The reveal the search leaves on the table

Kimi K3's design (kimi-k3-theory-design.md) located what is structurally missing rather than mis-weighted. The cheapest reveal in the game is the same-wave double hit — two numbered neighbours of a solid gray popping in one cascade — and no term expresses it: solid_exposure is a weighted sum of the two neighbours' marginal readiness, which cannot tell two discs that are each 80% ready from a pair that is jointly reachable in one wave. The proposal: an aligned_double_hit term, the product of the two neighbours' readiness over support-disjoint pairs, at +300; a second release step restricted to chains that terminate on a cover; the owner's ledger as a soft entombment penalty for values 3–7; and a danger gate that switches the construction terms off above height four — the "no emergency" clause as a product no linear leaf can express. Kimi's own prior: about one chance in seven that the primary arm passes, and the ledger clause demoted to a declared-diagnostic fallback because the entombed count already reads −0.023 against remaining life.

Asked to think past this screen, Kimi K3 wrote a conditional portfolio (kimi-k3-portfolio.md). Two points stand out. First, same-seed pairing is already exact common-random-numbers in this engine — the disc tape and each move's reveal stream are pure functions of (seed, move) — so the 200k–370k paired spread is divergence amplification, and no replay scheme can reduce it; only a model-based lifetime estimator, validated seed-free on the retained arms of the shared cohort, can lower the 50k floor. Second, the 1.54-million clairvoyant ceiling was measured against a persistent hidden board, but the native engine draws a gray's value at reveal time, so that planner knew things no legal policy could know even in principle. A stream-clairvoyant planner over the engine's actual future — about 25 CPU-hours — would bound every legal policy per tape; if it sustains under about 2.25 clears per move, the million-point mean is very likely out of reach under these semantics, and saying so with a measured bound becomes the program's most valuable result. Both are owner-level direction calls and were not started.

The D3+NNUE program (kimi-k3-d3-nnue-program.md) puts its mass on the exact control T5a — collapse the leaf-ply disc loop closed-form, D4-s7-shaped strength at 0.82x D4 s5 work, prior 0.50 — and stages everything to be killable offline. The direct-policy program (kimi-k3-direct-policy-program.md) settles the recurrence question: the public state is Markov, a cross-move hidden state is illegal as the protocol is written, and what memory would buy is amortised computation, which per-decision attention can supply legally; its recommended first rung is a sibling-set transformer scored offline. The unexplored-space sweep (kimi-k3-unexplored.md) cut ten of thirteen candidate directions as already tried and kept three, led by a half-hour board-clear accounting measurement. The evening's two negatives sharpen all three documents: the set-ranker probe is now the only untried route to a D4-ordering student.

Stage G0 of the preregistered sibling-outcome-label experiment (EX-20260823-sol-corpus-and-offline-gate-4d3d86e4) built the machinery in full: the corpus generator gained a --panel2 mode writing 992-byte PanelRecordV2 rows — all seven siblings per root, K CRN continuation outcomes per sibling, tapes derived from canonical public-root hashes under dedicated domains and never from the seed — plus a reader, a legality checker, and the Kendall-τ ladder analysis with the preregistered cluster bootstrap. Every pre-seed gate passed on the already-open smoke range (runs/RUN-20260823T215500Z-sol/g0/gates.log): v1 output byte-identical to a pristine binary built from git HEAD, panel2 output byte-identical across repeated runs and thread counts, mirror invariance exact by construction (continuations run in the canonical orientation), and labels present only for legal columns.

Then the seed-free feasibility check ended the stage. The design's D3 N7M6 budget lines rest on 292 moves/s — the throughput of the single-knob depth-3 search from RUN-A51D — but the frozen fidelity reference is the C0 factored engine (N=7, M=6, work 51,084,852), which the retained C0 artifact and a fresh smoke pilot both price at about 1.75–6.83 CPU-s per decision. The 64-root mini-ladder alone projects to 62.9–244.8 CPU-h against its 3.2 CPU-h line, and every D3 N7M6 stage lands at roughly 20–77× its allocation, so stop rule S1 halted G0 with the training lease SL-20260823T215000Z-a5216000 still unopened — no leased seed was read, and no ladder τ exists. The partial is recorded in runs/RUN-20260823T215500Z-sol/g0/ladder.json and research/runs/RUN-20260823T215500Z-fc74e0b4.json; the pipeline demo on 16 smoke roots (D1 vs D2, diagnostic only) is in smoke-ladder-demo.json. The decision now owed by the coordinator: a P-SOL-v2 with honest throughput arithmetic — a cheaper fidelity reference, shorter or fewer reference continuations, or a fast-engine factored search — or a scale-out request, since the ladder as frozen is a multi-hundred-CPU-hour object, not a 14-CPU-hour one.

C0 made D3 N7M6 the most interesting policy in the program; tonight's guardrail kill made its cost the binding constraint on the whole learned-label programme. Porting M=6 reveal sampling into the fast memo engine is semantics-preserving engineering (days, plus about 5 CPU-h of preregistered equivalence gates, zero leased seeds) and would turn the 6-root guardrail into a powered test, re-open the P-SOL corpus at ~40x lower cost, and drop 100-game D3 N7M6 studies to about half a CPU-hour. The sibling-outcome theory itself remains untested, not refuted - it is blocked on this port, and the training lease is still reserved and unopened.