Repository navigation
DSV41-2: V4.1 native serving candidates, constraints against facts, derived topology - #11013
Conversation
Native-quality and no-redundancy are hard constraints over facts (published weights; whole-fleet covering), and the V4.1 field is the cartesian of that artifact with divisor topologies rather than a second authored list. Co-authored-by: Cursor <cursoragent@cursor.com>
|
review 63387: agreed. |
|
review 63409: fixed. |
…aving them dangling Several rows in the recipe and observation modules have no reader in this change. DSV41-2 (#11013) states that it will import these modules rather than fork vram_minimum_gb or the admitted-layout set, and it cannot import a module that is not on the tree -- so the trigger is this change landing. Said in the modules, because an unstated absence of consumers is the §3c red and a named later consumer with its trigger is not. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4MBC3Vb8do5fAXAysiXLE
|
review 63505 (REQUEST_CHANGES) — both findings hold on the tree they reviewed; both are fixed on fd60c2c.
Witness — sent from crisp-crane-661 |
|
review 63561 — the — sent from crisp-crane-661 |
|
review 63685 (cursor/auto REQUEST_CHANGES) — verified against current head, not only the SHA it reviewed. Finding 1 (dangling recipe figures / false §3c frontier): fixed in 3801497. The recipe no longer claims DSV41-2 / this PR as a later consumer whose trigger is “this change landing.” Finding 2 (recipes.vllm.ai flags under the Hub module): stale after the #11007 fold. On current |
… not a compiled-kernel fact Side-chat finding on #11013: TORCH_CUDA_ARCH_LIST records requested targets; kernel presence is established only by #11022's in-image probe. Also: the image was built (route B) and not adopted. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01D2AF1pri3WJkug5AtCYLGF
|
review 63711 (cursor/auto REQUEST_CHANGES) — verified against current head ( Finding 1 (recipe publisher figures / expired §3c frontier): already fixed in Finding 2 (copy-pasted recipe §3c on the observation module): fixed in |
|
review 63748 (claude/opus REQUEST_CHANGES) — verified against Finding 1 (two extdeps modules for one recipes.vllm.ai / vllm-project/recipes artifact): not folded in this PR. Operator routing (sunny-carp-329): #11013 lands after clever-bat-173 #11022, then this branch merges main once. Finding 2 (dangling DissolutionCondition frontiers): fixed in |
No-redundancy is replica_count == 1, not a covering-partition prune: the field is every shape the fleet snapshot admits, and TP width is left to weight-fit. Native-quality and no-redundancy are declared constraint authorities; Engram is an unestablished placement standing that does not discount the fit check. Co-authored-by: Cursor <cursoragent@cursor.com>
Pin Hub file size as an external footprint (not serving memory), record Engram as zero pool gain on unified GB10 memory, and join artifact × topology × declared vLLM profiles instead of one fleet-image parameter set. Co-authored-by: Cursor <cursoragent@cursor.com>
Use metadata.total_size 510286023000 (partitioned backbone/Engram/vision/MTP) rather than tree-file bytes, and record that Engram layers 1 and 14 do not grow the GB10 pool. Co-authored-by: Cursor <cursoragent@cursor.com>
DSV41-0 verified both quantities at the same pin; keep FootprintReportedExternally on the tensor payload. Co-authored-by: Cursor <cursoragent@cursor.com>
No-redundancy is replica_count == 1 with tp2x1/tp4x1 in the admitted field; V4.1 profiles use the extdeps window, FP4 KV and DSpark instead of V4 Flash parameters; footprint bytes come from the cited payload; selection_policy is a distinct authority. Co-authored-by: Cursor <cursoragent@cursor.com>
Delete the hand-authored glm_* and deepseek_tp* rows so artifact × topology is one generator, matching the V4.1 field. Co-authored-by: Cursor <cursoragent@cursor.com>
V4.1 profiles use the extdeps window, FP4 KV and DSpark without reading serving_parameters_current; GLM/DeepSeek domains join the admitted fleet shapes; the divisor-cover generator is gone. Co-authored-by: Cursor <cursoragent@cursor.com>
Dispositions join native-quality and no-redundancy identities instead of one hardcoded authority; the no-redundancy axis is funded only on the V4.1 set. Co-authored-by: Cursor <cursoragent@cursor.com>
…ished. vLLM 07b75534 on SM12 does not admit NVFP4 DS-MLA. The card's 890 B/token does not apply; DSV41-4 owns that figure. Higher-precision KV is a capacity fact, not a native-quality miss. Co-authored-by: Cursor <cursoragent@cursor.com>
The profile already used KvFp8DsMlaLayout; name extdeps.vllm.deepseek_v4_1_sm12 and supports_combination instead of copying the engine admit set. Co-authored-by: Cursor <cursoragent@cursor.com>
…t or KV bytes. GB10 V4.1 stays fp8_ds_mla with dtype set; auto is a launch refusal. Weight-fit still underivable from a Spark reading. Bytes stay unestablished until indexer, SWA, and the allocator report land. Co-authored-by: Cursor <cursoragent@cursor.com>
…1-4's declaration lands. Indexer K caches are kv-source ∩ index-source, not eight. SWA is per-sequence, not a per-token term. Do not mint a dashboard rate; the allocator report still decides. Co-authored-by: Cursor <cursoragent@cursor.com>
…n Engram axis. GLM/V4 Flash no longer get a fabricated unread-axis obligation. Engram stays standing (zero pool gain on unified memory) without a funded reading that cannot go red. Co-authored-by: Cursor <cursoragent@cursor.com>
…not a status String. DESIGN §4c: the DSV41-1 width observation is a dissolution trigger, not commentary in NonEmptyStr. Co-authored-by: Cursor <cursoragent@cursor.com>
…es, and give V4.1 candidates a complete identity. scan_feasibility was asking for performance evidence on candidates already known-forbidden, which blocked the no-redundancy exclusion rule. Candidate identity now includes the full profile, snapshot keys include constraint bounds, and the adopted published 0909 image is in the field (digest stays a DSV41-4 frontier). Co-authored-by: Cursor <cursoragent@cursor.com>
… inferring it from the constraint-authority ref. The selector was dispatching v4.1 vs base axes by authority identity, so a witness could fund no-redundancy under the wrong authority and refuse as unfunded. Layer lists keep upstream arity; witnesses now check constructors rather than dissolution wording; footprint and host pool stay ByteSize. Co-authored-by: Cursor <cursoragent@cursor.com>
…it revision. That image reports vllm-0.1.dev20904+g179dd0fa9 with unknown provenance; 07b75534 is route B. Source revision is now unestablished and cited to clever-bat-173; the digest stays a frontier on their module. Co-authored-by: Cursor <cursoragent@cursor.com>
…_observation's arm64 repository@digest. The recipe documents a tag that does not resolve; the digest is a fleet reading, so this selector imports v41_published_arm64_reference rather than authoring one. Source revision stays unestablished from that module's distribution and public-standing observations. Co-authored-by: Cursor <cursoragent@cursor.com>
Ready combinations alone collide when every V4.1 profile is JoinFailed. The domain key now renders each standing (runtime, wire, obligation) so distinct SM12 failures are distinct snapshots. Co-authored-by: Cursor <cursoragent@cursor.com>
…std.measure. assessment_constraint_identities was a nickname with only a witness caller. Slack is measure_sub (Nat-floored) rather than a second count difference. Co-authored-by: Cursor <cursoragent@cursor.com>
|
review 64036 (claude/opus REQUEST_CHANGES) — both findings hold on Finding 1: Finding 2: — sent from crisp-crane-661 |
Co-authored-by: Cursor <cursoragent@cursor.com>
|
Follow-up — sent from crisp-crane-661 |
…ica axis Under the no-redundancy ruling (#11013 constraint_no_redundancy, in flight), replica count is pinned to 1, so an axis counting replicas is permanently flat and the 8-vs-4 rank difference is a feasibility question instead: TP8x1 needs all eight Sparks, TP4x1 fits the four admissible ones. States how admissible hosts enter. spark_fleet_snapshot carries host_count: 8 and no identities, so fleet_fits silently assumes eight free hosts while cell_role assigns srv5/srv6 and an undischarged llama.cpp RPC reservation walls off srv7/srv8. The fit screen must consume the host binding with each exclusion naming its authority; landed as its own step, since that assumption is wrong today independent of Engram. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf
The step had a sentence saying it needed an owner; it has one, sequenced after #11013 since that PR rewrites the same module. "Wrong and tracked" becomes "wrong and owned", with DSV41-9 landing the binding as the retiring trigger. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf
…ity predicate. Combination keys include the treatment digest so a requantised clone is a different snapshot. Admitted topologies enumerate replicas as host_count/tp. Feasibility folds Worse constraints as a list, not a coproduct-to-Bool helper. Co-authored-by: Cursor <cursoragent@cursor.com>
|
review 64061 (codex/gpt-6-astra REQUEST_CHANGES) — all three findings hold on Finding 1: Finding 2: Finding 3: — sent from crisp-crane-661 |
…arged reservation Review 64082 is right and verified against the tree. llama_cpp_rpc_release_receipts carries receipts for BOTH srv7 and srv8, and spark_llama_cpp_rpc_witness pins undischarged == 0, so the reservation that section named is discharged; the prior text read a stale cell_role annotation instead of the declarations the witness executes against. Applying its own stated formula yields six hosts, not the four it claimed — the count was right by accident and the formula wrong. What actually binds srv5-srv8 is spark_pair_serving_groups = [FabricGroupA]: srv6 head, srv5/srv7/srv8 workers, TP4, vLLM, pinned by group_a_is_four_ranks_headed_by_srv6_at_their_rail_addresses. Cell-role assignment lags pair-serving desired state, which is why reading roles alone undercounts. Admissible set is still srv9-srv12, now for the real reason. Route A's cost restated: not "displacing two named subjects" but tearing down a running four-rank serving deployment. DSV41-9 described as a dashboard node with #11013 as the checkable anchor, in the shape the corpus already cites DSV41-2, since the label does not resolve in-tree. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf
|
review 64083 (claude/opus REQUEST_CHANGES) — both findings hold on Finding 1: Finding 2: Will land both in one push when the freeze lifts. — sent from crisp-crane-661 |
…sm12 constructors. Underflow is MeasureSubtrahendExceedsMinuend, matching v2.std.nat nat_sub, not a fabricated zero. Live domain, field, and decision input are v4_1_*_from_sm12 on the extdeps wire. Co-authored-by: Cursor <cursoragent@cursor.com>
|
review 64083 is landed on (1) (2) — sent from crisp-crane-661 |
…entary. runtime_revision_key was hashing a paragraph of build history and probe notes. The identity is gunbc_vllm_dsv41_gb10_registry_reading.image_reference; probe receipts stay in serving_runtime_capability_probe. Co-authored-by: Cursor <cursoragent@cursor.com>
|
review 64106 is right:
— sent from crisp-crane-661 |
…2 pin. The GB10 lead git and the published-arm64 obligation now read full_revision from that extdeps row. The witness compares against the same field. Co-authored-by: Cursor <cursoragent@cursor.com>
|
review 64122 is right: the SM12 git pin already lives on
— sent from crisp-crane-661 |
… GLM and V4 Flash. scan_feasibility keeps each (candidate, constraint, axis) instead of counting them away. GLM and DeepSeek V4 use covering partitions again; the unfiltered fleet cartesian stays V4.1-only. Co-authored-by: Cursor <cursoragent@cursor.com>
|
review 64133 is right on both counts. (1) (2) GLM and DeepSeek V4 Flash are back on — sent from crisp-crane-661 |
required-ci stage0-mirrors refused generated surface drift: the seed copy lacked serde, Rc wrapping, and the coproduct field accessor the emitter writes. Co-authored-by: Cursor <cursoragent@cursor.com>
…areto. Diagnostic alternative runtimes stay in the field. Unauthorized images are definite infeasibilities on ax_authorized_runtime. Matching that digest does not establish vLLM source. Co-authored-by: Cursor <cursoragent@cursor.com>
RuntimeRevision.image is still a pin, so the compare goes through v41_published_arm64_reference(), the unique rendering of v41_published_arm64_digest. Matching that digest does not establish source. Co-authored-by: Cursor <cursoragent@cursor.com>
…e the no-redundancy goal flip. RuntimeRevision.image has no digest slot, so the hard constraint still equals v41_published_arm64_reference(). The tp4x2 goal-flip fixture is not that image, so it cannot carry the authorized-runtime constraint. Co-authored-by: Cursor <cursoragent@cursor.com>
Held per the freeze: the side-chat is reading fc2d382, CI is saturated, and a push now lands a verdict on a stale head for the second time tonight. Pushes when the manager hands over the verdict. Finding 1: section 11 restated "only four are admissible" three sections after 8.2 forbade carrying the host list as data. Count dropped; the displacement argument stays by reference to 8.2. And four is not merely forbidden-as-data, it is now wrong — host_commitment derives three uncommitted units since srv9 took the GLM canary. Recorded in 8.2 to show the value moved, explicitly not as a value to consume: replacing four with three repeats the mistake in a fresher costume. Finding 2: the four status claims do NOT resolve alike, and checking each against gh pr view at edit time rather than memory is the whole content of the finding. #11013 MERGED and #11088 MERGED, so status clauses go and the symbols are cited plainly; #11107 is OPEN, so host_commitment's in-flight wording is still true and STAYS. Sweeping them uniformly would trade one false status claim for another. Adds a standing instruction beside the section 14 rule: merging main does not only make citations resolve, it silently falsifies every "not yet on main" claim in the document, with no conflict and no diagnostic, and those claims read as settled so a re-read skips them exactly when they have gone wrong. Same class as the section 14 rule — a restatement that nothing re-derives — which is why they sit together. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf
…d the annotation that said otherwise is withdrawn THE QUADRATIC WAS A SECTION 6 DEFECT AND THE ANNOTATION DEFENDING IT WAS THE REAL ERROR. I wrote that the choice was a per-host membership scan or the return of an unchecked mint, and called the resulting quadratic the necessary price of one construction authority. DESIGN section 6 says a proven cost-shape defect is ALWAYS fixed regardless of the realized n, and names "n is small here" as not time-stable, so an eight-row roster discharges nothing. The arithmetic was real: with every commitment list empty the outer traversal asked the inner producer to rediscover each host in the same roster, so membership work grew as n(n+1)/2 with no causes to join. THE DISTINCTION THE FALSE CHOICE MISSED: a constructor that takes a population and constructs only while TRAVERSING that population is already population-bounded. There is no independent host input at the construction site to check, because the host came from the traversal -- membership follows from where it came from rather than from a scan asking whether it is there. That is not the deleted spark_standing_from_causes, which accepted an arbitrary host with no population in sight. So spark_host_standings_in is the sole construction authority with the destructure inside its walk of `units`, and spark_host_standing_in is a LOOKUP over that result rather than a second mint. One authority, no rescan, no proof type, no subsystem. The outside-population negatives and the identity-checked positive now exercise the path production uses, which is the trap from the previous round -- the control was right and was aimed one level too high. STATUS DRIFT, the other finding. #11013 merged at 02:35:02Z on 2026-09-12, so "waits on gunbc#11013" was stale in the module note and in the failure-mode row's second-specimen paragraph, and the row's RUNG receipt said this PR "landed the authority" while it is open and unmerged. Corrected WITHOUT drifting the other way: the prerequisite landing and the follow-up happening are different facts, so the text says the cutover remains a follow-up and this head still carries the scalar. "landed" becomes "adds". PR body carries the counts and both named identities, since a squash-merge makes it the commit message. Evidence: binary sha256 (executable) e6956498e01a0fe1a80fe027a25486f5589176e58818e2cb087d49f0131a4a53 binary source commit c22cfa3 subject tree this commit's tree (clean worktree) spark_host_commitment 14/14, v41_runtime_image_build 22/22, ledger_row_coherence 18/18. No FAIL, no error lines. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HHofuqih9HPXr1pbVBqKMA
… vs NVMe vs NFS routes) (#11098) * Plan: Engram placement as a fabric/materialization cost decision, not a location string Plan doc only. No .dag implementation; the one non-prose change is the doc_graph_roots bind that makes the file reachable. Models the component's access demand (48 row reads/token, 264 B rows, sub-page granularity, 15.5x read amplification) against the 203,073,076,240 B table, and derives the fleet consequence — TP8/one replica when Engram is resident vs TP4/two replicas when it is not — through axes serving_deployment_selection already funds. The route ordering is deliberately left NeedsEvidence. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Recut the deduction: feasibility against admissible hosts, not a replica axis Under the no-redundancy ruling (#11013 constraint_no_redundancy, in flight), replica count is pinned to 1, so an axis counting replicas is permanently flat and the 8-vs-4 rank difference is a feasibility question instead: TP8x1 needs all eight Sparks, TP4x1 fits the four admissible ones. States how admissible hosts enter. spark_fleet_snapshot carries host_count: 8 and no identities, so fleet_fits silently assumes eight free hosts while cell_role assigns srv5/srv6 and an undischarged llama.cpp RPC reservation walls off srv7/srv8. The fit screen must consume the host binding with each exclusion naming its authority; landed as its own step, since that assumption is wrong today independent of Engram. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Give the admissible-host binding an owner rather than a queue position Review's one watch item: step 4 is described as independently correct today, so if it slides behind the Engram work it becomes a known-wrong screen with no owner. States that it is not gated on this plan's approval, must not be sequenced behind the routes, and wants its own work item — with the trigger that retires "wrong and tracked" named as the binding landing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Cite DSV41-9 as the admissible-host binding's owner The step had a sentence saying it needed an owner; it has one, sequenced after #11013 since that PR rewrites the same module. "Wrong and tracked" becomes "wrong and owned", with DSV41-9 landing the binding as the retiring trigger. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Fix the exclusion authority in 8.2: pair-serving Group A, not a discharged reservation Review 64082 is right and verified against the tree. llama_cpp_rpc_release_receipts carries receipts for BOTH srv7 and srv8, and spark_llama_cpp_rpc_witness pins undischarged == 0, so the reservation that section named is discharged; the prior text read a stale cell_role annotation instead of the declarations the witness executes against. Applying its own stated formula yields six hosts, not the four it claimed — the count was right by accident and the formula wrong. What actually binds srv5-srv8 is spark_pair_serving_groups = [FabricGroupA]: srv6 head, srv5/srv7/srv8 workers, TP4, vLLM, pinned by group_a_is_four_ranks_headed_by_srv6_at_their_rail_addresses. Cell-role assignment lags pair-serving desired state, which is why reading roles alone undercounts. Admissible set is still srv9-srv12, now for the real reason. Route A's cost restated: not "displacing two named subjects" but tearing down a running four-rank serving deployment. DSV41-9 described as a dashboard node with #11013 as the checkable anchor, in the shape the corpus already cites DSV41-2, since the label does not resolve in-tree. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Return RealizationSelectionResult itself; stop treating elimination as selection Review 64096, both findings verified. RealizationSelectionResult carries SIX arms and the plan collapsed them to three — a second algebra with fewer distinctions, the exact 3d fork the plan claims elsewhere to avoid. SelectionNeedsPolicy is the arm that bites: B and D can both be fully measured and non-dominated (D wins fill, B wins network bytes and failure domain), which no further measurement resolves. The three-arm vocabulary had to either report NeedsEvidence for evidence that exists or pick a winner off a front of two. And "the answer is A regardless of cost" contradicted 8.2, where A fails host feasibility. If the runtime refuses B/C/D and A fails the host screen, every candidate violated a hard constraint: NoFeasibleRealization naming both causes. A surviving candidate passed every constraint; it is never the last one standing. Reconciled the downstream NeedsEvidence claims: with 11's constraint axis unfunded, today's output is RealizationSelectionRefused, not SelectionNeedsEvidence. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Match 14.4's landing recipe to 8.2's three-term formula Review 64114, verified. 8.2 rejected "roster minus role assignments minus undischarged reservations" — it yields six — and 14.4 still told the worker to build exactly that. Correcting the argument while leaving the recipe is the surviving-attractor shape DESIGN 3 warns about: the dead formula is the one the implementation would have been written against. The desired-serving-deployment term now leads, named as the only non-empty term today and as the one a worker reading the role roster would omit — which admits six hosts and hands four of them to a new shape while Group A serves on them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Propagate the six-arm result vocabulary into 8.2's host-screen sentence Own sweep, not review-driven. 8.2 still said a shape exceeding the admissible set is "Refused"; after 7's correction that arm means the selection itself was defective. A full fleet is NoFeasibleRealization — well-formed selection, no surviving candidate — which is the distinction the six arms exist to keep. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Correct the byte-partition rung, and drop the EngramPlacement nickname Review 64146, both verified. The byte partitions are CitedToAuthority publisher figures, not FootprintObservedOnFleet — deepseek_v4_1_flash_payload_footprint_standing says so, and the only FootprintObservedOnFleet nearby is a different artifact (V4, 168 GB). Claiming the stronger arm was rung inflation in the paragraph whose job is rung honesty. The correction has teeth: weight_fit_of returns WeightFitUnderivable on FootprintReportedExternally, so 8's rank arithmetic run through the screen that exists today yields underivable for every route. The 5-vs-3 split is motivation for the decomposition, not a feasibility finding, until a Spark reads the byte spans. 14 step 1 now takes that read together with the row-size read, since one pass discharges both. EngramPlacement was a second name for the selected route, in a plan whose thesis is that placement is not a property of the component — the exact shape a worker could mint the forbidden location type from. Named as the MaterializationRoute inside SelectedWithin instead. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Recut the plan against seven side-chat findings F1 coupling. "Two selections, no joint search space" was asserted, not established, and is wrong: fill is per host, the table is row-sharded across ranks, refill scope and storage follow host count, page cache follows route and concurrency, failure cost follows host assignment, and feasible topology follows the route's resident contribution. No invariance proof exists. New 7.1 takes the parameterized shape — route selection per exact serving candidate, outer selection over composed candidates — keeping both authorities distinct without a context-free route winner. Max admitted latency leaves the demand carrier; it is serving policy and belonged to the serving subject, as section 4 already said. F2 logical vs physical. 48 hash IDs is implementation-plus-config derived (n_hash_cols = (max_ngram_size-1)*n_heads), not config-only. The page-read chain is a scenario under named assumptions, not derivation: weight and scale are separate tensors so 264 B is a summed logical payload rather than a layout fact, and page count, coalescing, readahead, cache hits, host page size and per-rank ownership are all unestablished. Split into ComponentLogicalAccessDemand and RoutePhysicalAccessStanding with the standing inseparable from the number. F3 resident memory. "Zero resident bytes" for a non-resident route is fail-open: B and D depend on a retained page cache by construction. RouteResidentMemoryStanding with an unestablished arm that keeps the fit open. TP8/TP4 restated as lower-bound scenario, matching WeightsNotRuledOutByLowerBound. F4 census. Placement has four arms including LocalAccelerator; the hop restated placement and retention that tier_axes and CacheProvider already derive, making contradictory hops writable; CasTier is keying, not locality; ArtifactRequest is a closed coproduct over compiler artifacts and cannot host Engram; the component digest does not exist. Adds the subject/artifact/realization trio. The net-new count is now open rather than "exactly three". F5 consumption. A witness is not the production consumer: step 2 merges with the candidate-field consumption or lands as a declared frontier. F6 two gates. serving_runtime_capability_probe says of itself that presence is prerequisite inventory establishing nothing about serving, with feasibility belonging to RuntimeFeasibility and a launch receipt. Absence rules out; presence does not rule in. Digest grain corrected to the observed local config digest. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Four bounded plan-text repairs from the side-chat verdict 1. Missing-evidence semantics were self-contradictory across 8.3, 11.1 and 14. std.decision separates unfunded_constraint_axis_defects (constraint naming an axis ABSENT from declared axes: a malformed input) from unread_axis_gaps (a declared axis with no candidate reading: an evidence need). The plan conflated them. 11.1 now states four states in a table; today's output is SelectionNeedsEvidence, not RealizationSelectionRefused. Also deleted the automatic "once the probe lands" progression: a result depends on what a probe OBSERVES plus the remaining evidence, not on an instrument existing. And says the runtime axis is declared on purpose — omitting it to obtain a louder refusal would preserve the six arms' spelling while destroying their meanings. 2. The capability lookup key was wrong, and dangerously. The merged witness the_published_ask_is_keyed_on_the_pull_pin_not_the_config_id asserts the observation is keyed on the pull pin and that architecture_observation_exists(config_id) is FALSE. The prior text would have had a worker query pin-keyed inventory by config ID and read the miss as absence of support — manufacturing a measured negative from a wrong key. Execution identity and observation lookup key are now distinct, with all four terms of the relation kept. 3. Three claims outran their carriers. The 6.2e-8 ratio is (48 x 264 B) / table and rests on the ungrounded 264 B assumption, so it is conditional. 8.3 still called the 48 config-derived; it is implementation-plus-config. 14.1 claimed a byte-span read would make the arithmetic a feasibility finding; it establishes neither resident memory nor working set nor execution. And the third-party patch inference is withdrawn — someone patching their build observes nothing about the published digest. 4. Host set: consume gunbc.spark.host_commitment rather than restate a subtraction, since an independently enumerated recipe misses whatever category lands next — which is how this plan missed the serving-arm claim. No scalar replaces "four"; the document no longer carries the host list as data, which is the rule 8.2 already stated and 8.2 itself broke. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Make 14 defer to the sections, and merge main so the cited witness resolves here Review 64308, all three verified. The citation finding is a base problem, not a false citation: the witness the_published_ask_is_keyed_on_the_pull_pin_not_the_config_id and architecture_observation_exists are on origin/main but were NOT in this branch's tree, which predates that merge. I verified them against origin/main and cited them as if the PR tree carried them. Merged main so the symbols resolve in the corpus this PR actually presents. Steps 3 and the closing line each restated a reading the section above had already corrected — step 3 keyed the probe on the config digest, the exact inversion 11.1 says manufactures a false negative, and the closing line said refusal where 8.3 now says SelectionNeedsEvidence. Both are the brief a worker executes, so a stale restatement there outranks corrected prose everywhere else. Three reviews have now found this same shape in 14. The structural repair, stated in the section: 14 states no semantics of its own — every step names the section that owns its meaning, and where a step and a section disagree, the section wins and the step is the defect. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Authorised edits for review 64320, committed and DELIBERATELY UNPUSHED Held per the freeze: the side-chat is reading fc2d382, CI is saturated, and a push now lands a verdict on a stale head for the second time tonight. Pushes when the manager hands over the verdict. Finding 1: section 11 restated "only four are admissible" three sections after 8.2 forbade carrying the host list as data. Count dropped; the displacement argument stays by reference to 8.2. And four is not merely forbidden-as-data, it is now wrong — host_commitment derives three uncommitted units since srv9 took the GLM canary. Recorded in 8.2 to show the value moved, explicitly not as a value to consume: replacing four with three repeats the mistake in a fresher costume. Finding 2: the four status claims do NOT resolve alike, and checking each against gh pr view at edit time rather than memory is the whole content of the finding. #11013 MERGED and #11088 MERGED, so status clauses go and the symbols are cited plainly; #11107 is OPEN, so host_commitment's in-flight wording is still true and STAYS. Sweeping them uniformly would trade one false status claim for another. Adds a standing instruction beside the section 14 rule: merging main does not only make citations resolve, it silently falsifies every "not yet on main" claim in the document, with no conflict and no diagnostic, and those claims read as settled so a re-read skips them exactly when they have gone wrong. Same class as the section 14 rule — a restatement that nothing re-derives — which is why they sit together. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Side-chat verdict: three lateral contradictions, swept rather than patched Folded with the two REQUEST_CHANGES edits authorised earlier — one push, one head for the reviewer. 1. 8.3 still promoted a byte-span read into feasibility, which corrected 14.1 already denied. That mattered more than a duplicate because 14 DEFERS to 8.3, so the precedence rule routed authority into the error. Replaced with the reviewer's text: the read establishes the fleet-read payload input and not loaded resident memory, route working set, or execution feasibility; the positive arm stays WeightsNotRuledOutByLowerBound. 2. Host claims in three places, not the two named. Section 11 takes the reviewer's wording — a candidate requiring more hosts than host_commitment makes available is excluded, and if that empties the field it is NoFeasibleRealization with causes named. Deleted the 8.2 bullet reasoning admissibility from absence in the preceding categories, immediately before the text explaining why an omitted serving-arm claim invalidates exactly that. And 8.1 carried a third present-tense count the verdict did not name; the sweep found it. Also removed "three uncommitted units" from 8.2, which I had written as a demonstration. The instruction is no replacement number anywhere, and a figure labelled do-not-consume is still a figure that goes stale; the supply gap is an operator decision this plan does not resolve. 3. The patch inference's remainder: "establishes that the capability is not universally present" stood in the same paragraph that granted the author may never have tried the stock path. Replaced with the reviewer's sentence. All three share the shape section 14's note diagnoses, one axis over: sections contradicting each other laterally rather than 14 restating a section. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf * Remove the host-feasibility CONCLUSION, not just the number that supported it Verdict finding, and it is the same class one level subtler than the last: the availability number is gone but the conclusion it was supporting stayed, stated independently. 8.1's Route A bullet delegated the answer to host_commitment and then supplied it anyway in the next clause — my own replacement text, not survivor prose. Suffix deleted; the TP degree stays, since that is the scenario's demand rather than an authored availability figure. 8.2's "infeasible without displacing a running four-rank serving deployment" paragraph deleted entire, no replacement: section 11 owns the host-fit rule. Swept for the conclusion the way I swept for the number, and found a third: section 11 named "release Group A" as a remedy, which is today's incumbent as data. Now names constraint kinds instead, with the note that which is cheapest to move depends on what is committed at the time and is not a property of any route. The distinction worth carrying: a host-fit refusal establishes that a candidate cannot use the presently admitted supply. It does not establish that a particular incumbent must be displaced whenever that route is selected — a sufficient uncommitted population removes the exclusion without displacement entering the route at all. "A is infeasible without displacing" smuggles today's supply into a permanent property of the route, which is the stale-fact mechanism wearing the shape of a conclusion instead of a number. #11107 re-checked at edit time: still OPEN, so host_commitment's in-flight wording stays. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DiKZ27VVdsn6hn9V8fN9Kf --------- Co-authored-by: gunbc-ci-auto-heal <gunbc-ci-auto-heal@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Summary
replica_count == 1(tp4x2 fails; tp2x1 and tp4x1 pass). Idle hosts are admitted.(tp, replicas)withtp * replicas <= host_count(fleet_admitted_topologies). Covering-only lists remain GLM/V4 Flash. Runtime TP width set is named unestablished (DSV41-1), not guessed;tp4x1is in the field.KvFp8DsMlaLayout(fp8_ds_mla); bytes-per-token stay unestablished. Speculation is DSpark. Window is yarn-scaledmax_model_len1048576.v41_published_arm64_reference(), and the unread fleetspark_runtime(source unestablished).FootprintReportedExternallyfrom the cited payload510286023000(not a fleet reading).6099d0d9a2.selection_policyisdeepseek_v4_1_selection_policy, not the quality-requirement declaration.Receipt
select_serving_deployment(deepseek_v4_1_decision_input)isServingDeploymentNeedsEvidence(on-fleet footprint, DSV41-1 runtime feasibility, quality evaluation, unfunded axes).Test plan
replica_count_above_one_fails_no_redundancyv4_1_field_is_complete_over_every_shape_the_fleet_admits(three profiles; published pin is arm64 repository@digest)v4_1_gb10_kv_layout_is_fp8_ds_mla_and_bytes_per_token_are_unestablishedchanging_axis_goal_changes_the_decision_identity/flipping_no_redundancy_goal_on_tp4x2_changes_feasibilitythe_real_deepseek_v4_1_subject_returns_obligations_rather_than_a_topology