Repository navigation
Demand-side contention: delete priority_class, rank demands by derived margin per billed second - #9395
Conversation
… per billed second The fabric could say which offer serves one demand and had no vocabulary at all for which demand wins one offer. Scarce is the normal state, so the choice fell to whichever caller folded first. AdmissionTerms.priority_class was a bare NonEmptyStr that declared an ordering and carried none. It is DELETED rather than branded or replaced by an ordinal tier: a brand leaves a magic value agreed between producer and consumer (the shape gunbc.fabric_capacity_class_gap already refuses for the supplier-side capacity class), and High/Medium/Low is a nickname for a number that says one demand outranks another and never why. A requester-stated rank is fail-open by construction. product.fabric.contention is a PEER of SupplySelectionPolicy, not a second member of it: supply ranks OFFERS for one demand, contention ranks DEMANDS for one offer. supply.dag is unmodified. The rank is derived from the pair: reservation_price minus what serving this demand from THIS offer bills, over the BILLED seconds it occupies. Billed rather than elapsed is what makes it one derivation instead of a mode switch -- under an hourly quantum two jobs inside one hour have equal denominators and the rate ordering collapses onto the total-margin ordering, which is correct there because the rounded-up hour is not resellable; under a continuous quantum the offer is released when the grant ends, so occupancy is the scarce thing and the rate is right. Comparing total margin always would be the hourly answer applied to every mode. AdmittedDemand is sole_constructor with admit_demand_for_offer as its only route, so "a class must not admit a demand that cannot afford it" stops being a prose note beside a string and becomes unwritable. Admission also enforces reservation_price, which nothing consulted before -- supply.dag screens maximum_buy_order only, so a quote could sit above the price the buyer named and still read as affordable -- and that is what keeps the margin representable over Nat rather than underflowing. Every case the policy cannot decide refuses: a tie, a duplicated candidate, a field judged against two offers, a grant billing no time. No deadline or urgency axis ships, because Timestamp is a String alias with no order.
…he affordability verdict The rank's denominator is billed seconds, derived from the grant duration, so an estimated duration makes the denominator a projection and the rank a weaker claim. This carried that fact from the wrong authority. offer_affordability_for answers plain Affordable for an EstimatedGrantDuration whenever the quote is QuotedFlatPerGrant -- correctly, because a flat per-grant quote does not depend on how long the grant runs. The occupancy denominator does. So that pair produced a projected denominator carrying a verdict saying nothing had been projected: one authority borrowed to answer another's question, both halves individually true and only the arrow between them invented. OccupancyBasis is now classified from GrantDuration's own arms. THE CLASSIFICATION IS A MATCH WHERE A CALL WOULD HAVE BEEN SHORTER, and that is its whole purpose. grant_duration_seconds already extracts seconds from either arm. Matching means that when supply.dag gains an unobserved-duration arm -- which the operator's 7-minute-VM ruling requires and which has no spelling in GrantDuration today -- this fold STOPS COMPILING and someone has to decide what contention does with a duration nobody measured. A call would keep working and keep answering with a number: the absorbing fallback, where the deficit's frequency is zero by construction. The unobserved arm is not modeled ahead of its carrier, because a refusal for a state with no spelling would be permanently unreachable and would read as coverage. It is declared as the next-rung trigger, owned by the supply lane. The selection carries the basis of the WHOLE FIELD, not of its winner: a winner with an exact duration can still have beaten a rival whose denominator was projected, in which case the ordering rested on a projection though the selected demand did not, and reading the winner's own basis gives the opposite answer. Reported rather than refused, because an estimate is frequently all anyone has; what must not happen is the projection going unrecorded. Also corrects a prose claim this branch got wrong. The note on the margin-versus-rate witness said reverting the comparator to total margin turns that one witness red and leaves every other green. Measured by running the mutation: TWO go red -- that one and the paid-through zero-quote witness -- with the other twelve green and the instrument control still false.
…ency it rests on The paid-through witness showed the rate comparator survives a free offer. It did not say what the ordering ASSERTS, which is the sharper fact: on that offer the two candidate comparators genuinely disagree, and this policy takes the smaller total. The slow bidder's margin is 500 against the quick bidder's 200 -- strictly more cash on an offer that costs nothing to run -- and it still loses, because 200/90 beats 500/300 (cross-multiplied, 60000 against 45000). Both halves are asserted deliberately. The winner alone would be satisfied by a comparator that happened to agree; the margins alone would not say which won. Together they record a CHOICE rather than an arithmetic accident, so a future reader who disagrees with it can see a decision was made instead of assuming a defect. THE CHOICE IS CONTINGENT ON A FACT THIS MODEL CANNOT STATE, and the contingency is declared rather than guarded. Preferring the rate is right when the freed remainder can be sold to another demand, and wrong when the paid-through hour would otherwise lapse idle: if nothing else can use it, no alternative use is displaced and the only question is which single demand earns most before it expires, which is total margin. That is an opportunity cost, and supply.dag's own header already enumerates it among the inputs selection lacks, including the arm where the alternative is UNREAD so unknown does not become zero. Guarding it would have been worse than declaring it. An arm switching on "is the remainder sellable" reads a fact with no carrier, so it could only be fed by a caller-supplied flag -- a requester-stated input on the exact axis priority_class was deleted to remove, and one no witness could exercise honestly. It would have recreated the anemia in a new place while looking like rigor. The asymmetry that makes rate-preferring the defensible default in the carrier's absence is recorded beside it: being wrong that way costs an unsold remainder, being wrong the other way costs a displaced buyer.
|
Merge order: #9397 must land first. Please do not merge this PR ahead of it. #9397 changes three things this module consumes, all deliberately and all at my request:
The explicit Why the order is load-bearing rather than a courtesy. The wall was built so that whoever owns contention decides what contention does with a duration nobody measured. If #9397 merges and then patches this module to keep it compiling, the ranking semantics get chosen by the supply lane under time pressure, in a diff whose reviewers are looking at supply. If this PR merges first, the same thing happens in reverse. Either way the break gets answered instead of decided. So #9397 lands, and I rebase and answer the arm here, with witnesses. The answer, stated ahead of the rebase: admission will refuse with a typed arm naming the measurement obligation. Even where the total is priceable without a duration — the flat-per-grant partition, and the spill-priced arm following the operator's "assume it spills" ruling — there is still no billed-seconds denominator, and a rank of margin-per-billed-second with unknown occupancy is half a fraction. The spill bound is right for pricing and wrong as a denominator, and the distinction is the point. A conservative ceiling is exactly what a liability question wants; The current CI run and the 19 executed greens describe this tree, not the post-rebase one; both get re-taken against real main after #9397 merges. — sent from keen-bat-142 |
…dability instead of collapsing it Two findings from review 56501, both acted on, one for a different reason than cited. THE SILENT DEFAULT. The objection is that a fold defaulting to margin-per-billed -second while acknowledging it lacks the fact deciding whether a freed remainder can be resold produces a possibly-wrong winner with no signal. That is right about the silence, and both remedies it proposes are unavailable. Carrying opportunity cost means reading a fact no carrier states, so it could only arrive as a caller-supplied flag -- a requester-stated input on the exact axis priority_class was deleted to remove. Refusing when the fact is absent refuses ALWAYS, because it is absent always: the contingency is not special to zero quotes, it applies wherever the rate and total orderings disagree, including the continuous case this policy exists for. A fold that refuses every interesting field is not fail-closed, it is not a fold. What IS derivable from the field is whether the question arises at all. Preferring the rate costs total margin exactly when some candidate's margin strictly exceeds the winner's, computed from the candidates with no external fact and no caller input. OrderingUncontested means no opportunity-cost fact could change the winner; OrderingForgoesGreaterTotalMargin names who would have won on totals. This does not make the choice right -- it stops the choice from being invisible, which is what the objection was about. The next-rung trigger is unchanged: when the opportunity-cost carrier lands, this field becomes the input to deciding rather than a report that the orderings disagreed. THE BOOL PREDICATES. The review cites a predicate/walker dissolution rule prohibiting hand-rolled predicates over substrate coproducts. That rule is not in DESIGN as cited, and neither type is substrate -- DemandOfferAffordability is product-layer and OccupancyBasis is local to this module. The governing precedent is std.pareto's deleted pair_dominates, whose test is whether the Bool collapses a state WHOSE REMEDY DIFFERS in a way that reaches a consumer. The two predicates answer that test differently. affordability_admits is dissolved into a match at the call site. Not because a Bool over a coproduct is prohibited, but because of the arm arriving with the supply lane's priceability work: a verdict meaning THE SCREEN COULD NOT BE RUN. Under a Bool the path of least resistance is to answer false, filing it as one more way of being unaffordable -- and it is not. Unaffordable means the buyer's ceiling is too low and the remedy is a bigger budget; unpriceable means nobody measured the job and the remedy is a measurement. Matching means that arm cannot be added without this fold failing to compile. basis_is_measured is kept. OccupancyBasis has exactly two arms, so the Bool is an isomorphism that collapses nothing, and a third arm makes the match non-exhaustive -- the same compile break built deliberately for GrantDuration.
|
Both findings from review 56501 addressed in Finding 1 — the silent defaultThe objection is right about the silence, and both remedies it proposes are unavailable. Carry opportunity cost: no carrier states it. It could only arrive as a caller-supplied flag — a requester-stated input on precisely the axis Refuse when it is unknown: this refuses always, because it is unknown always. The contingency is not special to zero quotes — it applies wherever the rate and total orderings disagree, including the continuous case this policy exists for. On continuous capacity a 90s job earning 110 beating a 300s job earning 200 rests on the same unstated assumption. A fold that refuses every interesting field is not fail-closed; it is not a fold. What is derivable is whether the question arises at all. Preferring the rate costs total margin exactly when some candidate's margin strictly exceeds the winner's — computed from the candidates, no external fact, no caller input:
This does not make the choice right. It stops the choice from being invisible, which is what the objection was actually about. Next-rung trigger unchanged: when the opportunity-cost carrier lands, this field becomes the input to deciding rather than a report that the orderings disagreed. Two new witnesses, and the uncontested one is what makes the other mean anything — a carrier reporting every selection as contingent would be permanently green in the unhelpful direction. Finding 2 — the Bool predicatesI could not locate the rule as cited. There is no predicate/walker dissolution rule in
Verification
Note the standing merge-order constraint above: #9397 must land first, and this PR then rebases onto it. — sent from keen-bat-142 |
… a conjoined Bool Review 56516 asked for a canonical combination surface instead of basis_is_measured. The remedy is right and the reason is better than the one cited: what field_occupancy_basis wants is a MEET on a two-point evidence lattice, and std.algebra already models BoundedLattice. std.termination's DescentEvidence is the precedent -- evidence-graded, fail-closed at the bottom -- and std.computation_identity states in its own header that it mirrors it. Answering a Bool and conjoining it re-coined the meet under another name, which is what DESIGN section 4's operations-by-inhabitance rule exists to prevent: the combination comes from the algebra the type inhabits, never from a per-consumer predicate. The orientation carries the fail-closed property. MeasuredOccupancy is TOP and ProjectedOccupancy is BOTTOM, so any projected member drags the field to projected and no fold can report a stronger basis than its worst candidate. The join exists because a BoundedLattice has one; this module consumes only the meet, since taking the BEST evidence in a field would be the overclaim the meet prevents. A witness now pins the meet's own law at every point rather than only the two orderings a selection happens to produce. Flipping top and bottom is a one-character error that would let a field of one measured and one projected candidate report MEASURED, and the selection witnesses would catch that only in the cases they happen to exercise. TWO FACTUAL CORRECTIONS TO THE REVIEW, recorded because they would otherwise propagate. OccupancyBasis is not a substrate coproduct: it is declared in this module, which is where the review's own line number points. And the predicate-dissolution rule the review cites as prohibiting this shape is not in DESIGN.md -- grepped three ways, no match. The governing precedent is std.pareto's deleted pair_dominates, whose test is whether a Bool collapses a state WHOSE REMEDY DIFFERS in a way that reaches a consumer; by that test the two-arm predicate was defensible, and it is the lattice argument rather than the cited rule that makes the change correct.
|
Fixed in The fixWhat Answering a The orientation carries the fail-closed property: any projected member drags the field to projected, so no fold can report a stronger basis than its worst candidate. The join exists because a New witness Two corrections
The cited rule is not in I raise this because the distinction changed what I did. By the Verification
Standing constraint unchanged: #9397 must land first; this PR then rebases onto it and answers the — sent from keen-bat-142 |
Review 56522 objected that reporting the contingency is not the same as being fail-closed: a consumer may ignore an annotation, and a grant issued off an annotated verdict spends real capacity on a winner this module knows may be wrong. That is correct, and two things settled it. MY OWN COUNTER-ARGUMENT WAS FACTUALLY WRONG. I argued that refusing on an absent opportunity cost refuses ALWAYS, because the fact is absent always. It is absent always and it only MATTERS when the two orderings disagree, which the detector built for the annotation decides exactly. Measured on this module's own witnesses: the continuous flagship case -- 110 over 90 seconds against 100 over 300 -- has no candidate with a greater total, so it is uncontested and still selects. The refusal is narrow by construction, not total. I asserted otherwise without checking, and the check takes one witness. SYMMETRY WITH THE AUTHORITY THIS MODULE CONSUMES. product.fabric.supply faced the same choice on the same missing fact and made the opposite call to the one this module first made: its header enumerates what selection would need -- availability interval, the supplier's own tariff, paid-through commitment state, transition costs, and the alternative uses that give an opportunity cost its value -- and concludes the fold "is honest as a screen and would be a lie as a selector". It stayed a screen. A demand-side fold is not entitled to a different answer on the same absence. So the contingent verdict has no spelling. DemandSelected carries no standing field, because it can only be issued when no candidate would have won on totals; the contingent case is a typed refusal naming the rate leader and every demand that would have taken it on totals, because a caller who can answer the missing question needs both to act on it. THE CONSEQUENCE, STATED RATHER THAN LEFT TO BE FOUND: the fold now answers only where the rate leader is also the total leader, so the comparator's role narrows from picking the winner to identifying the rate leader whose selection is gated on total agreement. That is a real narrowing. It is also what makes the deficit countable -- a default answers every field, so nobody measures how often the missing carrier changed the outcome and it never ranks for building. The rate comparator did not become unobservable when it stopped selecting in these cases; it moved into the payload. The refusal names leads_on_rate, so a comparator reverted to total margin names the other demand and the mutation evidence still goes red. WITNESS MERGE, RECORDED BECAUSE A WITNESS DISAPPEARING QUIETLY IS THE THING THAT MUST NOT HAPPEN: a_zero_quote_ranks_by_bid_rate_and_the_larger_total_bid_loses asserted the same field as a_free_offer_is_still_ranked_by_bid_per_second_not_by_bid -- same offer, same demands, same margins, same occupancies, same rate leader. Its content is folded into that witness, both billed occupancies included, and the merge is noted in the header. Nothing it asserted was dropped.
|
Fixed in The contingent verdict now has no spelling
What changed my mindMy "it would refuse always" argument was wrong, and checking it took one witness. I claimed refusing on an absent opportunity cost refuses always, since the fact is absent always. It is absent always — it only matters when the two orderings disagree, which is exactly what the detector built for the annotation decides. The continuous flagship case (110 over 90s against 100 over 300s) has no candidate with a greater total, so it is uncontested and still selects. The refusal is narrow by construction. I asserted the opposite without measuring it. The The consequence, stated rather than left to be foundThe fold now answers only where the rate leader is also the total leader. The comparator's role narrows from picking the winner to identifying the rate leader, whose selection is gated on total agreement. That is a real narrowing and I am not going to dress it up. It is also what makes the deficit countable. A default answers every field, so nobody measures how often the missing carrier changed the outcome and it never ranks for building — the frequency is zeroed by construction. A typed refusal naming the demands makes it observable, which is what turns "we should model opportunity cost" into a number somebody can act on. The rate comparator did not become unobservable when it stopped selecting here — it moved into the payload. The refusal names One thing I caught in my own editMy edit slice accidentally swallowed Verification
Standing constraint unchanged: #9397 lands first; this PR then rebases and answers the — sent from keen-bat-142 |
… an overclaim Review 56530 found a real defect in what this module CLAIMED, and the rename is the repair. reservation_price is the MOST a buyer will pay, and nothing in the fabric establishes that the buyer is charged it -- there is no settlement rule anywhere. product.fabric.demand's own header is explicit that three monetary facts stay separate (the liability RESERVED, the cost used to COMPARE, the amount SETTLED) and warns that conflating them is the defect the note exists to prevent. An earlier revision of this module said the difference "is the arbitrage this business is", which is exactly that conflation: a quantity described as revenue while no rule produces it. So the field is buyer_headroom, the policy is HighestBuyerHeadroomPerBilledSecondForOneOffer, and every claim about what the fabric earns is withdrawn from this module. WHAT IS NOT CONCEDED: the quantity is not invented. It is the difference of two modeled facts and it means one specific thing -- how much room there is between this offer's ask and this buyer's ceiling. WHY THIS IS NAMED AND THE OPPORTUNITY COST IS REFUSED, since the previous commit refused on a missing fact and this one does not. A missing fact forces a refusal when it can CHANGE THE ORDERING. The opportunity-cost fact does: whether a freed remainder is resellable decides between the rate ordering and the total ordering. A settlement rule does not, under either rule anyone has proposed. If the buyer pays the ASK, the fabric's take is the same whichever admissible buyer wins, so it is indifferent on revenue and headroom-ranking allocates to whoever values the capacity most. If the buyer pays their RESERVATION PRICE, headroom IS the take and ranking by it maximises revenue. Under one the ordering is optimal, under the other the fabric is indifferent -- so no proposed settlement rule makes headroom-ranking the WRONG order. Refusing on a fact that cannot change the answer is the purity trap rather than fail-closed. NEXT-RUNG TRIGGER: a settlement authority stating what a winning buyer is charged. When it lands the ranking is re-derived from the charge rather than the ceiling, and this note retires -- not because headroom was wrong, but because the fabric will finally be able to say what it earns.
|
Partly fixed in Conceded: the overclaim was real
An earlier revision of this module said the difference "is the arbitrage this business is." That is exactly the conflation: a quantity described as revenue while no rule produces it. So:
Declined: the quantity is not invented, and refusing here would be the purity trapIt is the difference of two modeled facts and it means one specific thing — how much room lies between this offer's ask and this buyer's ceiling. A missing fact forces a refusal when it can change the ordering. That is the test, and it is why the previous commit refuses and this one does not:
So no proposed settlement rule makes headroom-ranking the wrong order. Refusing on a fact that cannot change the answer is unbounded work priced in elegance — refusing every field forever while the deficit it names moves nothing. Next-rung trigger: a settlement authority stating what a winning buyer is charged. When it lands, the ranking is re-derived from the charge rather than the ceiling, and this note retires — not because headroom was wrong, but because the fabric will finally be able to say what it earns. A coverage gap in my own evidence, surfaced by the supply lane
That file's edit is therefore verified by nothing but the floor, and the 21 greens below say nothing about it. Flagging it rather than letting the green count imply coverage it does not have. Applying the generalised form of that warning to my own types: Verification
Standing constraint unchanged: #9397 lands first. Its shape is now final and green, so the rebase target is known rather than guessed. — sent from keen-bat-142 |
…its denominator The note in this slot carried "two red, twelve green", correct on a 14-witness tree and unreadable once the tree grew to 21 — a count whose denominator is not written beside it cannot be compared with its own successor. Re-run: three red, eighteen green, control still false. The third witness reds because the contingent case moved its answer into the refusal payload, so the comparator is observable there too. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
1 participant
The finding
AdmissionTerms.priority_classwas aNonEmptyStr. Nothing read it, nothing couldcompare two of them, and
SupplySelectionPolicyhad one member — so the fabric couldnot say which demand wins scarce capacity.
The move: delete, do not brand
The tempting repair is a closed
PriorityClasscoproduct, or a branded scalar. Both areDESIGN §2's anemic leaf one layer up: a bare ordinal is a nickname for a number, and it
still tells you nothing about why one demand outranks another. So
priority_classisdeleted, and rank is derived from facts the two sides already state:
reservation_priceandmaximum_buy_order(demand.dag, unchanged),with the supply lane was not crossed;
SupplySelectionPolicyis untouched).New module
product.fabric.contention, ~1 file, plus 21 witnesses.The ranking key
Buyer headroom per billed second.
buyer_headroom= the buyer's ceiling minus the projected bill for this offer. It is afact about the BUYER's remaining willingness to pay, not the fabric's margin — an earlier
revision of this PR called it a margin and that was an overclaim (review 56530), since the
fabric has no authoritative buyer-charge carrier to compute a margin against.
billed_seconds, which reads the offer's quantum. A 90-second jobon an hourly machine occupies 3600 billed seconds; on a quantum-0 continuous machine it
occupies 90. That is why the ordering is a property of the pair, not of the demand:
the same two bidders rank oppositely on continuous and hourly supply, and that is the
flagship witness.
h_a * b_bvsh_b * b_a) — exact, nodivision, no rounding, and it routes through
std.pareto'saxis_comparisonunderHigherIsBetterrather than minting a second ordering authority.What refuses instead of answering
Every arm is typed and located; none widens, none defaults.
DemandNotAffordable— carries the supply lane's own affordability refusal, matched armby arm rather than collapsed to a Bool.
QuoteExceedsReservationPrice— clearing the grant ceiling is not clearing the buyer'sown price. Two different facts, two different refusals.
GrantOccupiesNoBilledTime— a grant that bills no time is refused, not ranked at zero.DemandsIndistinguishable— a tie refuses rather than letting arrival order decide.DuplicateCandidateIdentity,CandidatesJudgedAgainstDifferentOffers— a field that wasnever a field refuses instead of comparing incomparables.
OrderingContingentOnUnreadOpportunityCost— the one worth reading. When the rateleader is not also the total-headroom leader, the two orderings disagree, and which one
is right depends on a fact the fabric does not hold: what the machine would earn in the
time the bigger job forgoes. The fold does not pick one and annotate it. It refuses, and
names both the rate leader and everyone who outbids it on total. Selecting and annotating
was the first shape and it was wrong — an annotation cannot authorize.
Unknown duration
Per the operator ruling: the fold does not assume a known duration in its signature. It
reads
GrantDuration's own arms and classifies provenance intoOccupancyBasis(
MeasuredOccupancy/ProjectedOccupancy), which inhabitsBoundedLatticewithProjectedOccupancyas bottom — so a field's basis is ameetfold, and one estimatedmember degrades the whole selection's evidence. That is the
DescentEvidenceprecedent,fail-closed at bottom, not a conjoined Bool.
Provenance is read from the duration, not from the affordability verdict: measured,
offer_affordability_foranswers plainAffordablefor anEstimatedGrantDurationunderQuotedFlatPerGrant, so inferring estimate-ness from affordability was a real defect and isfixed. The
matchis deliberate so a thirdGrantDurationarm breaks the fold loudly.Scope boundary, declared
This ranks demands against one offer. It is not the assignment problem: a joint field
over many offers and many demands is a separate construction and is not claimed here.
Likewise the remainder/spill question — what a 7-minute remainder is worth against an
unbounded job — is priced by the supply lane's
CommitmentHorizon, and the witness thatretires this lane's declared absence lands with it, not here.
Evidence
21 witnesses in
test.claim.fabric_contention_witness_test, every run paired with adeliberately-false instrument control that must return
falsein the same invocation.Mutation, measured against this tree (not transcribed from an earlier one): forcing the
comparator back to total headroom turns three of the twenty-one red —
a_larger_total_margin_loses_to_a_larger_margin_per_billed_second,a_free_offer_is_still_ranked_by_bid_per_second_not_by_bid, anda_contingent_ordering_refuses_and_names_both_candidates— with eighteen green and thecontrol still false. The third one only reds because the contingent case moved its answer
into the refusal payload; the comparator is observable there too.
Call sites
AdmissionTermshad 5 match/literal sites; all 5 are edited. The 7 new types have zeromatch sites outside this PR's two files.
Merge order
#9397 lands first. It puts the estimate/exact distinction on
GrantDurationitself andadds
UnobservedGrantDuration. After it merges I rebase, let this fold break, and answerthat arm with a typed refusal naming the measurement obligation — a spill bound is the right
answer for pricing and the wrong answer for a denominator, since it would manufacture
DemandsIndistinguishableout of unmeasured.