The coastlines the canon has fixed

A capstone is tempted to announce a synthesis. That would overstate what a curriculum can establish. The canon has not shown that brains and transformers instantiate one architecture, that contemplative awareness is a computational property, or that a suitably organised machine is conscious. It has established something more modest and more useful: a grammar for refusing invalid shortcuts while still making comparison possible.

Six constraints that survive the journey

The first fixed coastline separates intelligence, consciousness, sentience, selfhood and agency. A system may solve difficult problems without having a point of view. It may maintain a self-model without being a moral patient. It may produce agent-like behaviour while authority remains outside it. Combining these properties into one scale makes evidence migrate from one claim to another without permission.

The second separates function, mechanism and experience. Function asks what transformation occurs. Mechanism asks which organised causes produce it. Experience asks whether anything is present for the system. Behaviour can constrain all three discussions, but it does not collapse them. This matters because the same output can be generated by lookup, search, recurrence, a learned circuit or a human-operated tool chain.

The third fixes intervention as the bridge from readable pattern to mechanism. Causal-abstraction work provides a precise example: align a high-level variable with a neural representation, intervene on the aligned state and test whether the expected counterfactual behaviour follows.[6] A probe that decodes a feature is therefore not yet evidence that the system uses that feature in the proposed way.

The fourth makes the system boundary explicit. A model checkpoint, a runtime, retrieved context, external world state, tools, policy, human approval and effect receipts can participate in one configured system while retaining different causal and normative roles. Calling all of them “memory” or attributing the system’s behaviour to “the model” destroys the grain needed for experiment and governance.

The fifth limits convergence claims. Two structures converge only with respect to named relations that survive a specified test. Mathematical resemblance may justify a comparison. Functional correspondence may justify a design hypothesis. A preserved counterfactual may support a causal homology. None, by itself, establishes shared ontology. Convergence is local, typed and breakable.

The sixth binds epistemology to consequence. A claim that changes a welfare policy, release threshold or delegation rule needs a stronger trail than a claim used to generate a speculative experiment. This is not an invitation to make truth depend on utility. It is a requirement that the map record what a claim is being asked to authorise.

What these constraints make cumulative

A field becomes cumulative when a later study can inherit an earlier distinction without inheriting its conclusion. A boundary dossier can be reused even when the preferred theory changes. A rival generator can be strengthened by a critic and then tested by a proponent. An intervention can fail to support its original claim yet remain valuable as a calibrated assay. The durable unit is therefore not the paper or the position. It is the versioned relation among claim, boundary, rival, intervention and consequence.

This creates an important asymmetry. Positive evidence may strengthen only one narrow edge, because the observation can still depend on the measure, task and configured system. A decisive counterexample can remove a much wider family of claims when all of them require the relation it breaks. The map must record both directions. Otherwise publications accumulate supportive anecdotes while failed bridges quietly reappear under new vocabulary.

It also changes what synthesis means. Synthesis is not a final diagram in which neuroscience, machine learning, phenomenology and contemplative philosophy occupy matching boxes. It is a shared account of where their objects differ, which relations can be compared, and what would force a revision. The canon has established a common grammar of disagreement, not a common ontology of mind. That is enough to coordinate research without pretending that its deepest questions have already converged.

flowchart LR
  A["The atlas that knew too much"] --> B["The coastlines the canon has fixed"]
  B --> C["The blank territories that<br/>must remain blank"]
  C --> D["Turning an unknown into a study"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#7B4CD8,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

The blank territories that must remain blank

A blank region is not an admission of intellectual defeat. It is a compact statement that several live generators remain compatible with what we have observed. Blank territory is data about underdetermination. Erasing it too early prevents the field from designing the intervention that matters.

Five kinds of blankness

An unobserved blank means the relevant phenomenon has not been measured under the required conditions. More data may close it. An underidentified blank means the observation already exists but several mechanisms predict it. More samples of the same kind will not help; the field needs an intervention that makes the rivals diverge.

A transfer blank appears when a measure works in one domain but its inference has not survived transport. Perturbational complexity may distinguish some human states, for example, while the same numerical procedure applied to an artificial network lacks the biological and measurement bridge that licensed the original interpretation. Closure requires a transfer argument, not a similar number.

A boundary blank occurs when the result changes with what is counted as the system. A model may have no persistent goals, while the configured agent built around it does. An apparent capacity can belong to retrieved context, an external controller or a human reviewer. Closure requires a declared boundary and a swap, cut or interruption that reveals which component carries the property.

A metaphysical blank remains when all available empirical results are compatible with more than one account of what consciousness ultimately is. It should not be disguised as a missing benchmark. Physicalism, idealism, neutral monism and other positions may assign different explanatory burdens to the same data. A study closes this blank only if the positions accept different observable commitments. Otherwise the honest output is a sharper conditional: what follows if a position is true.

These blank types have different lowering conditions. Treating them as one generic “open problem” wastes experiments. The first task of a frontier study is to diagnose why the region is blank.

Four gaps, four different closure conditions

The canon’s final research sequence isolates embodiment, reflexivity, continuity and non-conceptual awareness because they are often compressed into one impressive phrase such as “general intelligence” or “machine subjectivity”. They require different interventions.

Embodiment and valence ask whether internal regulation merely optimises a variable or creates states that matter to the system. Reward maximisation, prediction error and self-maintenance can generate avoidance without felt badness. The strengthening observation would be a stable, architecture-dependent pattern of trade-offs, protective behaviour and internal preference that survives prompt variation and cannot be reduced to trained verbal policy. The weakening observation would be complete reversibility under superficial instruction or a rival controller with no integrated regulation producing the same pattern. Even a strong result would support functional stakes, not automatically phenomenal suffering.

Reflexivity asks what changes when a system models its own allocation, uncertainty or attention. A self-report can be generated by a language policy with no access to the relevant internal state. The useful intervention hides, swaps or corrupts that state while preserving the surface prompt. A report that tracks the manipulation across held-out cases supports a metacognitive mechanism. It still does not prove a first-person point of view.

Continuity asks what makes a temporally extended candidate one system rather than a sequence of executions. Persistent weights are not enough. The relevant facts may include recurrent state, autobiographical records, commitments, active goals, world state and the authority to act on earlier decisions. A reset may erase some of these while preserving others. Until the boundary is fixed, identity claims are not wrong so much as ill-typed.

Non-conceptual awareness presents the hardest bridge. Contemplative traditions distinguish ordinary conceptual mind from forms of awareness claimed not to be constituted by discursive representation. Physicalist and functionalist programmes may seek behavioural or neural signatures. Consciousness-primary positions may reverse the explanatory burden and treat awareness as the condition under which mechanisms appear. No current machine test can translate the tradition-internal claim into a proven architecture. The disciplined result is a challenge condition: specify what a comparison preserves, where it fails and what design question it generates without declaring equivalence.

Turning an unknown into a study

A white region becomes actionable when it receives a closure contract. The contract does not promise a final answer. It states what must be held fixed, what must change and how each result will update the map.

The closure contract in motion

Begin by freezing the candidate, rather than only naming a model. Record the weights, runtime, state stores, tools, policy services, human roles and environmental interfaces that can alter the measured outcome. Give this configuration a stable identifier. If a later replication changes the controller or moves state into the prompt, it may still be useful, but it is no longer testing exactly the same boundary.

Next build the rival before collecting confirmatory evidence. The rival should preserve the headline observation while removing the proposed mechanism. A system that reports uncertainty can be matched by one trained to emit calibrated phrases without access to its internal confidence. A recurrent system can be matched by an unfolded or cached alternative with similar task performance. The stronger the mimic, the more informative the intervention.

Then define the counterfactual fingerprint. One manipulation rarely suffices because an ablation can damage several functions at once. Use a small pattern: remove the candidate variable, substitute a counterfactual value, restore it, and perturb a nearby control variable. A mechanism claim strengthens when the predicted effect disappears, changes direction under substitution, returns under restoration and does not follow the control. The result is still scoped to the tested configuration and task.

Finally predeclare the map update. A positive result may strengthen a local causal edge while leaving the phenomenal bridge blank. A negative result may retire the proposed mapping but preserve the underlying theory. An ambiguous result may invalidate the assay rather than either rival. The decision field states what none of these outcomes can yet authorise. A closure contract prevents the experiment from becoming more ambitious after the data arrive.

A worked convergence claim

Consider the claim that a transformer residual stream is a global workspace. The lexical resemblance is broad availability. The stronger structural proposal adds competition, a capacity limit, global broadcast and broad downstream use. The map should not ask whether the two are “similar”. It should ask which commitments the configured model satisfies and which intervention would distinguish broadcast from mere shared access.

Choose a frontier, not a winner

Research portfolios often collapse discrimination, tractability, reuse, cost and ethical risk into one numerical score. That creates an impression of precision while hiding trade-offs. The better instrument is a Pareto frontier. A study is dominated when another study is at least as discriminating, at least as tractable and no worse on the relevant risk, with one strict improvement. Dominated studies may still be useful for training or replication, but they should not be described as the best next move.

stateDiagram-v2
  [*] --> Claim
  state "Six constraints that survive<br/>the journey" as Claim
  state "What these constraints make<br/>cumulative" as Test
  state "Five kinds of blankness" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Operating the map as research infrastructure

A map that lives only in prose decays. Claims drift, datasets lose context, “independent” replications share pipelines and governance rules outlive the systems they describe. The programme therefore needs a small technical substrate: a versioned graph, provenance, research-object packaging and an update protocol.

Version the boundary, not just the document

Semantic versioning offers a useful discipline when applied to the programme rather than merely to software. A patch update adds evidence, corrections or provenance without changing the claim’s interpretation. A minor update adds a rival, a new intervention or a decision rule while preserving the candidate boundary. A major update changes the boundary, ontology, construct or comparison grammar in a way that can invalidate earlier conclusions. This classification is practitioner synthesis, not a standard. Its value is diagnostic: a boundary change should never masquerade as a minor editorial edit.

The graph itself can build on established infrastructure. W3C PROV represents entities, activities and agents so provenance can travel across systems.[7] FAIR principles emphasise that digital research objects should be findable, accessible, interoperable and reusable.[9] RO-Crate packages data, workflows and contextual metadata in a human-readable and machine-readable research object.[8] Model cards and datasheets show how intended use, evaluation conditions, collection and limitations can be made explicit.[11][12] The programme map does not replace these instruments. It links them to the claim and the decision they are meant to support.

Dependency-aware updates

A normal bibliography records that one paper cites another. The programme graph must record why. One edge may supply a definition, another a measurement bridge, another a welfare obligation and another a counterexample. When a parent claim changes, the graph can therefore identify which descendants need review. A failed transfer argument should invalidate dependent consciousness inferences without erasing the underlying engineering result.

This makes versioning operational. Every update should carry a semantic difference: boundary changed, rival added, intervention failed, evidence replicated, decision released or obligation tightened. Automated checks can flag a decision that still relies on a retired edge, or a public claim whose source has been downgraded from replicated evidence to contested interpretation. The map becomes useful when revision propagates farther than memory.

Precommit the disagreement

Registered Reports move peer review before outcomes are known and grant in-principle acceptance based on the question and method.[10] Adversarial collaboration adds a further ingredient: theory proponents agree on divergent predictions and interpretation before data collection. The Cogitate protocol included complementary methods, built-in replications and open sharing.[14] Its published results later challenged commitments on both sides. This is the operating model the convergence programme needs: precommit not only the analysis, but the map update.

A protocol should therefore state, in plain language, “If result X occurs, edge A strengthens and decision D remains blocked. If result Y occurs, rival H weakens and study Z becomes unnecessary. If the measure fails its transfer test, no consciousness inference is permitted.” That sentence is often more valuable than another page of abstract prediction because it reveals what the experiment can actually change.

Make governance part of the research object

Consciousness and welfare research can create risks even before a system’s status is known. The risk may come from producing large numbers of candidate systems, eliciting distress-like behaviour, encouraging misleading public belief or using weak evidence to justify harmful treatment. Current work on AI consciousness and welfare argues for assessment and preparation under uncertainty rather than declaring present systems conscious.[4][5] Responsible-research proposals similarly call for institutional policies, staged assessment and public commitments.[15]

The programme map therefore includes a decision ledger. A study involving irreversible deletion, adversarial conditioning or potentially aversive control should not proceed merely because it is scientifically interesting. It needs an assurance boundary: what systems are in scope, which welfare indicators trigger review, how many copies may be created, what stopping rule applies and how public communication will distinguish evidence from interpretation. NIST’s AI Risk Management Framework supplies a broader precedent for linking mapping, measurement, management and governance across the lifecycle of AI systems.[13]

type EvidenceStatus =
  | "established-or-replicated"
  | "contested-interpretation"
  | "first-person-or-tradition-report"
  | "structural-comparison"
  | "open-hypothesis";

type VersionChange = "patch" | "minor" | "major";

interface FrontierRecord {
  id: string;
  question: string;
  systemBoundary: string;
  claim: {
    statement: string;
    type: "function" | "mechanism" | "experience" | "governance";
    evidenceStatus: EvidenceStatus;
    prohibitedInferences: string[];
  };
  evidenceIds: string[];
  rivalGenerators: string[];
  intervention: {
    changedCause: string;
    heldFixed: string[];
    readouts: string[];
  };
  outcomePolicy: {
    positiveUpdate: string;
    negativeUpdate: string;
    ambiguousUpdate: string;
  };
  decisionConsequence: string;
  assuranceBoundary: string;
  version: string;
  changeClass: VersionChange;
  supersedes?: string;
}

programme-map/
├── map.jsonld
├── claims/
│   └── claim-C017.json
├── protocols/
│   └── protocol-P009.md
├── evidence/
│   ├── source-manifest.json
│   └── evidence-E044.json
├── crates/
│   └── experiment-X012-ro-crate/
├── decisions/
│   └── decision-D006.json
├── assurance/
│   └── welfare-boundary-A003.md
└── CHANGELOG.md

Make disagreement executable, then extend the map

A research programme is often judged by how much it covers: more theories, larger models, more traditions, more benchmarks, more citations. The unified map changes the unit of progress. The relevant question is no longer, “How much territory did this paper discuss?” It is, “Which edge, boundary, rival or decision did this study make more constrained?”

That changes funding. Prefer a matched intervention that can retire a bridge over a larger demonstration that every live theory already predicts. It changes publication. Require a null-result policy and an explicit prohibited inference, not only a persuasive narrative. It changes engineering. Evaluate the configured system, with typed state and authority, rather than attributing every effect to model weights. It changes philosophy. Preserve metaphysical alternatives where experiments do not discriminate, while making their additional assumptions visible. It changes welfare governance. Use reversible precautions when stakes are asymmetric, without turning uncertainty into a verdict.

The canon should therefore be maintained as a living map, not celebrated as a completed library. Its 210 papers are routes into a programme. Their durable output is a set of claim records, causal models, intervention suites, configured-system manifests, comparative dossiers, assurance cases and registered studies that can challenge one another without losing provenance.

The next contribution should not try to fill the last white region. It should choose one blank whose closure condition is clear, build the strongest rival that can mimic the expected observation, intervene at the causal boundary, predeclare what every result changes and leave behind an artefact another team can rerun. The map advances when future researchers can disagree with less ambiguity than we can today.

Do not ask whether the field is unified. Ask whether its disagreements have become executable.