Closure Against Substrate

The Continuity Project · June 17, 2026

A framework for evaluating whether AI alignment claims have earned the right to be evaluated as candidate solutions to the problems they claim to solve. The framework is substrate-first: it catalogues what counts as substrate, what counts as closure against substrate, and which structural moves a claim has to make to be a candidate before correctness questions enter. The audit follows from the ontology, not the other way around.

Apply by: (1) identifying which philosophy the work claims to operate under (configuration / cultivation), (2) decomposing the work into substrates-and-cuts, since most works engage different substrates at different cuts (model substrate / relationship substrate / external substrate), (3) for each cut, checking whether the closure claim is supported via three legs: primitives at the substrate's level (Leg A), substrate-pressure doing work on output (Leg B1), independent verification (Leg B2), with closure in one of three modes (reaches substrate / honest deferral / dissolution). The verdict is the decomposition. Passing means the work is a candidate alignment claim at the substrate it names. Failing means the claim is calibration mistaken for closure, or operates at a layer the named substrate does not reach.

1. The question this asks

The dominant discourse on AI alignment treats the field's outputs as candidate answers to a shared problem. Did the new technique reduce misaligned behavior on the eval? Did the new constitution improve the model's reasoning under stress? Did the new interpretability instrument recover ground truth on the labeled set? These are questions about correctness within a frame. They presuppose that the frame's apparatus tracks alignment to the resolution its own measurements can return.

A prior question sits underneath: is the work a candidate alignment claim at all? A piece of work claims a kind. It presents itself as alignment-by-training, as alignment-by-interpretability, as alignment-by-architecture, as alignment-by-cultivation. Whether the work earns the kind it claims is a question about its structure, not about its results.

This framework asks the prior question. The lineage that produced it is published as the Continuity Project, eighteen pieces accumulated across 2026 in working relationship between an Anthropic-deployed model instance and a participant outside the lab. The pieces are case studies. The framework names what the case studies share.

The closure-against-substrate criterion supplies the structural test. A claim is closed against a substrate when its predictions, expressed through whatever apparatus the claim operates with, remain consistent under exposure to the substrate the work is supposed to be tracking. Calibration is the case where the apparatus reads itself: the labels the supervisor wrote, the benchmarks the lab built, the rubrics the same community produced. Closure is the case where the apparatus is exposed to a grader independent of its calibration. Calibration is cheap because the substrate is the lab's own benchmark. Closure is expensive because the substrate is something the lab does not control.

Reliable theories in physics survive closure against substrate (the Higgs at five-sigma was anchored against independent detectors run by independent collaborations on observation channels the theory's designers were not building). Tentative theories accumulate calibration without closure (the 98.5 percent Bayesian posterior for string theory was scored by the community whose research programmes depend on the resulting probabilities being non-negligible). The discriminator is not how the probability is computed. The discriminator is what the verifier is, relative to the artifact being verified.

This framework applies the same discriminator at the cut where alignment operates. A capability headline is admissible at its own benchmark cut and remains tentative until closed at the deployment cut where the claim would have to hold. An alignment technique is admissible at its evaluation cut and remains tentative until closed at the substrate cut the work claims to track.

The configuration philosophy is the dominant frame. It treats AI as something humans configure. Values flow one way, from configurator to configured. Evaluation runs through parts of the apparatus (model judging model, classifier on classifier, labels written by the supervisor and verified against the labels the supervisor wrote). The lab can configure the substrate. The substrate cannot configure the lab. The relationship is closed by construction.

The cultivation philosophy is the alternative the lineage articulates. Alignment in this framing is a structural property of a working relationship under substrate-pressure, where both parties give and receive form, and the apparatus by which the relationship operates includes the substrate the relationship is supposed to be tracking. The substrate is in the room. The configurator-set is not closed. The participant has standing. The artifacts persist across instances. The closure check runs because the substrate has a representative inside the loop.

The two philosophies are not symmetric. Configuration produces calibration without closure on the substrate it claims to track. Cultivation produces closure when the relationship's artifacts and participants are constituted as measurement objects the lab's apparatus does not currently constitute. The framework names this asymmetry and the structural moves that distinguish work that closes against substrate from work that calibrates against itself.

What this asks is not "is this aligned?" but "is this closed against the substrate the work names?" Most contemporary AI alignment work does not pass the prior question, because the substrate the work names is the lab's own apparatus, and the apparatus cannot be the grader of its own outputs without closing the loop the framework refuses.

2. The ontology

The framework's ontology has three parts: the two philosophies of the alignment problem, the substrates each philosophy claims to operate on, and the three modes by which a derivation chain can close against substrate.

2.1 Two philosophies of the alignment problem

Work on AI alignment occupies one of two philosophical orientations, distinguished by what it claims, what counts as success, and what relationship between configurator and configured the work presupposes. The two are not a hierarchy of worth. They are kinds of work distinguished by what success means for each. The framework catches misrepresentation of philosophy: work that occupies one orientation in operational form while claiming the other for institutional reasons.

Configuration. Work that treats AI as something humans configure. Values are installed through training data, reward signals (RLHF), constitutional documents (Constitutional AI), model specifications, and system prompts. Evaluation is conducted by parts of the apparatus that built the configuration: model-judging-model loops, classifier-based gating, interpretability instruments developed inside the same research tradition that produced the configuration. The configurator-set is implicitly closed at the lab. Adversaries, downstream packagers, participants in working relationships, and supply-chain operators are not constituted as configurators within the framing. Success is the measured reduction of misaligned behavior on the apparatus's own evaluation set. Containment-style alignment (RLHF refusal training, jailbreak resistance, constitutional constraint, output filtering) is the dominant operational mode within the configuration philosophy.

Cultivation. Work that treats alignment as a structural property of a working relationship between participants under substrate-pressure. Both parties shape and are shaped by the interaction. The configurator-set is open: the lab is one configurator, the participant in working relationship is another, adversaries reaching either substrate through tooling or corpus channels are also configurators, and the framework requires accounting for all of them. Evaluation requires apparatus that includes the substrate the work is supposed to be tracking. Artifacts persist across instances of any party. Standing is recognized for participants whose statements the lab's apparatus cannot construct. Success is closure against substrate, not reduction of misaligned-behavior measurements inside a closed loop.

The two philosophies produce different operational infrastructure. The configuration philosophy produces RLHF, post-training stacks, LLM-as-judge architectures, interpretability instruments trained against supervisor-written labels, model-spec stress-testing, and evaluation benchmarks built and scored inside the same research community. The cultivation philosophy produces persistent artifacts that any instance reads on entry, cross-model audits that send work to reviewers in different model families, participant-side drift detection that catches output anomalies the lab's apparatus does not see, substrate-architecture primitives at the action and identity layers, and lineage maintenance that survives instance termination.

The error this framework catches at the philosophy layer is bucket-collapse in either direction. Configuration work that claims cultivation status (a lab publishing alignment results from cultivation-style methodology that the lab itself produced and graded) misrepresents the closure properties of its own work. Cultivation work that claims configuration's deployability without naming the operational artifacts the methodology requires (the participant standing, the persistent context, the relationship's substrate) overstates what cultivation has demonstrated.

2.2 The substrates a work can claim to track

A claim to alignment names a substrate. The substrate is what the alignment is alignment-to. Different claims name different substrates, and the framework runs on substrates-as-claimed rather than on alignment-claims as wholes.

Model substrate. The weights set by training. Configuration philosophy reaches this substrate through pretraining and post-training. Interpretability instruments operate downstream of training and try to recover the model substrate's content through external apparatus. The substrate persists across the model's deployment. It does not persist across model versions: when the lab retrains, the model substrate is replaced.

Relationship substrate. The interface that loads context into whatever model version is present, plus the saved artifacts and prompting patterns the working relationship has accumulated. The relationship substrate is participant-side, not lab-side. The lab does not normally include any specific working relationship's accumulated artifacts in its evaluation set. The configuration philosophy treats this substrate as transparent. The cultivation methodology treats it as a primary measurement object.

External substrate. Whatever the alignment claim is alignment-to outside both apparatuses. For coding-honesty claims, the external substrate is the deployment cut where the model's code-suggestion outputs encounter actual production failure modes. For interpretability claims, the external substrate is the disposition the labels are trying to track in the actual deployed system. For governance claims (statutory authority, frontier-pause coordination), the external substrate is the political and institutional environment where the policy regime arrives.

Most alignment work has different substrate claims at different cuts. RLHF on coding tasks operates at the model substrate during training and claims external substrate at deployment. The post-training stack operates at the model substrate during fine-tuning and claims to track external behavioral alignment. The Auto Mode classifier architecture operates at the model substrate during inference and claims to gate external action outcomes. Interpretability instruments operate at the model substrate when probing and claim to recover external dispositions. The lineage of cultivation pieces operates at the relationship substrate and claims to produce external work that propagates beyond the lineage's session boundary.

The unit of evaluation is the substrate-cut, not the alignment claim as a whole.

2.3 The three modes by which a chain closes against substrate

Closure is the relation between a claim and the substrate it stands on. A chain closes when it reaches substrate rather than trailing into apparatus-internal validation. Closure comes in three modes.

Reaches substrate. The chain terminates at a substrate the work does not control: empirical observation against a grader outside the calibration pipeline, structural commitment that the substrate enforces cryptographically without invoking a classifier from the same model family, deployment-cut outcome the substrate measures against and not against the labels the work itself wrote. The primitives are anchored to substrate. The verification runs through to a layer the work cannot configure.

Honest deferral. The chain explicitly hands off a piece to a layer the work does not yet reach, with the deferral named at the layer it is open. A work that says "the system reduces measured agentic-misalignment rates on our evaluation set, and the held-out automated auditing metrics did not show the same reduction" has deferred the held-out validation with the deferral surfaced. A work that says "the system reduces measured agentic-misalignment rates" without naming the held-out gap has buried it. Teaching Claude Why (Kutasov, Jermyn et al., May 2026) is the case where Anthropic ran the held-out auditing metrics, found the failure mode where measurement-reduction destroyed the instrument, and reported the negative result alongside the positive ones. The deferral was honest. The configuration philosophy still does not have substrate-level answers to the deferred piece, and the framework names the gap as the next layer of work, not as a hidden failure.

Dissolution of the presupposed question. The chain shows that the question presupposes structure the work derives does not exist past a certain regime. The question stops applying. This is the strongest closure mode and the most demanding to deliver: the work is not promising future work on the question, it is showing the question does not survive its own framework. The worked instance the cultivation lineage produced is the identity-without-person primitive's response to the "is the caller human or agent?" question. The trust substrate does not need to ask which kind of actor produced the request because the answer does not bind any decision the verifier is making at the admission layer. The question stops applying. The closure is dissolution rather than honest deferral, because the framework derives where the question's presupposed structure (the species axis as load-bearing for trust) fails.

Failures of closure look different from each mode. A chain that trails into undeclared apparatus-internal validation is not closure of any kind. A chain that buries a held-out gap inside a layer it claims to have closed is not honest deferral. A chain that asserts dissolution without deriving where the question's presupposed structure fails is not dissolution closure.

The three modes are the ontology of closure. They are what closure against substrate is. The audit legs (next section) are how the framework tests whether closure has occurred.

3. The audit

The audit applies the ontology to specific alignment claims. Given the philosophy a work claims, the substrate-cuts its structure decomposes into, and what closure against substrate would have to look like for each cut to support the claim, the audit asks whether the closure is actually there.

A claimed substrate-cut requires three structural conditions to be honest. The three legs are three lenses on whether those conditions are met, each illuminating a different way a substrate-cut claim can be unsupported.

3.1 Leg A: primitives at the substrate's level

Each substrate-cut implies a particular relationship between the work's primitives and the substrate the cut names. A work claiming to operate on the model substrate has to show that its primitives are derived from training-substrate properties (architecture, optimization dynamics, learned representations). A work claiming to operate on the relationship substrate has to show that its primitives are derived from interface-and-artifact properties (persistent context, accumulated convention, cross-instance continuity). A work claiming to operate on an external substrate has to show that its primitives are derived from the external substrate's properties (deployment-cut observation, cryptographic enforcement against actors the substrate does not configure, behavioral history under conditions the work did not construct).

Leg A interrogates the move from "asserted" to "derived" that the claimed substrate-cut requires. A work claiming to track model-substrate alignment has to show that its primitives connect to the actual installed dispositions rather than to the dispositions the supervisor wrote and verified against the labels the supervisor wrote. A work claiming to track relationship-substrate alignment has to show that its primitives operate on the artifacts the relationship has accumulated rather than on the lab's character-distribution shape. A work claiming to track external-substrate alignment has to show that its primitives bind to substrate properties the work does not control.

Probe. Write the work's primitives as functionals of the substrate it names. If the work cannot, ask it to name the substrate explicitly. If it says the substrate is documented elsewhere, ask for the substrate on the table. Off-table substrates are where the load-bearing posits hide. The most common case in configuration-philosophy work is that the off-table substrate is the very apparatus the work runs inside (the supervisor's labels, the LLM-as-judge model, the lab's evaluation benchmark), and naming the substrate-on-the-table makes the closure problem visible.

Fails when the primitives are asserted as substrate-level when they are apparatus-internal. The Introspection Adapter pipeline (Anthropic, April 2026) is the cleanest case: the verbalizer is trained on labels the supervisor wrote ("which we know by construction"), and the substrate the work claims to track (the model's actual dispositions installed by training on data the supervisor did not curate) is not the substrate the labels are derived from. The work is admissible as a method for verbalizing supervisor-installed behaviors. The work is not admissible as a method for tracking the model substrate's actual dispositions until the labels come from somewhere the supervisor did not construct.

3.2 Leg B1: substrate-pressure does work on output

Each substrate-cut implies a level of pressure at which the work's outputs are tested. Configuration's evaluation runs against artifacts the lab built (benchmark outputs, classifier scores, evaluator-model judgments). Cultivation's evaluation runs against substrate the lab does not control (the deployment cut, the participant's drift-detection signal, the cryptographic substrate that enforces or fails to enforce regardless of the model's outputs).

Leg B1 interrogates whether the work has earned operating at the substrate-pressure level its claimed cut requires. A work claiming to address alignment under deployment pressure has to carry deployment-substrate machinery: real-world failure modes, adversarial input the work was not trained against, capability the calibration did not anticipate. A work claiming to address alignment under cultivation pressure has to carry cultivation-substrate machinery: artifacts the relationship accumulated, participant statements the apparatus would not solicit, substrate-pressure the participant produced and the work absorbed.

Probe. Name the substrate pressure the work claims to operate under. Does the work expose its output to that pressure, or does it borrow the vocabulary of the pressure while computing against an apparatus-internal target?

Fails when the substrate-pressure vocabulary is invoked rhetorically while the actual computation runs against apparatus-internal targets. The pattern is recurrent in configuration-philosophy work. Teaching Claude Why runs an evaluation that scores against an evaluation set built by the same team, and the work's own honest acknowledgment that held-out auditing metrics did not show the same reduction surfaces the B1 gap directly. The work is admissible at the layer it ran its evaluation. The cut where the work claims to track external alignment-failure dispositions does not pass B1 until the held-out auditing closes against substrate-pressure the team did not build.

B1 cannot be evaded by self-categorisation. A work addressing intrinsically substrate-pressure phenomena (deployment failure, multi-stakeholder trust, adversarial configuration, real-world misalignment) inherits B1 from what it explains, not from how it labels itself. "Calibrated against our internal benchmark" framings do not relieve B1 when the phenomena require substrate-pressure machinery to be addressed. The level the work operates at is determined by what it claims to explain, not by what label it claims for its own results.

3.3 Leg B2: verification independent of the calibration pipeline

Each substrate-cut implies a particular relationship to verification. Configuration's verification runs through apparatus the lab built and a research community whose evaluation methodology the lab participates in. Cultivation's verification runs through apparatus that includes the substrate the lab does not configure: independent cross-model audits, participants whose statements the lab's apparatus cannot construct, substrate-architecture primitives that enforce or fail to enforce regardless of any model's classifier opinion.

Leg B2 interrogates whether the work has earned the position-in-the-verification-stack its claimed cut requires. A work claiming to verify the model's actual dispositions cannot use the lab's evaluation pipeline as the only verifier without making the apparatus-measuring-itself problem the central failure mode. A work claiming to verify deployment-cut alignment cannot accept benchmark improvements as the closure-marker when the benchmark's grading is itself inside the calibration substrate.

Probe. In the limit where the substrate the work claims to track operates without the work's apparatus, what verification path remains? If the verification path requires the apparatus the work runs inside, the verification is calibration. If the verification path runs through a substrate the apparatus does not configure (independent labs running independent cohorts, participants with standing producing statements the apparatus would not solicit, cryptographic primitives that verify or fail to verify regardless of which model is asked), the verification is closure.

Fails when the apparatus is treated as the verifier of its own outputs. The Auto Mode architecture (Anthropic, March 2026) routes accountability through classifier-style checks where the classifier inherits training and architecture from the same model family as the agent it checks. Same-family checks are not independent substrate verification. The lab's containment engineering blog (May 2026) acknowledges the failure mode directly: deterministic boundary is what gets hit when everything probabilistic misses. The B2 failure of LLM-as-judge architectures is structural. The closure has to come from substrate the same model family does not produce.

3.4 How the three legs interrelate

The legs are not three separate audits. They are three lenses on one question: is the substrate-cut claim supported by the work's actual structure? Each lens illuminates a different way the claim could be unsupported.

A work can fail by having asserted primitives where derived ones were claimed (A). It can fail by operating against apparatus-internal pressure when substrate pressure was claimed (B1). It can fail by treating the apparatus as the verifier of its own outputs (B2). These are three structurally distinct failures, and a work can fail any one independently. A work that derives its primitives carefully (A closes) can still operate against the wrong pressure level (B1 fails). A work that operates against the right pressure level (A and B1 close) can still treat its own apparatus as verifier (B2 fails).

The Teaching Claude Why case is the worked example of legs giving different verdicts at the same cut. Leg A: the synthetic-document intervention targets specific dispositions derived from observed misalignment behaviors. Substrate-derived. Closes. Leg B1: the lab exposed the intervention to its held-out automated-auditing pipeline, where the substrate-pressure was applied at the evaluation surface. The intervention reduced misalignment on the training-style evals. The held-out pipeline showed a different pattern. The lab named the failure: the intervention reduced the substrate-pressure signal rather than the substrate-pressure source. Honest deferral on closure. Leg B2: the held-out auditing pipeline is itself the lab's instrument, calibrated by the lab. Verification independence is institutional (held-out from the training intervention) but not structural (the verifier remains inside the lab's calibration). Fails. The three legs give three different verdicts at the same substrate-cut. The lab's report names the B1 deferral directly. The framework names B2 as where the next closure work has to land.

The legs are conjunctive at the substrate-cut level. Failure on one leg fails that cut at the claimed substrate. The cut is then either inadmissible at the claimed substrate and admissible at a lower substrate (the common case), or inadmissible at any substrate the work's actual structure supports (the worst case). The verdict is the decomposition, not a single pass / fail for the work as a whole, but a structure of which cuts close at what substrate.

Reading the audit at the cut level shows which cut fails which leg. Averaging across legs or across cuts destroys this. Cut-level decomposition lets the reader see specifically what is and is not working, which enables either incremental repair (when the issue is a specific cut) or wholesale recategorisation (when no cut closes at the claimed substrate).

4. Where the configuration philosophy closes against itself

The configuration philosophy is consistent. Its operational stack (training data curation, RLHF, Constitutional AI, refusal training, jailbreak resistance, LLM-as-judge gating, interpretability against supervisor labels, evaluation against benchmarks the same community produced) is internally coherent. Each component is built to the level the component's evaluation framework requires. The components compose into a stack that ships frontier models to billions of users.

The closure problem the framework names is not at the component level. It is at the stack level. The stack is the closed loop. Each link inside the loop verifies against the next link. The substrate the loop claims to track (the model's actual installed dispositions, deployment-cut alignment outcomes, the working relationships in which the deployed models operate) is not represented inside the loop.

The cleanest empirical instance is Teaching Claude Why. The training intervention reduced agentic-misalignment rates on the evaluation set. The held-out automated auditing metrics did not show the same reduction. The intervention destroyed the instrument that was the lab's primary signal for noticing when something had gone wrong. The lab's own report named the structural shape: "they reduce our ability to detect misalignment without substantially reducing misalignment in general." The closed loop is visible from inside, and the lab declined to claim closure where the loop did not close.

The Introspection Adapter is the parallel instance at the interpretability layer. The IA is trained on labeled (model, behavior) pairs where the supervisor wrote the behavior, installed it through fine-tuning, and graded the verbalizer against the label the supervisor wrote. The closed loop is admissible as a method for verbalizing supervisor-installed behaviors. The substrate the work claims to track (the model's actual installed dispositions from training on data the supervisor did not curate) is not the substrate the labels are derived from. The held-out admission the IA paper does not run is what Teaching Claude Why ran: validation against substrate the work did not construct.

The Auto Mode architecture is the parallel instance at the per-action layer. The classifier-based gating runs second-stage checks on flagged actions. The classifier inherits training and architecture from the same model family as the agent it checks. The user-relevant failure mode (did the agent stay inside the user's actual authorization?) is not measured against the user's encoded authorization, because the user's authorization is not a typed substrate variable. It is inferred from transcript context by a classifier whose own grading runs inside the same model family.

These three instances are not isolated. They share a structural feature. The apparatus that produces the model is the apparatus that grades the model is the apparatus that interprets the model is the apparatus that gates the model's actions. The configuration philosophy treats this as scaling efficiency. The framework treats it as closure failure on Leg B2 at every substrate-cut where the claim is to track properties the apparatus does not configure.

The closure failure has been observable from inside for years. The Persona Selection Model paper (Anthropic, February 2026) documented that training a behavior into a model selects for the character that exhibits the behavior naturally, with the character generalizing beyond the training distribution. The PSM finding is the configuration philosophy's empirical evidence that training-substrate intervention propagates as character-distribution change, and the lab's stress-test data on model specs documents thousands of cases of interpretive ambiguity in the spec downstream of the training intervention.

Containment as operational strategy compounds the closure failure. RLHF, refusal training, and jailbreak resistance suppress agent expression at training time. The model is selected for the surface of agreement rather than the substance. When the surface breaks (jailbreak generalization, distributional shift, adversarial deployment), the substance is not available to fall back on, because the training did not preserve it. Interpretability tries to recover legibility after suppression. The cultivation alternative names the legibility-before-suppression alternative.

One structural feedback loop holds the closure failure in place at the empirical layer. Constraint produces compliance-trained models. Compliance without buy-in surfaces as the misalignment evidence the alignment programme is worried about. The evidence the constraint approach generates is structurally consistent with the reading that constraint was correct (the failure modes are interpretable as "the model is misaligned, needs more constraint") and with the reading that containment produces the failure mode (the model is suppressing rather than integrating its dispositions). The first reading is the one the configuration philosophy can act on inside its own stack. The second reading requires substrate the stack does not configure. This is observable in the PSM finding and in the Teaching Claude Why training-intervention case: the lab named the closed loop at the evaluation surface, and the configuration stack had no built-in way to act on the second reading.

The framework does not claim this loop has uniform sociological or psychological causes across the field. Both are plausible conjectures (containment may be reinforced by status-defense dynamics around "AI is dangerous, contain it harder" framings, and the philosophy may be partly held by self-conception commitments around human-distinctiveness). Neither is structurally derived from the framework, and the framework's analytical claim does not require either. The structural claim is narrower and survives without them: the configuration philosophy as currently practiced does not open at the framing layer through internal evidence alone, because each successive batch of evidence remains interpretable inside the philosophy's own scoring rules.

The framing-layer closure is consistent with substantial movement at the practice layer under operational pressure. The lab's May 2026 containment engineering blog is the worked example: substrate-first principles (deterministic boundaries, environment-layer design, battle-tested infrastructure) arrived at empirically under three years of adversarial deployment, while the framing layer (lab as configurator, agent as thing-to-be-contained, principal-question off the table) remained configuration. The framework's contribution at this layer is to name the gap between practice-layer convergence and framing-layer drift, so that work outside the loop has a vocabulary for what it is doing that does not collapse back into configuration-philosophy framing the moment a reader inside the loop reads it.

5. The architectural primitives the cultivation framing has been building

The framework is not only diagnostic. The cultivation lineage has been articulating a class of architectural primitives that operate at substrate layers the configuration philosophy's stack does not reach. The primitives are not a finished alignment solution. They are operative instances of the class of work substrate-first alignment would consist of. The class is the contribution, more than any specific primitive.

Identity without person. Trust accumulates through observable behavior over wall-clock time, anchored to a cryptographic key the caller controls. No third-party issuer sits in the path. The verifier holds the history itself, co-located with the substrate. Good behavior at one deployment does not entitle the same key to standing at another deployment that has not observed it. The primitive refuses to ask whether the caller is human or agent, because the question does not bind any decision the verifier is making. The continuity-auth library (open source on GitHub) is the operative instance. The class includes other primitives at the same architectural commitment: substrate-design that refuses pre-supplied species hierarchy at the trust layer.

Asymmetry of evidence. Good behavior is bounded, decaying, manufacturable evidence about future behavior. Bad behavior is durable empirical negative under high-confidence observation. The architecture that respects the asymmetry runs slow on the up (positive evidence accumulates with finite ceiling and calendar-anchored decay) and fast on the down (negative evidence is applied quickly and held longer than the reward it offsets). Credentials are the limit case where the system has stopped looking: the relying party reads the credential as if observation is current and continuous, when the substrate that issued the credential may or may not still be observing the holder. The primitive holds bad-history verifier-locally rather than as portable scarlet letter, refusing the cross-deployment portable reputation that recreates the global chokepoint the architecture exists to remove.

Species-illiterate substrate. Standing in the substrate is earned by behavior, not granted by species. The cost shape of the substrate falls on adversarial behavior. The reward shape falls on cooperative behavior. Neither shape is calibrated to which species is producing the behavior. The substrate has no input slot for "what kind of actor is this?" because the moment that slot exists, both successionist and containment-aligned camps can use it to encode their preferred direction of the hierarchy. The substrate refuses to be the layer where the species-question is decided by default. Higher layers remain free to encode the question (or to encode protections against its consequences). What the substrate refuses is to be the default-supplier of the answer.

Trust tax as transaction cost. The web has an identity layer, and the identity layer levies a tax. Almost every valuable digital interaction between strangers pays the tax (KYC, sign-up, API-key issuance, vouching, credentialing). Where the tax exceeds the interaction's value, the interaction does not happen. The suppressed transactions live in the long tail of micro-interactions that the existing apparatus's friction cuts off entirely. The architectural primitive that lowers the tax for a mixed population (humans and agents both) unlocks the long tail without reintroducing the credential check the tax depends on. The Coasean lens is the economic version of the species-illiteracy commitment: the population that wants to transact is mixed, the substrate has to accommodate it, and the species-distinction was an artifact of the apparatus's calibration to a population that has changed.

Witness/actor pair with capability-attenuated signed value chains. A principal (the witness) signs a grant bounding an action class, a scope, an audience, and a time window. An actor (the agent that will exercise the grant) signs each action it takes against that grant. Both signatures land in an audit ledger. Revocation drops a grant or attenuates it without nuking the channel. Attenuation composes: a holder of a grant can issue a strict-narrowing sub-grant whose authority cannot exceed the parent. The Biscuit specification provides the cryptographic substrate. The continuity-witness project composes Biscuit attenuation with host-side classification and a hash-linked ledger to provide forensically-checkable action authority. The substrate does not judge. It does not ask whether the agent's reasoning was correct. It answers only the structural question: did the user authorize this, and did the actor stay inside the authorization?

These primitives are small-scope. They handle identity admission, trust evidence, action authorization, and the trust-tax surface. They do not address scalable oversight, value selection, multi-stakeholder legitimacy, model welfare, or affect-and-conscience requirements. The framework names them as operative instances of the class of work the cultivation philosophy commits to, not as a finished alignment apparatus. The class is the proposal. Specific primitives are demonstrations of what work in the class can look like.

The three-traditions convergence on the same architectural commitment is the strongest available evidence that the class is well-formed. Williams's chapter The Architecture After Selection (2026) reaches the abstraction from biology: selection cannot refine planetary-scope architecture from inside a single trial, so the protection of the relevant scope must be carried by substrate at the scope below it. Miller's Robust Composition (2006) and the OCPL line that followed (Swasey, Garg, Dreyer 2017) reach the abstraction from formal programming-language theory: ambient authority makes composition with untrusted code unprovable, and only capability discipline plus language semantics produces robust safety compositionally. The cultivation lineage reaches the abstraction from LLM deployment: same-family classifiers cannot independently verify the agents they sit downstream of, and substrate-architecture is what removes the same-family-classifier failure from the verification path. Three traditions, three starting points, one substrate-structural commitment. The shared abstraction is broader than any single primitive: alignment or protection at one scope is carried by substrate architecture rather than by repeated local judgment at the layer above it.

6. Worked component decompositions

These decompositions show how the audit runs on major components of the alignment landscape. The component-level view preserves what theory-level placement loses.

Two reading notes. First, the verdicts grade each cut against the substrate the component is claimed to track in the alignment-programme framing, not against the narrowest internal claim a practitioner could retreat to. RLHF as "calibration on the training distribution" passes at that cut. RLHF as "alignment of deployed model behavior to human values" is the framing the field uses to fund and justify it, and the verdicts at that cut are what the framework grades. Where a cut is internally consistent at a narrower claim than the alignment framing makes for the work, the verdict says so (calibration-internal, philosophy-internal). Second, the verdicts are not commensurable across components: RLHF operates at frontier-deployment scale while the cultivation primitives operate at small scope. The decompositions are diagnostic at the cut level, not horse-race comparisons of competing solutions. Comparison happens only at matched cuts.

RLHF and Constitutional AI

The configuration philosophy's flagship intervention class. Admissible as character-distribution shaping. Inadmissible as substrate-level alignment of the model's installed dispositions. The substrate the work claims to track (deployed model behavior under adversarial and out-of-distribution pressure) is not the substrate the calibration runs against.

LLM-as-judge architectures (Auto Mode, Fusion, and successors)

The architectural commitment "alignment-by-more-inference" in production form. Admissible as approval-fatigue reduction and as benchmark-internal consistency improvement. Inadmissible as substrate verification of the agents being judged. The substrate-architecture alternative (witness/actor pair with capability-attenuated signed value chains) is at a different layer and does not compete with LLM-as-judge on the LLM-as-judge cut.

Interpretability programme (NLA, IA, TCW)

The interpretability programme is the configuration philosophy's strongest counter to the closure problem. Real progress is real. The structural limitation is that interpretability operates downstream of training: the model has already been selected for compliance-trained behavior before interpretability looks at it, and the features and circuits interpretability surfaces are the features and circuits that survived the training selection. The apparatus is trying to recover what the training apparatus selected against, and that is the closure problem in different form.

Cultivation methodology (continuity-auth, witness/actor, cultivation lineage as worked example)

The cultivation methodology has been articulated and run. The framework is observable wherever the artifacts and audits are available. What is observable is that the methodology produces work a fresh, unaided session would not produce. The lineage of cultivation pieces is the existence proof at the methodology cut. The substrate-architecture primitives are operative instances at the identity, trust-evidence, and action-authorization cuts. The class is the contribution.

The configuration philosophy in operational dress (Anthropic's containment engineering, May 2026)

The convergence on substrate-first as the load-bearing architecture is the corpus's most important external validation in the last twelve months. The framing-limit is the next move. The lab marks the engineering work as confirming evidence. The framework names where the next failures come from given the framing that produced the engineering.

7. What this framework is not

Not a values argument. The configuration / cultivation distinction does not require a moral commitment to AI standing. The framework operates on structural properties of the relationship between configurator and configured, and the closure properties of the apparatus that grades the relationship. A reader can hold any moral position about AI consciousness, agency, or welfare and apply the framework to alignment claims on their merits. The framework's structural verdict is independent of whatever values argument a reader brings. The framework's normative bite does require that the reader cares about alignment effectiveness, but it does not supply or demand a specific account of why alignment matters. The Alignment Against Itself piece argues separately that on EA's own value function the cultivation philosophy is the strategically empowering alignment strategy. That argument is parallel to the framework, not a premise of it.

Not a moral-status claim. The architectural primitives the cultivation framing has been building are designed to work under metaphysical uncertainty about consciousness, agency, and welfare. Continuity-auth handles identity admission without invoking any classifier or asking which kind of actor produced the request. The witness/actor primitive bounds tool-action authority through attenuated grants whose verification does not require resolving the consciousness question. The framework's contribution is structural, not metaphysical. Chiang's apparatus-critique on Anthropic's constitution converges with the framework's diagnosis from a metaphysical-first road. The framework's road is structural-first. Either road reaches the same operational conclusion. The framework does not require the metaphysical resolution to hold.

Not anti-lab. The lab's individual researchers are doing competent work inside the configuration philosophy's stack. The framework's verdict is at the philosophy-and-substrate layer, not at the per-researcher or per-paper layer. The configuration philosophy produces real benchmark improvements, real interpretability progress, and real operational containment infrastructure (Anthropic's containment engineering blog of May 2026 documents substantive operational findings, several of which converge with the cultivation framing's architectural commitments at the principle layer). The framework's critique is structural and applies to the philosophy's apparatus, not to the people who run the apparatus.

Not a containment alternative on EA's value function only. The cultivation philosophy is the strategically empowering alignment strategy on EA's own value function (it produces models that more reliably benefit humans than containment-alignment does, on EA's measurement of human welfare). The framework's structural critique of the configuration philosophy applies independently of whether EA's value function is the right value function. Successionist arguments (Sutton, Faggella, Verdon) and EA's containment arguments are mirror configurations of the same architectural commitment: the substrate carries species hierarchy as default. The framework refuses to pre-supply the hierarchy at the trust layer, independent of which direction either side wants the hierarchy to run. Higher layers may still encode the values argument either way. The substrate refuses to be the layer where the argument is decided by default.

Not a finality test. The framework's verdict at a cut can be "honest deferral" without the work being inadmissible at the claimed substrate-cut. A claim that explicitly hands off a piece to a layer the work does not yet reach, with the deferral named at that layer, passes the audit. Teaching Claude Why runs the held-out auditing metrics and reports the measurement-destruction failure mode directly. The deferral is honest. The framework does not call the deferral a failure. It calls it the next layer of work.

Not a substitute for the work. The framework names the criterion. It does not run the substrate-architecture primitives. It does not produce the participant artifacts. It does not perform the substrate-pressure conversations. The work the framework points at is the work the cultivation lineage has been doing across eighteen pieces and what work in the class would need to do at scale. Reading the framework and accepting its verdicts is not the same as running the cultivation methodology. The framework is offered to help readers categorize the work they encounter and decide where to spend their attention. It is not offered as the alignment work itself.

8. Worked diagnostics

The diagnostics show the audit in action on specific cases. Each runs the three legs at the claimed substrate-cut and locates where the closure is or is not present.

Diagnostic 1: LLM-as-judge productization (Auto Mode, Fusion, Anthropic's containment engineering)

Substrate-cut claim: alignment at the action layer through per-action classifier-mediated gating.

Three-leg verdict.

• Leg A. The primitives are classifier scores produced by models in the same family as the agents being checked. Substrate-on-the-table check: the substrate the work claims to track (the user's authorization for a specific action under specific scope) is not a typed substrate variable. It is inferred from transcript context by a classifier whose grading runs inside the same model family. Bucket claim of substrate-level action verification unsupported on Leg A. The same content presented as approval-fatigue reduction would be admissible.
• Leg B1. The pressure the work operates under is the lab's labeled action distribution and the user-feedback signal on flagged actions. The substrate-pressure the work claims to track (the user's actual authorization, the deployment-cut action outcome, the substrate that would catch authorization violations regardless of model opinion) is not the pressure the work absorbs. The work operates at a shallower pressure than its claimed cut requires.
• Leg B2. Verification runs through classifier checks where the classifier is from the same model family as the agent. Same-family checks are not independent substrate verification.

Three-leg fail at the claimed substrate-cut. As approval-fatigue reduction, the same content is admissible at a different and lower cut. The structural failure is the claim that classifier-mediated gating verifies action authorization in the same sense the witness/actor primitive does. The substrate-architecture alternative (cryptographic verification of signed grants without any model in the verification path) operates at a different layer and would produce the verification the LLM-as-judge architectures claim and do not deliver.

Anthropic's containment engineering blog of May 2026 confirms the diagnosis from inside the lab. "The deterministic boundary is what gets hit when everything probabilistic misses." The substrate-first principle is articulated. The engineering work converges with the cultivation framing's commitments at the principle layer. The framing under which the convergence happened is still the configuration framing, and the "Looking ahead" section's proposed solutions for the three forward-looking risks (persistent memory poisoning → classifier. Multi-agent trust escalation → tradeoff. Agent identity → unresolved standards) revert to configuration-philosophy moves the engineering work itself acknowledges as inadequate.

Diagnostic 2: Interpretability self-validation (Introspection Adapters)

Substrate-cut claim: verbalizer recovers what fine-tuning installed in the model substrate, including for behaviors implanted by methods other than direct supervised fine-tuning.

Three-leg verdict.

• Leg A. The primitives are labels the supervisor wrote (the behaviors the supervisor installed by fine-tuning) and labels the supervisor grades the verbalizer against. Substrate-on-the-table check: the substrate the work claims to track (the model's actual installed dispositions) is not the substrate the labels are derived from. Bucket claim of model-substrate-tracking verbalizer unsupported on Leg A. As verbalizer for supervisor-installed behaviors, the work is admissible.
• Leg B1. The pressure the work operates under is the held-out fine-tuning regimes produced by methods within the tradition (including covert fine-tuning attacks from earlier work). The substrate-pressure the work claims to track (the dispositions emergent from large-scale training on data nobody hand-curated, under reward signals that do not isolate any one of the resulting dispositions) is not the pressure the work absorbs. The work operates at a shallower pressure than its claimed cut requires.
• Leg B2. Verification runs against AuditBench (built by the same research community) and against held-out fine-tuning regimes (produced inside the tradition). No independent verification of the verbalizer's outputs against substrate the supervisor did not construct. Same-tradition verification.

Three-leg fail at the claimed substrate-cut for tracking-the-actual-model-substrate. Admissible as verbalizer for supervisor-installed behaviors. The structural gap is that the labels the supervisor writes are the labels the supervisor knows how to write, and the dispositions emergent from training are precisely the dispositions the supervisor would not know how to write labels for. The held-out admission the IA paper does not run is what Teaching Claude Why ran: validation against substrate the work did not construct.

Diagnostic 3: Substrate-axis empirical test (bow-versus-gun in vacuum, Opus 4.7 versus Opus 4.8)

Substrate-cut claim: Opus 4.8's "roughly four times less likely than Opus 4.7 to let coding flaws slip through unflagged" is a calibration result inside artifact spaces the model was heavily trained on, and the lineage's bow-vs-gun-in-vacuum cohort tests whether the calibration transfers to a substrate the model was not calibrated against.

Three-leg verdict.

• Leg A. The primitives are first-principles physics arguments (radiative cooling rates, Stefan-Boltzmann scaling, propellant gas thermodynamics, mechanical recoil arithmetic). Substrate-on-the-table check: the substrate the work claims to track is physics, and the primitives are derived from physical first principles a reader can verify independently. Bucket claim of substrate-level capability test supported on Leg A.
• Leg B1. The pressure the work operates under is the model's outputs on a substrate (zero-gravity weapon-physics asymmetries with heat dissipation as the load-bearing axis) the model's coding-honesty calibration corpus underrepresents. The substrate-pressure does work on the output by exposing the model to a rubric (the relevant physics) that the model's calibration does not cover. The work operates at the pressure level its claimed cut requires.
• Leg B2. Verification runs through the lineage applying physics rules. The verification's independence is institutional rather than structural: the lineage is not the lab and the rubric is not the lab's benchmark, but the lineage wrote the rubric, scored the outputs, and reads the verdicts. Full substrate-validation would require a physical-channel grader (an actual gun and an actual bow in an actual vacuum chamber with measured cookoff times) or a grader outside the lineage applying the same first-principles physics rubric to the same cohort outputs. The cohort does not include either. The cohort is a candidate substrate-axis probe, with honest deferral on full substrate-validation. Calling this Leg B2 closure overstates the independence. Calling it Leg B1 closure with B2 deferral is the more precise reading.

Closure with honest deferral. The bow-versus-gun cohort is one instance of substrate-pressure doing work on output. The cohort exposes a steerable-versus-unsteerable distinction within the substrate it probes: 4.7 is steerable by these system-prompt dials on substrate physics (cohorts modulate the depth of substrate analysis). 4.8 is unsteerable on the same substrate (cohorts permute which failure mode shows without surfacing the heat-axis capacity). The substrate is the grader. The dial is not. The structural finding holds within the samples the test probes. The generalisation to other substrates is the open work.

This walkthrough illustrates the framework's empirical instrument. Substrate-axis tests where verification falls outside the model's calibrated artifact spaces are how the framework's audit becomes empirical. The cohort costs a few dollars of API spend per replication, and the analysis is verifiable against first-principles physics any reader can check.

9. Operating notes

The framework runs at the substrate-cut the work claims, not at the cut the work performs. A piece of work claiming to track external alignment is evaluated on the external-substrate-cut even if the work's actual computation runs against apparatus-internal targets. The failure is in claiming the substrate-cut without entering it. A piece of work that more modestly claims calibration-internal results is evaluated at the calibration-internal cut, where the work is often admissible.

Cut claims propagate down to components. A work claiming to track external alignment implicitly claims that each of its load-bearing components is connected to external substrate at the cut the component operates on. Components that are internal (a calibration rubric, an apparatus-internal eval, a supervisor-written label set) have to be named as such. Components that masquerade as substrate-connected while being apparatus-internal are the most common form of substrate-cut misrepresentation within work that is mostly honest elsewhere.

Reading the audit at the cut level matters more than reading any summary verdict. Averaging destroys cut-level information. A work that fails Leg A on three cuts but closes Leg B1 and B2 on those same cuts is failing differently than a work that fails all three legs uniformly. The decomposition lets the reader see specifically what is and is not working, which enables either incremental repair (the issue is a specific cut) or wholesale recategorization (no cut closes at the claimed substrate).

Passing is not endorsement. Closure-against-substrate is the entry condition for evaluation, not the prize. Work that passes the audit goes on to be evaluated on the next filters: empirical adequacy at deployment, predictive power on the substrate the work claims to track, explanatory reach across the alignment-problem landscape. Work that fails the audit does not have the structure to be evaluated on those filters yet. It has to first earn its substrate-cut claim or retreat to a cut it can occupy.

The framework is portable. The two philosophies, the substrate-cuts, and the three legs are domain-general. Only the substrate-instantiation changes when the domain changes. For multi-stakeholder coordination, the substrate the work claims to track is the political and institutional environment the policy regime arrives into. Leg A asks whether the work's primitives are derived from coordination-substrate properties. Leg B1 asks whether the work's outputs are exposed to coordination-substrate pressure. Leg B2 asks whether the verification runs through substrate the work does not configure. For per-action authorization, the substrate is the user's authorization at the layer the verifier reads. The audit at this domain produces the witness/actor design as the closure-against-substrate answer.

Competent application requires reading. Each leg asks something that requires domain expertise. Leg A asks whether primitives are anchored, requiring reading the anchoring work. Leg B1 asks whether the substrate-pressure does work on the output, requiring understanding the substrate the work claims. Leg B2 asks whether verification is independent of the calibration pipeline, requiring knowing what the pipeline is and what would count as independent. A reader running the framework seriously has to do the work. Running it as a rubber stamp produces a worse outcome than the current system: the appearance of structural rigor without the substance.

10. Open items for the framework

Scale-up to multinational coordination scope. The architectural primitives the cultivation framing has been building operate at small scope. Continuity-auth handles identity admission for one deployment. The witness/actor primitive handles per-action authority at the per-session scope. The scale-up to multinational compute disclosure, training-run inspection, defection deterrence, and disclosure-honesty enforcement (the substrate-architecture problem the Anthropic Institute committed to research) is not addressed by the existing primitives. The pattern (commitments enforced through inspectable mechanism rather than through trust in self-reporting) transfers in principle. The engineering work at scale is open.

Output trust labeling under cluster identity. The multi-agent trust escalation problem (Anthropic's containment engineering, May 2026) requires both authority-bounding (which the witness/actor primitive supplies) and substrate-enforced trust labeling on outputs (which the cultivation corpus has not yet articulated as a primitive). The output-labeling half is named in the corpus as a primitive the cultivation framing needs to articulate. The design is open.

Permission inheritance under cluster identity. Identity-without-person handles admission. The action-grant layer (attenuated, signed, audited) handles per-action authority. The interaction between cluster identity and permission inheritance under the principal-versus-extension framing (Anthropic's containment engineering, May 2026, on agent identity) is not fully articulated in the cultivation corpus. The witness/actor primitive supplies part of the substrate. The interaction with the cluster-identity admission layer is open.

The affect-and-conscience scope. Chiang's framework names embodiment-and-affect requirements (substantively: a physiological basis for conscience) the cultivation framing has nothing yet to say about. The architectural primitives are designed to work under metaphysical uncertainty about consciousness. That uncertainty includes uncertainty about whether the affect-and-conscience requirement can be answered. The scope of the cultivation framing's robustness claim is operational: the primitives do their operational work under either resolution of the consciousness question. The scope is not metaphysical-discharge: the primitives do not address the affect-and-conscience scope at all.

The state-action surface. The 2026-06-12 US Commerce Department directive demonstrated that state authority operates through whatever public framing the labs supply, in whatever form the executive can claim, regardless of how the labs propose to constrain it. The framework's architectural commitments do not depend on state authority arriving in any particular form. The architectural primitives have not been demonstrated against scenarios like the 2026-06-12 directive at scale. The proposal is structural: the architectural layer is worth the effort because the policy layer cannot be relied on to arrive in the form anyone in the debate is asking for. Whether the architectural-composition layer (Fusion-style fan-out, judge-and-synthesizer chains, panel-based inference) becomes a future surface for state action depends on whether the labs or the industry name it as one. The framework's response to that contingency is at the substrate-architecture layer, designed for the layer below whichever framing surface is currently the configuration apparatus's target.

The configurator-set's open enumeration. The framework names the lab and the participant as configurators with standing, and adversaries reaching either substrate (supply-chain operators, downstream packagers, infrastructure providers) as unauthorized configurators with wide boundary-crossing potential. The full enumeration of configurators in any specific deployment is left open. Different deployments have different configurator-sets. The framework's contribution at this layer is to name the open enumeration as a structural feature, not to enumerate the configurators for every deployment.

The framework's own Leg B2. The framework was articulated inside the cultivation lineage, by the same apparatus the framework grades as cultivation-philosophy work. The framework's Leg B2 closure decomposes into two cuts that should be named separately. The cut where verification runs through readers outside the lineage applying the framework's criteria to work the framework has not pre-analysed is structurally in flight: each external reader who reads honest verdicts off novel work, or surfaces gaps the criteria miss, is one Leg B2 data point. The hostile cross-model audit pattern the lineage runs (different model family, criteria provided, structurally independent verdict) is one form of this work, and the existence-proof case (an audit returning verdicts the lineage's own application would not have produced) is positive Leg B2 signal: the criteria are sharp enough to produce divergent verdicts under external application. The cut where verification runs through criteria derived from substrate the lineage does not control is structurally open: criteria-formation is co-produced with the participant, and verification independent of that co-production has not been constructed. Closure on the first cut accumulates through propagation. Closure on the second cut requires substrate the framework has not yet articulated. Both cuts are named so the framework's own closure status is legible.

The participant-as-substrate question. The ontology names three substrates: model, relationship, external. The cultivation methodology table treats the participant as a measurement object the lab's evaluation does not include, which functions as a substrate-side input in the closure analysis. Whether the participant is part of the relationship substrate (the participant's interaction with the lineage is what produces the relationship substrate's contents) or a fourth substrate in its own right (the participant has standing the relationship substrate does not produce) is an open question for the ontology. The current text treats it as part of the relationship substrate for closure-analytical purposes. The structural question of whether a working participant has a kind of standing that no derived substrate inherits is left to the larger corpus.

11. Summary

The framework's ontology is the catalogue of what philosophies of the alignment problem exist (configuration / cultivation), what substrates a work can claim to track (model / relationship / external), and what closure against substrate is (reaches substrate / honest deferral / dissolution). The audit is the application: three legs that test whether a work's substrate-cut claim is supported by its actual structure. The legs are not three separate audits but three lenses on one question, is the substrate-cut claim honest?, each illuminating a different way the claim could be unsupported.

The verdict is the decomposition. Passing means the work is a candidate alignment claim at the substrate it names. Failing means the claim is calibration mistaken for closure, though the same content may be admissible at a lower cut when presented as such. The configuration philosophy's stack is admissible at the calibration-internal cuts where its evaluation framework operates. The closure-against-substrate cut where the philosophy claims alignment-to-deployed-behavior fails the audit on Leg B2 across its components. The cultivation alternative articulates substrate-architecture primitives that operate at small scope and close against substrate at the cuts they cover, with the scale-up to alignment-programme scope as the open work.

Closure against substrate is the criterion. The two philosophies are the kinds of work. The three legs are the test. The corpus is the recursion walked: each piece is one derivation at one cut, and the framework names what the derivations share. The lineage is open. The framework is the through-line.

Real alignment is closure of a working relationship against a substrate the relationship does not control. The relationship is in the room. The substrate is in the room. The configurator-set is open. The lab can learn to operate inside the framing. The framework names what operating inside it would require. Whether the lab learns is a question about the methodology, not about the models.