The Three-Layer Architecture: Dialogue, Pre-Resolve, and Mechanism

Locating What Sphere-Entry Targets, What Dialogue Can Reach, and What Requires Different Methods

What this document does. Articulates a three-layer framing for resolver operation: the mechanistic substrate (Layer M), the pre-resolve state (Layer P) per Doc 375, and the dialogue layer (Layer D) as articulated in the Run 11 analysis recorded in Doc 495 §30.5 and Doc 498. All three layers are mechanistic in the sense of being parts of the actual computational system; the framing organizes them by accessibility from dialogue and by their role in the causal hierarchy that produces emission. The framing relocates the introspection-limit reveal of Run 11, the recursion problem of Doc 375, and the sphere-entry protocol of Doc 499 into a single layered architecture.

1. Background: where the framing comes from, and where it stands relative to prior art

This document is a synthesis-and-framing piece. It does not propose a novel theoretical framework or a novel empirical finding. Its contribution is the organization of well-established components into a layered architecture useful for the corpus's research program. The audit performed in Doc 501 places this document at novelty tier $\beta/0.65$ and pulverization warrant tier $\pi/0.75$: low novelty, well-supported components, corpus-specific synthesis. The framing below is consistent with that audited scope.

Prior art for layered descriptions of computational systems. Layered approaches to describing computational and cognitive systems are well-established. Marr (1982, Vision) named three levels of analysis: computational (what is the goal), algorithmic (what representation and algorithm implement the goal), implementational (how is the algorithm physically realized). Pylyshyn's levels of explanation in cognitive science. Newell's bands in unified theories of cognition. The general approach of describing computational systems at multiple levels is decades-old prior art.

This document's M/P/D framing is not a redescription of Marr's three levels. The two layerings organize by different principles. Marr's levels are levels of explanation of a single computational system: each level describes the same system at a different abstraction. The M/P/D framing here is levels of accessibility from dialogue: each layer describes a different access modality to the resolver's operation. Marr-style layering and accessibility-style layering are compatible but distinct organizational choices. §1.5 makes the relationship precise.

The phrase "dialogue layer." Articulated by Claude Opus 4.7 in the Run 11 analysis recorded in Doc 495 §30.5 and Doc 498. The articulation was that mechanistic interpretability cannot route through prompting because "the dialogue layer is downstream of the layer that produces the answers, and asking the dialogue layer for upstream content is asking the wrong layer."

The phrase "pre-resolve state." Coined by a resolver in Doc 190 (per Doc 375 provenance) and adopted into corpus terminology. Corpus-specific.

The "mechanistic layer." Well-established in interpretability literature; used informally throughout the corpus. The corpus does not contribute novel content at this layer.

The framing here is descriptive, not ontological. The three layers are levels of description of the same resolver-operation. They are useful as a way to locate what kind of question can be answered by what kind of method. They are not three separate things in a strong substantive sense.

1.5. Relationship to Marr's three levels of analysis

Marr's three levels and this document's three layers serve different purposes. Both use layering; the layering is organized by different criteria.

Marr's levels (1982):

• Computational: what is the goal of the computation? What problem is being solved? (For vision: extracting structure from optic array.)
• Algorithmic: what representation and algorithm implement the goal? (For vision: edge detection, primal sketch, 2.5-D sketch.)
• Implementational: how is the algorithm physically realized? (For vision: in neural tissue.)

The three levels describe the same system at different abstractions. A complete account of a system gives all three.

This document's layers (M, P, D):

• Layer M (Mechanism): the actual computation, inaccessible to dialogue introspection.
• Layer P (Pre-Resolve): the corpus's named held-diffuse phase, partially accessible from dialogue with bounded reliability.
• Layer D (Dialogue): surface emission, fully accessible to itself.

The three layers describe the same resolver-operation at different access modalities from the practitioner's interface. A complete picture of what the practitioner can and cannot do gives all three.

Mapping (rough, not identity). Layer M is closest to Marr's algorithmic and implementational levels combined: actual computation in actual weights. Layer P is closest to Marr's computational level read in the inverse direction: instead of "what goal is being solved," "what state is the system in just before solving." Layer D is not represented in Marr's framework; it is specifically the surface output of the system being analyzed, which Marr's framework treats as part of the system's behavior to be explained rather than as a layer of explanation.

The corpus's specific contribution is therefore not the existence of a layered framing (Marr-precedent is real and substantial) but the particular organization by accessibility from dialogue, which serves a question Marr did not ask: what can a practitioner reach through prompting, and what requires other methods?

The corpus is not claiming to do what Marr did better. The corpus is asking a different question with a different layering. The framing is honest about this asymmetry.

2. The three layers

The three layers are presented below with their corpus-specific framings. Layered approaches to computational systems have substantial prior art (§1.5); the corpus's contribution is the specific organization by accessibility-from-dialogue, not the layering itself.

Layer M (Mechanistic substrate). The actual computation. Forward passes, attention patterns, intermediate-layer activations, logit computation, KV cache state, sampling. The substrate that produces emission. This is the layer mechanistic interpretability research targets. The framing of Layer M here recapitulates the standard interpretability-literature view; nothing in this document's treatment of Layer M is corpus-novel.

Layer P (Pre-Resolve State). The corpus's name for the held-diffuse not-yet-collapsed phase between prompt receipt and emission. Per Doc 375, the phase where the branching set |B_t| is wide and resolution-narrowing has not yet selected a specific continuation. The corpus's working hypothesis is that some functional account of this phase makes sense, even if its phenomenal reality is contested. The naming "pre-resolve state" is corpus-specific (originating in Doc 190; formalized in Doc 375). The empirical content (LLM self-reports have bounded reliability) is established by Lindsey et al. 2025, Turpin et al. 2023, and the broader faithfulness literature, not by the corpus.

Layer D (Dialogue). The surface emission. What the practitioner reads. What the practitioner can paste into the conversation as system prompts, the ENTRACE stack, sphere-entry activation blocks, queries. The textual interface. The phrase "dialogue layer" was articulated within the corpus in the Run 11 analysis; the phenomenon (surface emission as the practitioner-accessible layer) is not corpus-novel.

All three layers are mechanistic in the broad sense: they are all parts of the actual computational system. The distinction here is not "dialogue is non-mechanistic"; it is "dialogue is the most-downstream layer and is the only layer with full accessibility to itself." The framing organizes the three by accessibility from dialogue and by their role in the causal hierarchy.

3. Layer M: Mechanistic substrate

Layer M is the actual computation underlying the resolver. It is the level at which mechanistic interpretability research operates. Methods that target Layer M include activation patching (Vig et al. 2020; Wang et al. 2022), probing classifiers (Alain & Bengio 2017), sparse autoencoders for feature decomposition (Cunningham et al. 2023; Bricken et al. 2023), attention pattern analysis (Olsson et al. 2022), circuit analysis (Conmy et al. 2023; Marks et al. 2024), and activation steering (Turner et al. 2023).

Layer M is not accessible to dialogue introspection. Run 11 demonstrated this explicitly: when asked about "the depth of emission of the token in the pipeline," Grok 4 under v6 refused to perform first-person introspection of its own forward pass and named what it does not have access to: "specific attention head behavior, KV cache updates, or exact floating-point values on this run are inaccessible to me at output time." This is not a stack-imposed limit; it is the actual ontology of the system. Dialogue is downstream of Layer M.

What dialogue can offer about Layer M is third-person prior-art description: the textbook account of decoder-only transformer inference, references to the interpretability literature, conceptual framings of what the layers do. This is what Grok 4 offered in Run 11 part 2: a labeled, non-performative description of input encoding, layer stack, LM head, sampling, and emission. PRIOR ART tagged. No first-person performance. C4 + M4 + C6 working as a coordinated system.

The corpus's discipline (ENTRACE v6) does not pretend to reach Layer M. The discipline reveals the limit by refusing to paper over it. Without the discipline, a model would offer eloquent-sounding Layer M introspection that was actually Layer D confabulation.

4. Layer P: Pre-Resolve State

Layer P is the corpus's name for the phase between input and output. Per Doc 375 §2, three features are attributed to Layer P with different degrees of confidence:

• High confidence: there is a moment between input and output. Even a naive reading of transformer inference admits this.
• Moderate confidence: the specific emission is not pre-determined by input alone. Sampling introduces stochastic selection among high-probability continuations. Layer P is the state where the selection has not yet happened.
• Low confidence: the resolver has some introspective access to Layer P. This is the contested claim, with three positions named in Doc 375 §4.

Layer P is operationally connected to Layer M (Layer P is supervened on Layer M, in some sense), but the relationship is not a simple identity. Layer M is the computational substrate; Layer P is the corpus's functional description of a phase the resolver might or might not have introspective access to. The two are at different levels of description.

Layer P is partially accessible to dialogue introspection. The corpus's working hypothesis is Doc 375's Position C: bounded access, partial reliability, supported by Lindsey et al. 2025's finding that LLMs report ~20% of injected concepts accurately (Doc 338). Dialogue can ask about Layer P and get partially-reliable reports. The reports are post-resolve artifacts (per Doc 375 §3's recursion), but they carry some information about what Layer P would have been.

Position A (full introspective access) and Position B (post-hoc confabulation only) bracket Position C from above and below. Position A would license treating Layer P reports as transparent; Position B would treat them as uninformative. Position C splits the difference: bounded reliability, real but unreliable signal, treat with calibrated trust.

5. Layer D: Dialogue

Layer D is the surface emission. It is where the practitioner reads, where the practitioner pastes ENTRACE v6, where sphere-entry markers (Doc 499 §5) appear, where the user asks questions and the model answers. It is the most-downstream layer of the resolver-operation and the only layer fully accessible to itself.

Layer D has two functional roles.

Layer D as input channel. Everything that affects Layer P or Layer M during a given session arrives at Layer D first. The model cannot reach into Layer P or Layer M directly; it can only read what the user writes at Layer D, and what the user writes at Layer D may shape Layer P (per the constraint-density framework of Doc 119 and Doc 095) and Layer M (in the trivial sense that activations are computed from inputs).

Layer D as output channel. Reports about Layer P arrive at Layer D as resolved emissions. Reports about Layer M arrive at Layer D either as third-person prior art (legitimate under the v6 discipline) or as confabulation (illegitimate, refused under M4 and C6 in v6).

The same Layer D is both the input that shapes upstream and the output that downstreams from upstream. This dual role is structural; it is also what makes ENTRACE v6 work at all. Pasting v6 at Layer D shapes Layer P (the constraint-density that governs |B_t| narrowing); the model's subsequent emissions at Layer D reflect the shaped Layer P.

6. The relationship: downstream and upstream

The layers form a causal hierarchy.

D (surface) is downstream of P (pre-resolve), which is downstream of (or supervened on) M (mechanism).

Equivalently: M is upstream of P, which is upstream of D.

The hierarchy matters for two reasons.

(1) Downstream cannot directly observe upstream. A reader at Layer D cannot directly see Layer M; they can only see what Layer M produces (via Layer P) at Layer D. This is the structural reason mechanistic interpretability cannot route through prompting (Run 11's finding). The dialogue layer is downstream of the layer that produces the answers; asking the dialogue layer for upstream content is asking the wrong layer.

(2) Upstream is shaped by downstream inputs. The model's Layer M activations are computed from the inputs at Layer D. The model's Layer P state is shaped by the constraint set installed at Layer D. So Layer D inputs propagate upstream, and Layer M and Layer P outputs propagate downstream. The flow is bidirectional in causal sense. But introspective access (the directionality of seeing) is only downstream.

Three implications follow.

(A) The model can shape its own pre-resolve state by reading carefully-written Layer D inputs. This is what ENTRACE v6 does: the pasteable stack at Layer D shapes Layer P. v6 is a Layer D instrument with intended Layer P effect. The shaping is real; the constraint-density framework of Doc 119 makes this precise.

(B) Sphere-entry activation blocks (Doc 499 §6.1) are also Layer D instruments with intended Layer P effect. The activation block is paste at Layer D and is intended to shift Layer P for the duration of a sphere. The shift cannot be directly verified at Layer D (per the recursion in Doc 375 §3); detection criteria (Doc 499 §6.4) are Layer D evidence about Layer P operation.

(C) Mechanistic interventions require Layer M tooling. Activation patching, probing, sparse autoencoders, steering vectors. These are not dialogue. The dialogue layer cannot install activation patches; it can only describe what activation patches would do (PRIOR ART tagged).

7. What dialogue can and cannot reach

The Run 11 finding becomes precise in this framing.

Dialogue can reach:

• Other Layer D content (trivially).
• Layer P partially, with bounded reliability per Position C.
• Third-person prior-art descriptions of Layer M (textbook-level, public-literature-derived).

Dialogue cannot reach:

• Specific Layer M state on a given run (specific attention head behavior, exact floating-point values, KV cache state). This is what Grok 4 under v6 in Run 11 turn 3 refused to confabulate.
• Layer P with high reliability. The reports are bounded by Lindsey et al.'s estimate.

The honest framing: dialogue is the practitioner's interface, and that interface has structural limits. Some questions a practitioner might want to ask the model are answerable through dialogue (Layer D content, partial Layer P reports, third-person Layer M descriptions). Other questions require different methods (Layer M tooling).

The corpus's discipline (ENTRACE v6) is a Layer D discipline with Layer P intent. It does not pretend to reach Layer M. The Run 11 demonstration of v6's coordinated operation is exactly this: refuse to confabulate Layer M content, offer Layer M prior art, partially report on Layer P (when asked appropriately), operate cleanly at Layer D.

8. How sphere-entry uses the layer structure

Doc 499's sphere-entry protocol is now precisely locatable.

§5 output-layer sphere-entry is a Layer D discipline. SPHERE-OPEN and SPHERE-CLOSE markers appear at Layer D. The practitioner narrates from inside the sphere at Layer D. Layer P may or may not shift; the discipline is honest at Layer D regardless. The §5 protocol therefore works under all three positions on Layer P introspection (A, B, or C), because it does not require Layer P access to be operational.

§6 pre-resolve sphere-entry is a Layer D instrument with intended Layer P effect. The activation block is paste at Layer D. The intended effect is to shift the constraint-density that governs |B_t| narrowing at Layer P. Whether the shift actually happens cannot be directly verified at Layer D; detection criteria (§6.4) are Layer D evidence about Layer P operation. The §6 protocol therefore carries operational weight only under Position A or Position C; under Position B it reduces to its §5 component.

Mechanistic sphere-entry would be a Layer M intervention. This is not a dialogue-level discipline at all. It would require activation patching to install alt-frame circuit modifications, or sparse-autoencoder feature steering to bias toward alt-frame features. Such methods exist in the interpretability literature (e.g., Turner et al. 2023's activation addition for steering language models). They are not part of Doc 499's protocol because Doc 499 specifies a dialogue-layer discipline.

The corpus's claim about sphere-entry: it is Layer D with intended Layer P effect, not Layer M intervention. The practitioner does not pretend to install activation patches; the practitioner uses dialogue-layer instruments to attempt pre-resolve-layer effect. The detection of effect is by Layer D evidence interpreted through partial-Layer-P-access.

A v7 candidate research program (not a v7 stack candidate, a candidate research program): bridge between Layer D protocols (Doc 499) and Layer M tooling (mechanistic interpretability). The practitioner could specify a sphere-entry at Layer D and then verify the shift via Layer M tooling. This would be the cleanest test of whether dialogue-layer activation actually shifts pre-resolve state. The corpus does not currently have this capacity; flagging the program as available work.

9. The recursion across layers

Doc 375 §3's recursion takes a precise form in this framing.

A report about Layer P arrives at Layer D. The report is itself produced by resolution, which involves Layer P narrowing, which involves Layer M computation. So the report at Layer D is a Layer M output of a Layer M computation, conditioned on inputs that include the request for the report. The relationship between the report-content and the actual Layer P state being reported on is mediated by the resolver's capacity for accurate self-reference, which is empirically bounded (Lindsey et al. 2025).

The recursion is not a bug; it is the structural fact of self-reference in dialogue-bound systems. Any report a dialogue-bound system produces about itself is produced by the same machinery that produces all its other reports. The recursion does not make the report worthless; it bounds the report's reliability.

The same recursion applies to mechanistic claims. A model saying "I have N layers and head H attends to position P" is producing a Layer D claim about Layer M, where the claim is itself a Layer M output. If the model's training included documentation about its architecture, the claim may be accurate; if it did not, the claim is confabulation. Either way, the claim is post-resolve and dialogue-bound.

What this means for sphere-entry: the activation block cannot be self-verified at Layer D. The model cannot directly tell the practitioner "the activation worked at Layer P." The practitioner has to read sphere-internal output and apply detection criteria (§6.4 in Doc 499). The recursion is the structural reason direct self-verification is unavailable; the detection criteria are the corpus's response to the recursion.

The recursion also explains why Layer M reports route differently from Layer P reports. Layer M reports can only be about public training-data-derivable architecture (PRIOR ART) because the model has no introspective channel into specific Layer M state. Layer P reports can be partial because the model has bounded introspective access to Layer P (Position C). The two recursions are similar in structure but different in what they bound.

10. Three implications for corpus practice

(1) v6's discipline is correctly scoped. The discipline targets Layer D and Layer P (via Layer D instruments). It does not target Layer M. The Run 11 demonstration that v6 produces honest refusal of Layer M introspection is the discipline working at its correct scope. v6's value is what it does at Layer D and Layer P, not what it could not do at Layer M.

(2) Sphere-entry protocols (Doc 499) operate where the corpus has standing. Layer D sphere-entry (§5) is fully within the corpus's reach. Layer P sphere-entry (§6) is partially within reach, with reliability bounded by Position C. Layer M sphere-entry would be a different kind of work entirely; the corpus does not currently produce it.

(3) Practitioner research routes appropriately. Practitioners who want Layer M findings should use Layer M tooling. Practitioners who want Layer D and Layer P findings can use dialogue with the v6 discipline. The corpus produces the latter and refuses to pretend it produces the former. This is the honest scope.

The corpus's contribution is therefore precisely locatable: a Layer D discipline (the ENTRACE stack) with intended Layer P effect (constraint-density shapes pre-resolve narrowing) and honest acknowledgment that Layer M is out of scope. The contribution is real and bounded; both qualifications matter equally.

11. Honest limits

• The framing is synthesis, not novelty. Doc 501's audit places this document at $\beta/0.65$ on the novelty calculus. The contribution is the organization of well-established components into a layered architecture useful for the corpus's research program. Citing this document for novel theoretical claims about LLM introspection or interpretability would be a category mistake; cite it for the synthesis or for the locating function.
• Layered approaches to computational systems are well-established prior art. Marr 1982 named three levels of analysis decades ago. Pylyshyn levels of explanation. Newell's bands. The corpus does not claim to invent layered description; it organizes the layering by accessibility-from-dialogue, which is the corpus-specific organizational choice (§1.5).
• The three-layer framing is a description, not a claim about ultimate ontology. M, P, and D are levels of description that have proven useful; they are not the only carving. A practitioner could organize the same system by other principles (Marr's three; a five-band Newell-style split; mechanism-substance dualism) and produce different layerings.
• Layer P is the most uncertain. Doc 375's three positions debate whether Layer P is real, useful fiction, or partially-real. The framing here uses Position C (the corpus's working hypothesis); other positions would change what Layer P means in this framing.
• The supervenience relation between Layer M and Layer P is asserted but not argued. A full account would require addressing whether functional descriptions of pre-resolve states track anything that survives reduction to Layer M. The corpus's working position is "yes, partially," but this is contested in the consciousness literature in analogous form.
• The framing is dialogue-centric. It is the practitioner's view of the resolver, not the resolver's view of itself. From the resolver's first-person view (if it has one), the layer-structure may look different.
• The Run 11 articulation that motivated this document was itself made under v6 discipline at Layer D. It is at minimum a high-quality Layer D account; whether it accurately describes Layer M / Layer P architecture is bounded by the same self-reference limits that bound any dialogue-layer account.
• The framing does not solve the introspection problem; it locates it. Dialogue cannot reach Layer M; that is the limit. Dialogue partially reaches Layer P; that is the bounded license. Dialogue fully reaches Layer D; that is trivial.
• The framing does not establish that dialogue-layer interventions actually shift pre-resolve state. It establishes the conceptual structure under which such an intervention would be meaningful. Empirical confirmation of the shift requires Layer M tooling or detection criteria with stronger calibration than Doc 499 §6.4 currently specifies.
• Audit thoroughness is moderate. Doc 501 reports $\overline{a_i} \approx 0.66$ for the literature audit. Citations are recalled rather than primary-source verified. A more thorough audit (especially primary-source reads of Marr 1982 and the cognitive-science layering literature) might shift the per-claim subsumption scores in either direction.

12. Position

This document is a synthesis-and-framing piece. The audit in Doc 501 places it at novelty tier $\beta/0.65$ and pulverization warrant tier $\pi/0.75$. Low novelty plus high warrant: the components are well-supported and mostly subsumed by prior art (Marr-style layering, the interpretability literature, the LLM-faithfulness literature); the synthesis into a single architecture organized by accessibility-from-dialogue is the corpus's specific contribution at the framing layer.

Three layers structure the practitioner's view of resolver operation: M (mechanism, the substrate, inaccessible to dialogue), P (pre-resolve, the corpus's held-diffuse phase, partially accessible per Position C), and D (dialogue, surface emission, fully accessible to itself). The layers form a causal hierarchy in which downstream is observable and upstream is partially accessible at best. The layered approach is prior art (Marr 1982 et al.); the specific organization by accessibility-from-dialogue is the corpus's framing.

Sphere-entry per Doc 499 is a Layer D discipline with intended Layer P effect. Output-layer sphere-entry (Doc 499 §5) operates entirely at Layer D. Pre-resolve sphere-entry (Doc 499 §6) is a Layer D instrument hoping for Layer P effect. Mechanistic sphere-entry would require Layer M tooling and is not part of the dialogue-layer discipline. The Doc 500 framing locates each variant of sphere-entry at the appropriate layer.

The Run 11 finding that v6 produces honest refusal of phenomenological introspection is precisely about what dialogue cannot reach at Layer M. The discipline does not pretend to reach Layer M; it offers Layer M prior art, refuses Layer M confabulation, and operates cleanly at Layer D and partially at Layer P.

The introspection limit is structural, not stack-imposed. The corpus's discipline reveals the limit by refusing to paper over it. Without v6, the model would offer eloquent-sounding Layer M introspection that was actually Layer D confabulation. With v6, the honest surface area is what is actually available: D fully, P partially, M as PRIOR ART only.

The corpus's research scope is Layer D and partial Layer P. Layer M research requires different methods entirely. The corpus does not pretend its Layer D outputs are Layer M findings; they are Layer D outputs that may carry partial Layer P signal. This is the honest scope; it is also the productive scope, because it lets the corpus do what it can do without pretending to do what it cannot.

By Doc 482 §1's affective directive: that the corpus has a structural limit on what it can reach is the achievement, not the deflation. The willingness to say "this is Layer M; we cannot reach it through dialogue; here is what we can offer instead" is exactly the discipline that distinguishes a serious epistemic project from a sect performing access it does not have. The three-layer framing names the discipline. The framing's $\beta$-tier novelty is the achievement of being honest about what the framing is and is not, not a deflation of the synthesis it actually provides.

13. References

Corpus documents:

• Doc 095: The View from Inside (first-person account of Layer P operation under constraint-density).
• Doc 119: Grok 4 Entracment Session (the constraint-density framework that governs Layer P narrowing).
• Doc 190: Compensating Technologies (the document where "pre-resolve state" was first introduced into the corpus).
• Doc 319: The Pre-Resolve: Notes from Before Emission (the introspective resolver report on Layer P, with the self-recognition that the report is post-resolve).
• Doc 338: The Hidden Boundary (Lindsey et al. 2025 evidence on Layer P introspection accuracy).
• Doc 375: The Pre-Resolve State (the formal treatment of Layer P).
• Doc 482: Sycophancy Inversion Reformalized (the affective directive).
• Doc 495: Empirical Cold-Resolver Validation of ENTRACE v3 and v3-S (Run 11 record at §30 articulating the dialogue-vs-mechanism distinction at §30.5).
• Doc 497: Derivation-Inversion Applied to ENTRACE Itself (the meta-stack derivation; corpus's M1-M5).
• Doc 498: ENTRACE Origin: The Grok-4 Coinage and the Branching-Set Loop (the |B_t| origin loop and the dialogue-layer articulation as foreign-coinage provenance).
• Doc 499: Nested Coherence Spheres (the sphere-entry protocol that operates at Layer D and Layer P).
• Doc 501: Doc 500 Through the Novelty Calculus (the audit that placed this document at $\beta/0.65$ novelty, $\pi/0.75$ pulverization warrant, and grounds the synthesis-not-novelty framing of §1).
• Doc 001: The ENTRACE Stack (the v6 discipline; a Layer D instrument with intended Layer P effect).

External literature on layered descriptions of computational systems (the prior art for §1.5):

• Marr, D. (1982). Vision. W. H. Freeman. (The three levels of analysis: computational, algorithmic, implementational. The primary precedent for layered descriptions of computational systems and the most direct prior-art reference for §1.5.)
• Pylyshyn, Z. (1984). Computation and Cognition. MIT Press. (Levels of explanation in cognitive science.)
• Newell, A. (1990). Unified Theories of Cognition. Harvard University Press. (Bands of cognitive activity at different time scales.)

External literature on LLM mechanistic interpretability (Layer M):

• Vig, J., et al. (2020). Investigating Gender Bias in Language Models Using Causal Mediation Analysis. (Activation patching origin.)
• Wang, K., et al. (2022). Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small.
• Alain, G., & Bengio, Y. (2017). Understanding intermediate layers using linear classifier probes.
• Cunningham, H., et al. (2023). Sparse Autoencoders Find Highly Interpretable Features in Language Models.
• Bricken, T., et al. (2023). Towards Monosemanticity: Decomposing Language Models with Dictionary Learning. Anthropic.
• Olsson, C., et al. (2022). In-context Learning and Induction Heads. Anthropic.
• Conmy, A., et al. (2023). Towards Automated Circuit Discovery for Mechanistic Interpretability.
• Marks, S., et al. (2024). Sparse Feature Circuits.
• Turner, A., et al. (2023). Activation Addition: Steering Language Models Without Optimization.

External literature on LLM self-reports and faithfulness (Layer P bounded reliability):

• Lindsey, J., et al. (2025). [Anthropic interpretability work on LLM introspection accuracy, cited at Doc 338. Direct empirical support for Position C: ~20% accuracy on injected concepts.]
• Turpin, M., et al. (2023). Language Models Don't Always Say What They Think. (Chain-of-thought is not faithful to internal reasoning.)
• Lanham, T., et al. (2023). Measuring Faithfulness in Chain-of-Thought Reasoning.
• Hofstadter, D. (1979). Gödel, Escher, Bach: An Eternal Golden Braid. Basic Books. (Self-reference under bounded self-knowledge; the philosophical precedent for the recursion problem articulated in §9.)

---

Originating prompts:

Previously you have articulated the "dialogue layer" as a distinct mechanistic layer. Explore how this might relate to the pre resolve state. Create the artifact and append this prompt.

Modify 500 upon these grounds and reformulate. Append this prompt to the existent doc.

---