13. Move + WDL; later human continuations used the 53-month pool and then October.

Human-only stages

CM15

Adds one sampled future move from t+2…t+9. Initially no winner bias; later continuation favored winner proposals. Stopped 04 Nov at 18:11:22 UTC. No board reconstruction.

Current unified experiment

CM16

All eligible available sources, exact two-pass ordinary coverage, normalized outcome-weighted ordinary loss and fixed64 board readout. The earlier winner-sampling proposal is superseded. Broader than an isolated board-only ablation.

Do not inherit the old 80/20 human/engine target mix: it relied on heavy engine repetition. Any CM16 source weighting must fit the new per-task caps, preserve full coverage and be disclosed.

05 Verification

What is verified, and what comes next.

The corrected pilot benchmark, repaired history path and native initialization have receipts. Pilot holdout checks passed against 8,282 frozen games. The full frozen union/count is now certified. The production NCCL/atomic restore smoke passed. The full run is active with logical rollback, clip 1 and depth-8 rolling preparation. Verify sustained delivery, first checkpoint/evaluation, retention behavior, and recovery in operation. No automatic promotion or live-route change.

Portable verification completed: 49/49 tests pass. These comprise 25 board-contract checks and 24 outcome-weighting checks covering serialization, shifted layout, output scoring, exposure accounting, masks and global-update loss arithmetic. They are not substitutes for actual corpus/model validation. Separate actual receipts, repaired history regressions and remaining validation limits are reflected in the gates below.

Source inventory & global dedup CERTIFIED

The frozen eligible processed-source union is certified: 26,703,686 unique games, 2,263,679,463 ordinary anchors and 4,527,358,926 ordinary targets across two epochs. All 256 shards and independent audits completed. The older 8,137,577 references are contained in original Elite and add no games; this does not establish previous training consumption. October aliases are counted once; incomplete March is excluded. Historical exact-history/first-30-ply exclusions passed; additional family filtering relies on sealed sources. Independent replay sampled 512 games / 44,459 plies, not every game. Production readiness remains separate.

Two-epoch task allocation PARTIAL

Separate exact two-epoch coverage/cap tests passed. The timed pilot consumed about 1.618 ordinary passes, not two complete epochs. Cursor reconciliation passed; FUTURE8 destination maximum was 2 and WDL/FUTURE8 quota skips were 0. The full-union counts are certified and the selected run retains the exact two-epoch ordinary plan. Reconcile actual consumed per-task counts and separate FUTURE8 destination credits throughout the run. Preserve exactly one ordinary use per epoch for every eligible anchor, including losers, with no outcome-biased selection and independent auxiliary ≤2-use caps. Report natural outcome counts, collisions and consumed examples, plus correct ordinary weight totals.

Replay, context & branch isolation REPAIRED / TESTED

The overflow-history bug was reproduced, fixed and covered by multi-row regressions. Packed-gradient equivalence used tiny double-precision reference models, including overflow. Actual full-model BF16 finite/rank/committed-resume checks and global accumulated-gradient parity passed within unchanged tolerances. Preserve the full-history contract, 446-ply board limit and correct overlong skip behavior.

Tokenizer & native initialization OWNER PASS

The owner reports every parameter and 12 causal buffers checked, exact source/target hidden states, fresh tied 1,990-token table, first 512 positions copied and exact native cold-restore logits. Pristine initialization remains separate from disposable benchmark weights. Pristine hash remained unchanged after the corrected disposable benchmark. Initialization and committed restore checks do not prove playing strength or arbitrary-crash behavior.

Loss, quota & resume integrity PARTIAL

Actual two-GPU BF16 finite/rank/committed-checkpoint resume checks and terminal reconciliation passed. Production NCCL/atomic model + ledger restore passed a two-update functional smoke. Logical committed-trajectory rollback/replay is selected; full-scale state costs and arbitrary-crash fault coverage remain open. Average 65 board labels into one unit-weight example. Apply outcome weighting only to ordinary CE, normalize over all ordinary examples in the complete global optimizer update, then divide adjusted example losses by global N. Verify masks, accumulation and DDP gradient scaling. Keep the chosen crash/exposure semantics separate from retained-consumption mixture accounting; do not silently adopt conservative physical-attempt reservations as proof of exact retained coverage. Test distributed assignment, fault/restart behavior and steady 5% board allocation from the beginning.

Free-generation & legacy regression OPEN

Compare teacher-forced and free-generated results. Measure whole-board exactness, per-square and occupied-square accuracy, piece/color errors, valid termination and long-history/special-move slices. Before launch, verify this evaluation pipeline with smoke tests; do not require high accuracy from untrained board rows. Later measure ordinary move, WDL and FUTURE8 regressions during an authorized experiment. If isolating history readout from answer completion, add a history-only square-query or earlier-answer visibility ablation.

Representative runtime measurement TESTS COMPLETE

The corrected 27m 14.84s pilot and subsequent matched quick comparison are complete. Selected 8-row cache + prefetch achieved 9,528 ordinary / 12,513 task examples per second, 16.0% above its matched uncached control. Six warmup + 30 measured updates per passing case do not establish week-scale speed. Latest conditional full two-epoch forecast: 132 central / 167 slow + 10% hours. Verify rolling producer capacity and account for full-scale checkpoints/evaluations. MFU remains unmeasured; busy-time GPU utilization is different.

In training. Unpromoted. No eight-hour ceiling.
Started 04 Nov at 22:52:23.587 UTC from pristine initialization. The retained trajectory targets exactly two ordinary epochs with independent task caps. Promotion remains the user’s decision.

132 / 167 hours remains conditional.
Every 2 hours: planned checkpoint and retained model snapshot. Every 6 hours: planned 100 Sunfish + 25 frozen-live-CM14 games and report. No scheduled 2-hour checkpoint or full evaluation completed yet at this snapshot.

Immutable launch record

The verified launch receipt records 67 updates and 552,336 ordinary targets at 2026-11-04 22:53:12 UTC. The later observations above are separately dated.

Download the launch receipt · Download the adaptation recipe and historical evidence

Launch receipt SHA-256 3fb4127dd36f0a8f03972fbbe0641de45f85cbce56a0799ca67a97509238230f