Repository navigation
HOST-1: program-derived assimilation completion gate (obligations as a derived list, not a fixed product) - #7877
Merged
Merged
Conversation
…rived First slice of the local-inference economics lane. Lands the receipt carrier that makes the hardware question measurable, in the observing product layer (DESIGN external-upstream-decomposition: extdeps.llm.* owns provider wire contracts, this owns what WE observed when we ran something). Mechanism: - TokenCount instantiates the std.measure Count axis (HardwareThreadCount / CharacterCount precedent), declared here not in std.measure because an LLM token is domain vocabulary and std/measure.dag is seed-emitted. - No throughput, rate, or cost field exists in the record. Tokens-per-second and accepted goodput are DERIVED as exact FieldOfFractions<Nat> over the same authority gunbc.econ.acquisition ConversionRate consumes; the floor projection is a declared lossy view, total by typed refusal. - Usage is three-valued: an attempt whose provider reported no usage is not an attempt that consumed nothing. Missing usage refuses the economic derivation without refusing the attempt or its disposition. - Goodput counts ACCEPTED dispositions only, so a host that doubles token rate while halving acceptance shows no gain. - Admitted sessions and active sequences are separate observations; a server sustaining four sequences can admit considerably more sessions. - SubstrateIdentity binds the compiler revision, because escalation rate is endogenous to both the model improving and gunbc needing less judgment. Evidence: 13 witnesses green by execution, incl. RED controls for zero and absent decode duration (no fabricated throughput), unobserved usage, empty roster, and zero observation window. Honest limits recorded in-carrier, both measured rather than assumed: - TokenCount is structurally identical to HardwareThreadCount, so it buys reading clarity, not a construction wall. - Requested/resolved model brands do NOT wall substitution. A scratch probe passing RequestedModelIdentity into a ResolvedModelIdentity parameter resolved and ran green. An earlier draft of the note claimed a typecheck refusal and cited a witness that was never written; both were false and are retracted in the carrier with the next-rung trigger named. No cost derivation lands here: the token-rate rows are not yet cited, and fabricating a rate to make the arithmetic run is what the follow-on prevents. This slice declares its frontier rather than claiming a production consumer. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…y, don't fork it review 48816 found the merge-blocking defect and was right. TokenCount was declared locally in gunbc.econ.llm_attempt_receipt as Measure<Count, One, Nat>, citing extdeps.pricing.billing_units as precedent for downstream placement. Verified against the tree, the citation does not cover it. The billing_units aliases introduce their own phantom markers (PerStoredUnitMonth et al), so MoneyRate<PerStoredUnitMonth> is a genuinely new point in the type space that cannot be written in std without importing vendor vocabulary. Measure<Count, One, Nat> takes only arguments std.measure already owns, and that module already names the shape three times: HardwareThreadCount, CharacterCount, AmortizationMonths. A local fourth name was therefore a DESIGN section 3 fork, not a downstream instantiation. The review offered a tracked deferral as an alternative. Not taken: the only argument for deferring was the seed-regeneration cost, which is the cost of doing this correctly rather than a reason to avoid it, and deferring for the author's convenience is the unmarked-workaround shape DESIGN section 5 names as a line-stop signal. - TokenCount, token_count, token_count_value now live in dag/std/measure.dag beside their three siblings, under token_count_note, which records the billing_units distinction so the reasoning is not rediscovered, and repeats the siblings' honest limit: all four are structurally identical, so the family buys reading clarity and one construction site each, never a wall. - llm_attempt_receipt consumes them and retracts the bad citation in llm_token_count_authority_note. - src/v1/stage0/src/std_measure.rs regenerated (seed-emitted module). Evidence: regen_stage0 wrote 132 generated files with exactly one changed; the regenerated seed compiles; regen_stage0 --verify reports regen_divergence_count=0 with a binary rebuilt FROM the new seed, so the self-host fixed point holds rather than being self-confirmed by a stale seed. All 13 witnesses re-run green against the std-homed carrier. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
llm_model_identity_brand_note cited llm_token_count_placement_note, which the tree no longer contains: the previous commit renamed that note to llm_token_count_authority_note while re-homing TokenCount, and the honest-limit claim the sentence points at moved to std.measure token_count_note. Verified as dangling — the citation was the only occurrence of the name anywhere in dag/. Now cites std.measure token_count_note, where the Count-axis family's mitigatable-rung limit actually lives. This is the DESIGN section 3 stale-citation class in a note whose own subject is a false citation, and it is the failure mode that rule predicts: a rename invalidated a name-level reference with nobody touching the referring line. A symbol citation is at least decidable by grep, which is how review 48829 caught it; the positional form would not have been reachable at all. Swept the rest of both files for the same defect rather than fixing only the reported instance: dispatch_execution_receipt_bar_note, token_count_note, llm_token_count_authority_note, llm_attempt_receipt_frontier_note, ConversionRate, FieldOfFractions, character_count_note, HardwareThreadCount, CharacterCount, AmortizationMonths, dispatch_usage_unobserved, DispatchUsageObserved, MoneyRate and PerStoredUnitMonth all resolve. Witnesses re-run green. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… spine cross-product Two DGX Spark inference appliances (srv5, srv6) are on order. Enrolling them against the current model would manufacture obligations they can never satisfy: host_standup_spine is ONE universal list -- BMC credential convergence, virtual-media OS install, Actions runner slots, build cache, green-place pinning -- and host_phase_status crossed every enrolled host with all of it. An appliance has no BMC and no runner user, so those cells could never be converged and could never honestly be marked so. A completeness number computed over invented obligations is worse than no number. The correction is program scoping, NOT an Applicable | NotApplicable verdict. A verdict arm would keep the false cells and label them -- the parallel representation shape (DESIGN section 3) -- and leave every consumer to remember to filter. Participation now SELECTS which obligations exist, so an obligation that does not apply is not a labelled cell, it is not a cell. - New gunbc.host_assimilation_program: FleetParticipation (CiRunner | InferenceServing), HostEnrollmentIntent, HostAssimilationProgram, assimilation_program_for. Participation is a LIST because the axes are independent, and it is a role rather than a hardware kind, so a future Altra dedicated to serving does not inherit BMC obligations from what it is made of. - Phase-to-participation classification is an EXHAUSTIVE match over both closed phase enums, so a newly added phase cannot silently default into or out of a program -- it does not compile until classified. Construction, not validation. - It lives in the new module rather than as a field on HostStandupStep: that would reshape fourteen rows in load-bearing host_standup.dag, and that module's own gap-ledger note already ruled its row-reshaping change stays separable. - The program is DERIVED by filtering the spine, never a second hand-authored step list that could drift. - host_phase_status is enrollment-driven; the global-spine path is DELETED, not retained beside the new one. fleet_subsumption_hosts is now a projection of fleet_subsumption_enrollments, so consumers that only care which hosts exist are untouched and the two lists cannot disagree. Evidence, green by execution (12 witnesses on the refactored module): - IDENTITY: a CI enrollment selects the whole spine, and the runner-fleet matrix is still hosts x spine. Every pre-existing regression witness passes unchanged -- no-producer-is-UnknownRefused, the absorbing-fallback RED control, the reconcile upsert/teardown/noop diff, observation coverage, remaining work. - SCOPING: an inference enrollment carries zero BMC/runner/build-cache/ green-place/session-placement cells and exactly the four shared obligations. - TOTALITY: no spine step is unclaimed by any participation. - Downstream consumers re-run green: host_swap_backing, host_network_diagnosis, host_toolchain_components. Prefix phases are CI-runner only, deliberately. A Spark has real arrival obligations, but they are its own and get authored where that arrival is modeled; inheriting a differently-shaped host's provisioning steps because they sound generic is the exact false-obligation defect this removes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Replace the fixed four-field AssimilationCompleteGate product with a participation-derived obligation roster, roster validation (missing/unrequired refusal), and first-failure-wins evaluation. Fixes parse error from invalid ++ list concat in the WIP commit that broke regen and heal_generated_artifacts. Co-authored-by: Cursor <cursoragent@cursor.com>
… HOST-0 files)
|
Bugbot is not enabled for your account, so this pull request was not reviewed. Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs. |
Route assimilation completion obligations through assimilation_completion_inputs_for_one with match on FleetParticipation; document empty-participation enrollment invariant. Co-authored-by: Cursor <cursoragent@cursor.com>
Fold enrolled participations into a deduplicated obligation set by input identity so multi-role hosts derive all required inputs regardless of list order. Witness: inference-first both-participation enrollment must still require CI-only obligations. Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacks on HOST-0 (#7876, merged via
session/host-0-program-scoped-assimilationat c560361).Summary
Replace the fixed four-field
AssimilationCompleteGateproduct with a participation-derived obligation list. CI runner enrollment requires all fourAssimilationCompleteInputarms; inference serving requires onlyNodeRegisteredandResourcesPresent. Roster validation refuses when a required obligation is missing or an unrequired one is present (first-failure-wins in declared order). Green-place projection preserves exact refusal reasons including the#5948path for the fail-closed P3 scaffold.Test plan
Executed on head
024fdced0after merging updated HOST-0 base (post-#7851):Identity conjunct: existing spine assimilation witnesses pass unchanged;
p3_green_place_contributes_via_gunbc_gate_after_pinned_tree_landsand fail-closed gap/refusal controls green.Out of scope (separate PR): gap ledger
PhaseDispositionproduct change.