The newer q4096/deferred actor completed all 2,000 external games with independent replay. Against unrestricted Stockfish 19 it scored 56 wins, two draws and 608 losses, a relative difference of -411.50 Elo with a 95% interval of [-460.14,-370.68] at 100 milliseconds, one thread and 16 MiB. This is not an absolute or human rating. Its separate neural comparison recorded all 4,000 attempts, but three games remained unfinished at the frozen 512-ply ceiling. Independent full-sample acceptance failed. The 3,997 completed games gave a diagnostic -63.86 relative difference against classical, interval [-73.80,-54.07]; it is not a certified full neural result. Neither failure is repaired by the passed full data/learning audit or performance measurements.
The full original neural comparison completed all 4000 games with 1274 wins, 746 draws and 1980 losses. Independent replay measured -61.97 Elo relative to the unchanged classical evaluator, with a 95% opening-cluster interval of [-71.25,-52.87]. There were no crashes, illegal moves or time/protocol forfeits. The required +50 gain failed. Validation-loss improvement therefore did not establish stronger equal-time chess, even on the full corpus. The experiment does not identify evaluation quality, quantization or execution cost as the sole cause. Classical remains the default; the q4096 encoding has not inherited this measurement or demonstrated a positive gain.
Correctness, throughput, validation loss and match strength are separate measurements.
The legal generator agreed with the independent mailbox and Python oracle across the recorded reachable-position campaigns. Exact restoration and accumulator refresh checks exercised captures, promotions, king buckets, en passant and Chess960 overlap. Those results remain tied to their actual executables; final-source acceptance must rerun or explicitly establish an unchanged relevant input boundary.
The original coarse q64 network export produced excessive float/integer errors. The q1024 v2 contract reduced the measured 100000-probe maximum to 11.3652 cp and mean to 1.5691 cp. Integer native/scalar/NumPy outputs agreed exactly. This supports the numerical implementation on those probes, not playing strength or a universally established float error bound.
Three larger diagnostic seeds completed 20000 updates each. The selected seed was 17 at saved step 750, with validation loss 0.580516 versus untrained 0.695267. Later updates worsened held-out loss despite lower training loss. Selection retained the early checkpoint instead of the final checkpoint. The test arrays were not opened. This dataset predates functional-input reflection screening and remains diagnostic.
The selected larger diagnostic network scored 4 wins, 8 draws and 84 losses against the unchanged classical evaluator at 2000 nodes per move. The estimated difference was -416.56 Elo with 95% interval [-infinity, -254.73], using 12 independent opening clusters. That is a failed neural-contribution result. Increasing data and decreasing validation loss did not establish an improvement in chess decisions.
A residual evaluator retained the unchanged classical terms and trained three independent 20000-update seeds on the same older diagnostic data. Validation selected seed 17 at step 1000, loss 0.587335 versus the untrained classical-backed value 0.793343. Its fixed 96-game equal-node pilot scored 8W/4D/84L, estimated -373.80 Elo with interval [-infinity,-210.72]. All games completed without protocol/time failures, and separate replay reproduced the result. Residual initialization alone did not fix the learned decision problem. Raw games, settings and selection preserve this negative result. The initial pilot configuration enabled neural evaluation before loading its model and was rejected before any game; the corrected run used a new identity.
A later fixed ten-percent residual mixture had 40W/20D/36L at equal 2000-node budgets, +14.48 Elo with interval [-88.74,104.37]. At matched 20 ms per move it had 9W/17D/70L, -260.73 with interval [-364.31,-181.33]. Both retained all 96 games with no protocol/time failures and used the same executable and unchanged classical terms. The equal-node estimate is inconclusive; the equal-time sample is negative. Evaluator overhead remains a central limitation. Both fixed samples and raw games retain this distinction.
Warm short external diagnostics scored 42W/3D/3L against Stockfish 19 with its 1320 strength setting and 0W/0D/48L against unrestricted Stockfish 19. Both used one thread, 16 MiB, 20 ms per move, a one-node warmup and color-reversed openings. The former gives +394.11 relative Elo with interval [292.96, 544.69]. It does not establish a 1714 absolute engine rating. The unrestricted sample has no finite maximum-likelihood difference; its guarded upper 95% limit is -76.21. Forty-eight losses do not demonstrate proximity to Stockfish.
Cold-start matches contained opponent time forfeits. They remain recorded but cannot be sold as on-board strength gains. Warmup was introduced in separate run identities; the earlier games were not edited or discarded. Degenerate bootstrap samples initially yielded misleading zero-width uncertainty, so statistics version 2 adds a conservative bounded-score guard. Old summaries remain available.
The present conclusion is narrow: Caissa runs as an original engine and diagnostic laboratory, while its learned-strength objective has not passed. The classical default preserves the stronger observed evaluator.