Why believe any particular statement here?

In every one of the ten areas the corpus names, four things can be named separately and are not the same thing: the party that composes the object, the party that validates it, the act that binds it, and the party whose system holds the authoritative instance.

The strongest thing against it None. This is a statement about the instrument's coverage, not about the world.

What this does not tell you That the four positions matter equally, that they are equally stable, or that any of them is worth more than another.

Agentic capability is evidenced composing, preparing, recommending, routing and extracting on consequential procurement objects in every area where any agentic capability was inspected at all.

The strongest thing against it Goods receipt and acceptance CL-03: one dated announcement names fourteen supply-chain AI agents and NOT ONE is evidenced acting on a goods receipt, an acceptance state, an inspection disposition, a return, a three-way-match receipt state or a service entry sheet. Composition is not universal. supplier-onboarding codes D08 explicitly absent on the same point.

What this does not tell you How much composition happens, by how many capabilities, at how many buyers. There is no denominator.

No inspected artifact evidences an agentic capability performing the binding act on a consequential procurement object, in ten areas and three probes.

The strongest thing against it Invoice to payment C-003 - SAP Joule releasing an invoice from a payment block and cancelling a posted document. An agentic capability, coded execution, qualified at both registers, acting on CONSEQUENTIAL records in scope for the area. It survives this claim on two qualifications the area supplied itself: both objects are reversible-by-the-acting-party, and a person instigates. THE FIRST QUALIFICATION LIVES IN A FIELD CODED IN ONE AREA OF TEN.

What this does not tell you That agentic capabilities CANNOT bind, or that none does anywhere. This rests on an estate with 94 nil routes across 74 named estates, 31 of them hard retrieval blocks.

Where a machine performs the binding act in this corpus, the machine is a configured deterministic rule, in two areas on two unrelated object families.

The strongest thing against it Purchase order and fulfilment C-003, CONTESTED - a scheduled job finalizes an autonomously awarded negotiation and generates purchasing documents under an integration user with Associated Person Type set to None. Refused on THREE NAMED GAPS, not on absence.

What this does not tell you Any deployment, adoption or prevalence. Both cases are documented product behaviour on a configured path; neither artifact names a customer or a date of use.

In Microsoft Marketplace, the publisher's Activate Subscription API call was removed from the path between a customer's purchase and the start of billing, and the party holding the authoritative subscription state was the same party before and after.

The strongest thing against it The interface-only-change counterexample is UNRESOLVED, not refuted: the archive holds no capture of the plan-configuration page, so whether auto activation existed as a setting before it was announced is open.

What this does not tell you How many plans are configured this way. Existing plans default OFF, only new plans flipped ON, and the announcement states no adoption level. Nothing here is an observation about revenue: that billing begins earlier is an observation about a record state.

In the Microsoft Marketplace case the publisher held a request that gated a transition, not the state itself; the state, the suspension power and the unilateral 30-day void were the incumbent's in the baseline period.

The strongest thing against it Recorded rather than suppressed: a JSON field comment in the same baseline artifact reads 'This is the date when the subscription was activated by the ISV and the billing started.' It is a field comment on startDate describing the manual flow, in a document whose normative text calls the publisher's call a request. It does not establish that the publisher held the state.

What this does not tell you That the publisher gained or lost anything economically. No pricing power, value capture, market share or bargaining outcome is established.

Across the nine areas on the shared infrastructure schema, the 68 adjudicated infrastructure rows carry seven distinct role values, of which transaction infrastructure that holds the official state (18) and public registers (13) are the two that hold state, form-defining-body (10) sets grammar without holding an instance, and transaction-enablement-layer (6) does neither.

The strongest thing against it requisition C-009 - GSA's FAS Catalog Platform HOLDS the MAS catalog record AND, in its own channel, DISPLACES one of the document forms that governs it. Form-defining and record-holding are not always separate parties.

What this does not tell you That these roles are exhaustive, or that 'unresolved' at 18 of 68 is a role. It is an evidence state and the second most common value in the field.

In seven of ten areas the party that defines the permissible form of the consequential object holds no instance of it, and names the systems that do.

The strongest thing against it requisition C-009 again, and it is the same counterexample: one party occupies both.

What this does not tell you That form-setting is worth less or more than holding. No economic comparison is available.

In the one historical comparison with an admitted baseline on both sides, the counterparty's dependency on the incumbent system increased while the counterparty's control did not decrease, because it had none of the relevant control to begin with.

The strongest thing against it requisition-guided-buying O-2, and it is UNRESOLVED: there the binding act itself became machine-performable inside the incumbent - a review step ELIMINATED rather than routed. HC-02 was the case positioned to adjudicate it and returned not-adjudicable-on-inspected-evidence.

What this does not tell you That this generalises. One case, one piece of transaction infrastructure, one vendor. HC-04's own limitation states it is NOT portable to other transaction infrastructure, and this area's own record contains the contrary path where the vendor activates.

The authoritative record holder is named in first-party documentation in ten of ten areas; it is the only element of the seven-element chain that is fully populated with no not-established cell and no derived value.

The strongest thing against it None on coverage. The strategic inference that holding is therefore the decisive position is the PRODUCT'S inference and is filed as P2-16, a Hypothesis.

What this does not tell you That holding confers advantage. Coverage of a field is not evidence about the value of the position it names.

Nine adjudicated rows across six areas are coded participation_mode 'execution', and that field records THAT a named subject performs an operation, not what kind of performer the subject is; three of the nine disclaim an agentic mechanism in their own limitation fields and a fourth has had its mechanism ruled unresolved.

The strongest thing against it None. This is a statement about coding, verifiable by re-running the count.

What this does not tell you That nine is a small number or a large one. There is no denominator.

Where an area's grammar attaches the binding act to a RECORD STATE, software has been evidenced performing it. Where the grammar attaches the act to a NAMED OFFICE or to a COUNTERPARTY'S ACCEPTANCE, no software of any kind has been evidenced performing it.

The strongest thing against it FILED BY THE AREA THAT PROPOSED THE MECHANISM, and it is the corpus's first Hypothesis-level claim: 'The Peppol Receipt Advice attaches acceptance to a document state rather than to a named office, and this area found no capability setting it either.' A record-state binding act with nothing performing it REFUTES THE SUFFICIENCY of the pattern. It must appear wherever this claim appears.

What this does not tell you That record-state grammars CAUSE machine binding, or that named-office grammars PREVENT it. HC-03 was commissioned to test this historically and returned not-adjudicable-on-inspected-evidence. THIS CLAIM IS CURRENT-STATE ONLY.

Across the areas where an agentic capability was inspected at the binding step, the capability stops on the composition side of the actor, system or institution that makes the outcome official while the system holding the object remains unchanged.

The strongest thing against it FOUR OF TEN STOPS ARE 'MERELY UNEVIDENCED' - inferred from silence rather than described by an artifact. Only ONE area has a stop documented by the vendor at the binding step itself. The pattern is partly a statement about what vendors publish.

What this does not tell you Anything causal, and anything about how far capabilities will go. No trend, no rate, no direction.

The artifacts that place an operation on an object and the artifacts that carry a date are largely different artifacts, and three areas report this independently as their own finding.

The strongest thing against it contract-lifecycle is 25% institutional while its grammar is published by the UK Government Commercial Function - the split is not clean in the middle. The candidate that generalised this into a grammar claim was tested against the records and REJECTED.

What this does not tell you That undocumented means absent. A missing capture is a property of the archive, not of the world.

Deploying an agent into an area whose official record, obligation or other binding object is held by another party may increase reliance on that party's system while reducing the deploying company's own work.

The strongest thing against it The mechanism it generalises is drawn from a case with NO agent in it. Generalising from a configuration change to agentic deployment is the step this claim has NOT earned, which is why it is a Hypothesis.

What this does not tell you Anything causal, anything about prevalence, and anything about who gains. HC-04 establishes no value capture and no bargaining outcome.

The durable competitive position in these areas is control of the official transaction record rather than possession of the most capable agent.

The strongest thing against it requisition C-009 - GSA's FAS Catalog Platform holds the MAS catalog record AND displaces a governing document form. If holding were the whole position, holder and form-setter would be different parties; here they are one.

What this does not tell you That holders capture more value, earn more, or win. Every economic limb of this claim is barred by the economic evidence boundary and by the absence of a denominator.

A layer that neither holds the official record, obligation or other binding object nor sets its form may still be defensible if it becomes the route by which many parties reach the holder.

The strongest thing against it None available, which is itself the problem: a hypothesis no row can bear on is not testable from this corpus.

What this does not tell you Everything economic. This is a position description, not a prediction.

No single stable answerable party has been evidenced for a consequential procurement object, in ten areas and three probes.

The strongest thing against it The nearest thing, recorded so it is not mistaken for one: the UK debarment regime names a Minister of the Crown and the FAR names the contracting officer. Both are named authorities for ONE act, on ONE object, in ONE jurisdiction. Neither is a stable answerable party across the object's life.

What this does not tell you NEVER 'no one is accountable' and NEVER 'agentic procurement has an accountability gap'. Publishable only as: answerability was tested in ten areas and was not evidenced in any of them.

Whether a completed binding act can be undone, by whom and with what remainder, is coded as a first-class property in one area of ten.

The strongest thing against it None. This is a statement about the instrument.

What this does not tell you That the other nine areas are irreversible, or reversible. THIS IS THE CORPUS'S MOST FRAGILE LOAD-BEARING CELL: the qualification that rescues P2-03's counterexample lives in a field nine areas do not carry. If the other nine had coded it, P2-03 would be stronger or narrower, and nothing in this corpus says which.

What performed the work before the current mechanism is not established in three of the four historical cases.

The strongest thing against it None. Each case records the exact routes that returned nothing.

What this does not tell you That nothing existed before. A route that returned nothing establishes that the route returned nothing. A missing capture is a property of the archive, not of the world.

Whether the SAP Fieldglass 'GLA agent' is a person, a role, a scheduled process or a model-driven capability is not established.

The strongest thing against it None. The refusal to resolve is the finding.

What this does not tell you That it is not agentic. NO inspected artifact states that either, and 'not established' is not a negative fact. THIS CASE IS REMOVED FROM EVERY HEADLINE REQUIRING A DETERMINATE CLASSIFICATION.

No causal link is established between any institutional event in the corpus and the behaviour of any capability, platform, product or buying system, in either direction.

The strongest thing against it None available, and none constructible from this method.

What this does not tell you Any sentence in which a structural property is the subject of a verb of causation.

No capability recorded in any register is established as being in production use by any named buyer.

The strongest thing against it None. No retrieved artifact names a buyer using any capability recorded.

What this does not tell you Any deployment, adoption, usage, customer count, share or momentum statement whatsoever.

Whether the buying organisation's own ERP or master-data system belongs in the infrastructure population is not established, because no such system was ever retrieved.

The strongest thing against it None.

What this does not tell you That buyer systems do not hold official records, obligations or other binding objects. They very plausibly do, and the corpus cannot see them.

The corpus establishes no prevalence, share, rate, ranking, momentum, forecast, pricing power, value capture, revenue effect, switching cost or winner-or-loser status of any kind.

The strongest thing against it The nearest economic-sounding observation in the programme is HC-04's earlier billing start, and that case's own record refuses the reading: 'That billing begins earlier is an observation about a record state and is not an observation about revenue.'

What this does not tell you Everything in the sentence above.

Across ten areas and three probes, no inspected artifact evidences an AGENTIC capability performing the binding act on a consequential procurement object.

The strongest thing against it purchase-order-fulfilment C-003. Oracle's Autonomous Sourcing Assistant: 'After an autonomously awarded negotiation is approved, the subsequent ESS job run finalizes the award and generates the purchasing documents', with the write privilege held by 'an integration user (a new user with the Associated Person Type set to None)'. It is refused on three NAMED GAPS in the artifact — which document type is created, the status it is created in, and whether any buyer runs it with approvals disabled — and not on absence.

DETERMINISTIC configured automation HAS performed the binding act, in two of the ten areas, on two unrelated object families, evidenced in each vendor's own first-party documentation.

The strongest thing against it None to the claim itself. The nearest limit is that BOTH cases are documented product behaviour on a CONFIGURED path: neither artifact names a customer, a deployment or a date, and supplier-onboarding's carries no page-visible date at all. requisition's own artifact poses the question 'What expenditures can be automatically approved?', which is a setting, not a practice.

Where an area's grammar attaches the binding act to a RECORD STATE, software has been performing it. Where the grammar attaches the binding act to a NAMED OFFICE or to a COUNTERPARTY'S ACCEPTANCE, no software of any kind has been evidenced performing it.

The strongest thing against it goods-receipt-acceptance CL-12's own attached counterexample, filed by the area that proposed the mechanism: 'The Peppol Receipt Advice attaches acceptance to a document state rather than to a named office, and this area found no capability setting it either.' A record-state binding act with no software performing it REFUTES the sufficiency of the pattern. It is the strongest objection to this claim and it comes from inside the corpus.

Agentic capability is evidenced composing, preparing, recommending, routing and extracting on consequential procurement objects in every area where any agentic capability was inspected at all.

The strongest thing against it goods-receipt-acceptance CL-03. In one dated announcement naming FOURTEEN supply-chain AI agents, not one is evidenced acting on a goods receipt, an acceptance state, an inspection disposition, a return, a three-way-match receipt state or a service entry sheet. Composition is not universal — in some areas the agentic capabilities do not reach the object at all.

The party that defines the permissible FORM of a consequential object very often holds no instance of it, and names the systems that do.

The strongest thing against it requisition-guided-buying C-009. GSA's FAS Catalog Platform BOTH holds the MAS catalog record AND, in its own channel, displaces one of the document forms that governs it. Form-defining and record-holding are not always separate parties, and the area records this against its own CL-04.

Where an area's admissibility state is held by an institution, that institution transacts with none of the parties the state binds.

The strongest thing against it THE ABSENT ONE, and supplier-onboarding names it against its own claim. Its frozen contract names the buying organisation's own ERP or MDM as a candidate holder, NR-22 records that no artifact naming a specific buyer's master was ever retrieved, and the area states that if that unit were inspected the claim could fail. the infrastructure examined uniform institutional orientation is an artifact of that gap, not a finding about the area.

Across ten areas and three probes, no single stable answerable party has been evidenced for a consequential procurement object.

The strongest thing against it supplier-risk and supplier-onboarding come closest: the UK debarment regime names A MINISTER OF THE CROWN as the party whose decision creates the designation. That is a named authority for one act on one object in one jurisdiction. It is not a stable answerable party for the object across its life, and neither area claims it is.

Reversibility is coded as a first-class property in only one of the ten areas, and in that area it is the area's central finding. In the other nine it is not observable and has never been filled by analogy.

The strongest thing against it Two areas carry PARTIAL reversibility that was found without being sought. contingent-workforce: work orders and revisions can be withdrawn before a supplier takes action. goods-receipt-acceptance: the Reverse Goods Receipt quick action is documented and Joule performs it. Neither area coded reversibility as a field, and both facts surfaced anyway.

The two-register design across ten areas measures the availability of dated evidence, not change in capability.

The strongest thing against it purchase-order-fulfilment, the one area where the gap is partly substantive: the capabilities examined had 0 members at the cutoff and 2 now, and C-003 is not-qualified-at-cutoff because its artifact belongs to a 2026 release rather than because it carries no date.

The disclosure boundary closed measurably between Wave 2 and Wave 3: the same method applied to three fresh areas returned 38 percent fewer dated events against 87 percent more nil routes.

The strongest thing against it Area selection is not held constant. Wave 3's three areas were chosen for structural contrast across the procure-to-pay spine, not for estate richness, and requisition's shortfall is partly a cap artifact — two publishers reached the three-event-per-actor cap with further qualifying artifacts available. A yield comparison across differently-selected waves is not a measurement of the estate.

A machine may be able to perform a binding act only where the grammar of the object supplies a state to set WITHOUT also naming who is answering for setting it.

The strongest thing against it Supplied by the area that filed the hypothesis: 'The Peppol Receipt Advice attaches acceptance to a document state rather than to a named office, and this area found no capability setting it either.'

No causal link is established between any institutional event in the ten-area corpus and the behaviour of any capability, platform, product or buying system.

The strongest thing against it requisition-guided-buying EV-07 and C-012 both name the federal micro-purchase, which is the closest the corpus comes to a link. It establishes that the threshold governing a channel changed; it establishes nothing about how the channel or its operators behaved.