When Apparatus Sharpens Taxonomy: What Implementation Granularity Reveals About Type A, B, and C Failures

cafebedouin@gmail.com

—

Abstract

The trifurcation framework classifies reasoning failures into three operationally distinct types: Type A (drift across reasoning stages), Type B (axiomatic inconsistency), and Type C (indexical underspecification). The taxonomy was originally diagnostic: each type has a corresponding repair (frame-fix, axiom-revise, index-specify), and the diagnostic move is to identify which repair the failure calls for. A subsequent paper, The Asymmetry of Failure Types, extended the taxonomy architecturally: the same three types divide unevenly along treatment lines when built into a working apparatus, with Type B detected by formal computation, Type C prevented by required indexing, and Type A governed by paired synchronic and diachronic discipline.

This paper extends the architectural move with what an audit chain on a working apparatus revealed about the internal architecture of each type — granularity the diagnostic and architectural framings did not anticipate. Type C indexical specification has functional decomposition: in the apparatus, the (P, T, E, S) index isn’t a flat 4D space but a two-hub architecture where T and E feed Hub 2 (mountain/rope immutability) while P and S feed Hub 1 (χ via sigmoid), and Hub 2 captures more between-slice structural variance than Hub 1 under four of five tested metrics. Type A drift can occur between specification and implementation, not only across reasoning stages: the apparatus’s framework paper wrote χ = ε × f(d(P)) × σ(S(P)) for years while the implementation had d(P, E) and σ(S), an unmarked specification-vs-code mutation invisible until an audit closed the gap. Type B axiomatic-consistency claims have a scope-design dimension that the original account elided: the framework’s site-stability claim — H¹ binary preservation under observer-site expansion — survives with a precise qualification (scope ≤ global at the analytical observer position) rather than as universal, and identifying the qualification required testing the claim on a third site that violated the design constraint.

The synthesis is that implementing a Type-A/B/C diagnostic apparatus forces decisions the abstract diagnostic doesn’t make, and those decisions reveal that the categories have internal structure the original paper didn’t capture. The audits don’t displace the trifurcation; they sharpen what each type’s repair operation actually requires when the apparatus is mechanical rather than rhetorical.

—

1. From Diagnostic to Architecture to Granularity

The trifurcation framework — Type A drift, Type B structural inconsistency, Type C indexical underspecification — was introduced in Debugging Philosophy as a diagnostic move on philosophical paradoxes. A grammatically singular question like “What is Sleeping Beauty’s credence?” packages multiple distinct queries indexed by different coordinate systems; specify the index and the paradox dissolves. A diminishing heap that “remains a heap” through grain-by-grain removal trades on an unmarked threshold; fix the frame and the paradox dissolves. A self-referential set R = {x | x ∉ x} produces immediate contradiction without process or perspective; the axioms themselves are inconsistent and the system requires revision. Three failure modes, three repair operations, three different kinds of problem.

The Asymmetry of Failure Types extended this from taxonomy to architecture. When the same three types are built into a working apparatus — the Deferential Realism (DR) engine, which classifies social constraints across observer positions — they divide unevenly along treatment lines. Type B has multiple dedicated detectors at distinct measurement layers (Boltzmann compliance, structural signatures, False CI Rope, False Summit Mountain, False Natural Law). Type C has two architectural implementations (gauge orbits and cohomology) that operate as schema-layer prevention plus output-layer measurement. Type A has one primary instrument (the drift-events module), and its coverage is principled rather than accidental: drift in the frame of analysis cannot be exhaustively detected from within the same frame. The asymmetry tracks three different relations between failure and formalism — contradiction-to-structure, contradiction-to-specification, contradiction-to-governance — and validation is not a single operation but a family of operations with different points of entry and different limits.

This paper extends the architectural move with a third pass. Asymmetry established that the three types have different architectural shapes. The audit chain that followed Asymmetry — BC coupling, position-geometry metric sensitivity, the metric audit on the apparatus’s two-hub architecture, and the binary sheaf/presheaf boundary audit on a 10-slice working family — established that each type has internal architecture below the level the prior framings captured. The taxonomic frame says what can go wrong. The architectural frame says how each kind of wrongness can be known, prevented, or governed. The granularity frame says what the repair operation actually consists of when the apparatus is mechanical rather than rhetorical, and what kinds of failures can occur within the repair operation itself.

Three findings drive the synthesis. First, Type C indexical specification has functional decomposition that the abstract index-specification move doesn’t see. Second, Type A drift can occur between specification and implementation in ways the original taxonomy didn’t anticipate. Third, Type B axiomatic-consistency claims have a scope-design dimension that the original account elided. Each is developed in turn (§§3–5), with §6 drawing them together and §7 noting the specific ways in which the audits did not extend the trifurcation.

The thesis is not that the taxonomy was wrong. It is that operationalizing it discloses structure within each type, and that the structure matters for what the repair operations are and what they can claim to have accomplished.

—

2. Background: The Apparatus and Its Audit Trail

The DR apparatus operates on a corpus of 3,335 social constraints (laws, norms, institutional arrangements, professional licensing requirements). Each constraint has measured properties: base extractiveness ε, suppression level, theater ratio, beneficiary/victim structure, structural signatures. The apparatus classifies each constraint at observer positions — tuples of (Power, TimeHorizon, ExitOptions, Scope) — through a rule cascade that operationalizes the trifurcation. Type B detection runs as multi-layer formal checks against the structural axioms. Type C prevention runs as schema-layer requirements that no classification proceed without complete indexing. Cohomological obstruction (H¹) measures whether per-observer classifications glue into a global section (sheaf, classification observer-independent) or fail to glue (presheaf, classification observer-dependent in a structurally measurable way). Type A monitoring runs as diachronic drift-event detection against historical states.

The framework paper (v6.11) reports findings on a canonical 4-point observer site (U₁–U₄, designed observer positions) and a 156-point product-site expansion (all permissible (P, T, E, S) combinations minus 24 category-error exclusions). Two of v6.11’s central empirical claims are the dimensional hierarchy P > E ≈ S > T (within-block variation analysis on the product site) and binary site-stability (the H¹ = 0 vs H¹ > 0 classification is preserved with zero crossings between the canonical and product sites).

Three audit passes followed v6.11. The BC coupling audit established that observer specification and structural variation in the extractive subgraph are coupled and forward-asymmetric (ρ = 0.350 forward, ρ = −0.121 reverse), with the coupling geometry-driven rather than label-driven (replacing ordinal PTES distance with empirical classification disagreement collapses the forward correlation to ρ = 0.010). The position-geometry metric-sensitivity audit established that axis dominance under structural-distance metrics is metric-specific: T leads under extractive fraction, E leads under three other metrics, S leads under the mountain-fraction negative control, and cross-metric agreement is weak. The metric audit traced these findings to the apparatus’s two-hub architecture: Hub 1 (P, S → χ via sigmoid f and scope modifier σ) and Hub 2 (T, E → discrete mountain/rope immutability via lookup table). Hub 2 captures more between-slice structural variance than Hub 1 under four of five tested metrics, and the two hubs are statistically independent predictors. The sheaf audit then tested the binary-boundary claim on a 10-slice working family adjacent to but distinct from the canonical site, and found a 68.98% crossing rate driven by a specific scope-modifier mechanism at the analytical observer position.

The audit chain produced four artifacts that bear on the trifurcation: (1) the two-hub architecture of indexical specification; (2) a code-vs-paper notation drift in v6.11’s Axiom 2 (d(P) → d(P, E); σ(S(P)) → σ(S)) that had survived multiple paper revisions invisible to its readers; (3) the scope-conditional site-stability finding; (4) a within-rope-group T-axis effect that recasts what “T-dominance under extractive fraction” actually means. The first three of these are the synthesis paper’s substrate. The fourth is a sub-finding that compose into the §3 development without warranting its own section.

—

3. Type C Has Functional Decomposition

The original trifurcation account treats indexical specification as a single move: the query “What is Beauty’s credence?” maps to multiple distinct queries indexed by observer (self vs external), time-slice (before sleep vs upon awakening), or reference domain (coin outcomes vs awakening experiences). Specify the index and the paradox dissolves into multiple consistent answers.

When Splitting Isn’t Solving refined this with the sheaf/presheaf criterion. Some indexed answers glue (clean splits, sheaves) and the paradox dissolves. Others don’t (structured splits, presheaves) and the disagreement is structurally forced by a shared generative mechanism. The Coupling Diagnostic — vary the shared parameter vector θ and watch whether indexed outputs move together — is the operational test. Specification is necessary but not always sufficient.

The metric audit on the DR apparatus’s implementation reveals a third layer below this: the indexical parameters themselves are not interchangeable slots. The apparatus implements indexical specification through (P, T, E, S), and the metric audit found that this 4-tuple is processed through a two-hub functional decomposition rather than a flat 4D space. Hub 1 maps observer position to directionality d through power\_role\_heuristic(P, HasBen, HasVic, BaseD) plus marginal exit\_modulation(E), then through sigmoid f, then through scope\_modifier(S) to produce χ. Hub 2 maps (T, E) to mountain or rope through a discrete lookup table. P appears only in Hub 1; T appears only in Hub 2; S appears only in Hub 1; E is the only axis present in both, with marginal Hub 1 contribution and major Hub 2 contribution.

Empirically, Hub 2 captures more between-slice structural variance than Hub 1 under four of five tested metrics. When per-axis predictors are aggregated to hub-level (hub1_diff = 1 if slices differ on P or S; hub2_diff = 1 if slices differ on T or E), Hub 2 partial ρ exceeds Hub 1 partial ρ by 0.16–0.24 under metrics A (extractive fraction), B (type entropy), D (total variation), and E (cover-story flip rate). The exception is metric C (mountain-fraction negative control), where Hub 1 leads slightly. The two hub-level predictors are statistically independent (ρ between them = −0.088). The §4.1 finding from the metric-sensitivity audit that “T and E are consistently top-two across structural-distance metrics” reflects this hub-level decomposition: T and E dominate jointly because they jointly constitute Hub 2.

This sharpens what indexical specification is doing in operational terms. When the apparatus indexes a constraint at a specific (P, T, E, S) position, two functionally distinct subsystems are being queried in parallel, not a single 4D coordinate being looked up. The Type C resolution move “specify the index and get consistent answers” is, when implemented, a request for two different kinds of information: an extraction-magnitude reading from Hub 1 (P-dominated, scope-modulated) and an immutability-classification reading from Hub 2 (T-and-E discrete switch). Whether the indexed answers cohere depends on what each hub returns, and those returns are governed by different mechanisms.

The empirical consequence is visible in the within-rope-group T-axis effect surfaced by the metric audit and confirmed by a covariation robustness check. T-axis dominance under extractive fraction is not a continuous T-axis effect across the corpus; it concentrates in pairs where Hub 2 returns the same output (both rope, n = 112 of 253 non-degenerate pairs), with T partial ρ = 0.577 in those pairs and 0.147 in pairs where Hub 2 returns different outputs. Under the E-fixed restriction (n = 36, with T_diff varying via Tier-2 slices that break Tier-1 collinearity), T partial ρ rises further to 0.762. The mechanism is variation in Hub 1’s χ output across time-horizon values within constraints that Hub 2 has already routed to rope. Time-horizon “matters” empirically by modulating extraction magnitude within the rope-immutability regime, not by modulating the mountain/rope boundary itself.

The diagnostic move “specify the index” doesn’t capture this. The architectural move “require complete indexing at the schema layer” captures it formally but not mechanistically. The granularity move says: which axis you specify shapes which subsystem you’re querying, and the apparatus’s classification behavior depends on the parallel readouts of the two subsystems rather than on any single axis. Type C resolution in an implemented system is therefore a request for coordinated readouts across functionally distinct subsystems, and the architecture of those subsystems determines what kinds of disagreement specification can resolve and what kinds it can only characterize.

This composes with the sheaf/presheaf distinction rather than replacing it. The clean-split case (sheaf) is one where both hubs return information consistent with a single global classification. The structured-split case (presheaf) is one where at least one of the hubs returns information that cannot be reconciled with the other across observer positions. The cohomological obstruction H¹ measures the failure to glue regardless of which hub the disagreement comes from; the hub decomposition tells you, for any given presheaf, which subsystem’s variation is producing the obstruction. That is information the abstract sheaf/presheaf criterion does not surface, and that the audit chain produced because operationalization required tracking each hub’s behavior separately.

—

4. Type A Drift Occurs Between Specification and Implementation

The original Type A account locates drift across reasoning stages within a single agent: the Sorites argument proceeds through “10,000 grains is a heap” → “9,999 grains is a heap” → … → “1 grain is a heap,” where the predicate is reapplied at each stage but the threshold is never declared. The frame is treated as fixed while operationally allowing it to mutate. Frame-fixing — declaring the threshold at t₀ and maintaining it throughout — dissolves the contradiction.

Asymmetry of Failure Types situated Type A architecturally as the failure mode that cannot be exhaustively detected from within the same frame that may itself be drifting. Synchronic discipline, diachronic monitoring, and external review are the appropriate response — three different mechanisms operating at three different positions relative to the frame, not a single detection step.

The audit chain found a Type A drift pattern that neither account anticipated: drift between a project’s specification artifact (the framework paper) and its implementation artifact (the apparatus codebase). v6.11’s Axiom 2 was written χ = ε × f(d(P)) × σ(S(P)). The implementation, in prolog/constraint\_indexing.pl, computes d via power\_role\_heuristic(P, HasBen, HasVic, BaseD) + exit\_modulation(E) for the structural derivation path (used for constraints with explicit beneficiary/victim data — most of the corpus), giving d = g(P, E) rather than d = d(P). Separately, the implementation reads σ as scope\_modifier(S) directly from the context tuple, with no reference to P; the v6.11 notation σ(S(P)) is incorrect.

The discrepancy survived multiple revisions of v6.11 and was not visible to the framework’s readers, the framework’s author, or the other Claude instances that had been reasoning about the apparatus using v6.11’s notation. It was not visible to the apparatus either, because the apparatus matches its own code rather than the paper. The discrepancy was first surfaced by the metric audit’s reconnaissance pass, which inspected the χ formula at the codebase level rather than at the framework level, and it was corrected in v6.11 by an eight-location edit pass after the audit reported it.

This is a Type A drift in the strict sense of the original taxonomy. A frame (the formula notation in v6.11) was treated as fixed while operationally allowing it to mutate (the implementation evolved past the notation as exit_modulation and scope_modifier were added, and the paper notation was not updated to track). The mechanism is the same as the Sorites: a predicate (the formula) reapplied without re-declaring its content. The fix is the same: frame-fix by reconciliation. What is novel is the substrate. The original Type A examples concern drift across reasoning stages within a single agent’s chain of inference. This drift was between artifacts within a single project — between a paper’s specification and a codebase’s implementation, where the paper claimed to describe the codebase and did not.

The substrate matters for what the diagnostic and architectural framings can reach. The diagnostic move “freeze the frame at t₀ and maintain it” presupposes a single agent inspecting their own reasoning chain. It doesn’t address the case where the frame is split across a paper and a codebase, and where each is internally consistent (the paper’s notation is consistent with itself; the codebase’s logic is consistent with itself), but the two have drifted apart. The architectural move “synchronic discipline + diachronic monitoring + external review” partly addresses this: external review is what a paper-vs-implementation audit performs. But Asymmetry described external review as a check on the apparatus from outside its current frame, while what the metric audit performed was a check on a specification artifact from outside the artifact, with the implementation as the ground truth. That’s a slightly different operation — not “is the apparatus drifting from its design intent” but “is the design specification drifting from the apparatus.”

The granularity finding for Type A is that drift across artifacts is a real failure mode that the original taxonomy treats only obliquely. It exhibits the diagnostic signature (frame treated as fixed while operationally mutating) and admits the diagnostic repair (frame-fix by reconciliation). It is governed by the architectural treatment (external review catches what the system cannot inspect from within). But it operates at a substrate level — between papers and code, between specification and implementation, between artifacts of different kinds — that the trifurcation’s standard examples don’t make visible.

There are operational consequences. A project that maintains both a specification document and an implementation should expect Type A drift between them and should plan for it: not as something to prevent (the drift is often the natural consequence of implementation work outpacing documentation), but as something to audit periodically by inspecting the substrates against each other. The metric audit found two notation drifts in v6.11’s Axiom 2 within a single short formula. It is unlikely those are the only drifts in v6.11; they are the ones the audit’s reconnaissance pass surfaced because they sat at the apparatus’s most-touched code junction (the χ formula at the two-hub junction). Other drift may exist in less-inspected parts of v6.11. The audit chain did not look for it. A future audit could.

The deeper point is that Type A drift in implemented systems has a cross-artifact dimension that the trifurcation’s reasoning-stage examples don’t capture. The repair operation (frame-fix by reconciliation) survives unchanged; what changes is the recognition that frames can drift across boundaries within a project, not only within a chain of reasoning. Operationalizing the trifurcation surfaces this because operationalization requires multiple artifacts and multiple loci where the frame can be specified.

4.1 Not All Cross-Artifact Disagreement Is Type A

A subsequent observation, surfaced when reviewing the project’s own orientation document against the source papers it summarized, sharpens what the cross-artifact framing does and does not include. The orientation document cited the corpus as having 3,254 constraints, while a later paper cited 3,301, and a still-later document cited 3,335. Read at the surface, this looks like Type A drift — a frame (“the corpus”) treated as fixed while operationally allowing it to mutate. The repair would be frame-fix by reconciliation: pick a number, propagate it.

That diagnosis is wrong, and the right diagnosis sharpens the granularity finding.

The corpus is not a fixed dataset that occasionally gets revised. It is a continuously-extending artifact: each essay generated using the apparatus writes new constraints into it, on the order of three constraints per essay. The 3,254 → 3,301 → 3,335 progression is not three readings of a stable thing. It is three samples of a moving target, taken at three different times when three different documents were written. Other intermediate values existed at other moments. The number is well-defined at any specific timestamp; it is undefined as a singular value of “the corpus.”

The Sorites pattern is the diagnostic: treating a predicate (“the corpus”) as stable while the underlying state mutates. But the Sorites move generates Type A drift in the original taxonomy because the predicate’s underlying state is implicitly being held fixed. Here the predicate’s underlying state is explicitly mutating — that’s the apparatus’s design, not a failure mode. What’s actually happening is unmarked Type C: “the corpus” is grammatically singular but operationally indexes whatever-the-corpus-was-at-time-t, and the time-index is implicit. Different artifacts cite different time-indexed values of “the corpus” while sharing the implicit assumption that they’re referring to a single object.

The right repair, accordingly, is not Type A frame-fix by reconciliation (which would freeze a number that should not be frozen) but Type C schema-layer index specification: declare the time-index visible. In practice this means each load-bearing numerical claim about the corpus carries its sample timestamp, and audits cite the timestamp of the corpus state they ran on. The orientation document doesn’t pick a corpus number; it points to the source documents and notes that each cites a different timestamp.

This refines the §4 cross-artifact-drift finding. Cross-artifact disagreement in synthesis documents has at least two distinct generative mechanisms, with different repairs. Substantive claim drift (the Axiom 2 case): a single claim diverges between two artifacts that share a frame; repair is reconciliation. Continuous-extension drift (the corpus-count case): different artifacts cite different time-indexed values of an artifact that is monotonically extending; repair is making the time-index visible. Treating the second as the first misapplies the expensive repair operation — Type A reconciliation work — to a problem that needs only the cheap schema-layer move. Treating the first as the second leaves substantive disagreements unresolved.

The deeper observation: when a project produces artifacts that are themselves subject to continuous extension (a corpus that grows with use, a paper sequence that accumulates, an audit chain that accretes findings), some of what looks like cross-artifact Type A drift in synthesis documents is unmarked Type C — implicit shared-frame assumptions across artifacts that don’t actually share a frame because the frames are versioned. The granularity finding for Type A in §4 concerns substantive claims; the operational complement is that numerical and versioned claims about extending artifacts need to be handled at the schema layer, with explicit version-indexing rather than reconciliation. The cost of getting this wrong is asymmetric: misapplying reconciliation to versioned numbers is expensive but not load-bearing for findings; missing reconciliation on substantive claims (the Axiom 2 case) is what produces invisible drift between specification and implementation.

The corresponding operational practice for the DR project: pipeline output JSONs carry timestamps; audits cite the pipeline state they ran on by timestamp; cross-audit synthesis verifies the comparability of pipeline states before composing findings. None of this is recursive governance. It is one schema-layer move that converts an apparent Type A problem into a tractable Type C problem.

—

5. Type B Axiomatic Consistency Has a Scope-Design Dimension

The original Type B account treats axiomatic inconsistency as a property of the formal system: Russell’s R = {x | x ∉ x} produces R ∈ R ↔ R ∉ R immediately, regardless of process or perspective, and the repair is system-level revision (ZFC’s restriction on set formation). The diagnostic move is the structural check: the contradiction is in the formal vocabulary the apparatus already speaks, and the apparatus can detect it because detection operates in the same language as the system’s claims.

Asymmetry of Failure Types developed this architecturally: Type B detection is appropriate at multiple measurement layers because structure can be inspected at several depths, and a single detector at one layer would miss what detectors at other layers catch. The DR apparatus’s multi-layer Type B coverage (Boltzmann compliance, structural signatures, False CI Rope, False Summit Mountain, False Natural Law) is the right response to a class of failure that can present itself in more than one formal guise.

The sheaf audit on the 10-slice working family revealed a third aspect that neither framing captured: axiomatic consistency, in implemented systems, has a scope-design dimension. The axioms can hold on the design domain and produce binary-classification crossings on adjacent domains that someone might reasonably mistake for the design domain.

The framework’s primary empirical commitment is the binary site-stability claim: H¹ = 0 vs H¹ > 0 classification is preserved with zero crossings between the canonical 4-point site and the 156-point product-site expansion. The sheaf audit tested whether the same preservation holds on a third site, the 10-slice Tier-1 working family selected for population coverage rather than as a superset of the canonical site. Result: 68.98% binary crossing rate (1,940 sheaf→presheaf crossings of 2,834 working-set constraints, with 15 presheaf→sheaf crossings as variable-subsite artifacts).

The mechanism is identifiable and specific. Of the 1,940 sheaf→presheaf crossings, 91.0% (1,766) are driven by the 10-slice family’s U_4 context = (analytical, civilizational, analytical, universal). The canonical site’s analogous context, U₄ = (analytical, civilizational, analytical, global), differs in scope. Scope modifier values are σ(global) = 1.2 and σ(universal) = 1.0. At the analytical observer position, where canonical d(analytical) ≈ 0 and the sigmoid f(0) is at the extreme low end of its range, the residual scope multiplier σ is what determines whether χ exceeds the rope_chi_ceiling threshold. With σ = 1.2 (global), constraints with extraction-adjacent ε classify as tangled_rope or snare. With σ = 1.0 (universal), the same constraints, with the same ε and the same d, produce χ at or below rope_chi_ceiling and reclassify as mountain. Of the 1,766 U_4-driven crossings, 1,758 produce mountain at U_4 — exactly the predicted shift.

The product site’s construction explicitly excludes universal scope. The site_contexts_product/1 predicate carries a comment: “Excluded scope values: regional, continental, universal.” The exclusion has no published rationale in v6.11, but the scope-modifier mechanics make it clear: universal scope at the analytical position would produce mountain for any constraint near the rope_chi_ceiling, creating precisely the crossings the sheaf audit found. The canonical site similarly avoids universal scope. Both the canonical and product sites are scoped to keep σ ≥ σ(global) at the analytical observer position. The 10-slice family’s U_4 violates this constraint because the family was selected for population coverage (the analytical/civilizational/analytical/universal slice is the most-populated single slice in the corpus, with 2,543 constraints) rather than for axiomatic consistency with the design site.

The framework’s site-stability claim is not falsified. The canonical-to-product result remains correct on its own terms; the apparatus does preserve the binary boundary across that expansion. What the sheaf audit established is that the claim holds only on sites respecting the canonical scope-design constraint: σ ≥ σ(global) at the analytical observer position. Sites that include universal scope at analytical produce systematic mountain reclassification of extraction-adjacent constraints.

This is not axiomatic inconsistency in the original Type B sense. The framework’s axioms are not self-contradictory. The DR formalism does not derive both Mountain and not-Mountain for any single constraint at any single observer position. What the sheaf audit revealed is a distinction the original Type B account elided: in implemented systems, “axioms hold” and “axioms produce well-defined output” are not the same condition. The first is a property of the formal system. The second is a property of the formal system on a specified domain. The framework’s site-stability claim is meaningful only relative to its design domain, and identifying the design domain explicitly was not part of v6.11’s site-stability claim until the sheaf audit forced it.

The granularity finding for Type B is that axiomatic-consistency claims in implemented systems have a scope-design dimension, and the appropriate repair when the scope is violated is not axiom revision (the original Type B move) but domain specification (state explicitly which sites the axioms produce well-defined output on, and reject sites that violate the design constraint). The repair operation is structurally between Type B and Type C: like Type C, it’s about specification (specifying the design domain rather than revising the axioms); like Type B, it’s about the formal system rather than the input query. The apparatus does not currently enforce this repair; the sheaf audit’s revised site-stability claim (“scope ≤ global at the analytical observer position”) is testable, and any future site predicate added to the codebase could be checked against it programmatically. That’s a small implementation move that hardens the apparatus against the failure mode the audit surfaced.

The deeper point is that the original Type B account’s framing — “the contradiction is in the formal vocabulary the apparatus already speaks” — assumes the apparatus is operating on a domain where the vocabulary applies. When the apparatus is given inputs from an adjacent domain (a site that crosses a design constraint), the apparatus continues to compute, but its outputs no longer carry the empirical guarantees the design domain provides. The axioms hold; the empirical claim does not generalize. Distinguishing these is a third Type B repair that the original taxonomy doesn’t make explicit.

—

6. Synthesis: What Operationalization Reveals

The trifurcation framework distinguishes three failure modes by their generative mechanism: Type A (drift across stages), Type B (axiomatic inconsistency), Type C (indexical underspecification). Asymmetry of Failure Types extended this architecturally: the three types correspond to three different relations between failure and formalism, and validation is a family of operations rather than a single one.

The audit chain on the DR apparatus extends both framings with internal architecture for each type:

• Type C indexical specification has functional decomposition. The (P, T, E, S) index in the implemented apparatus is processed through a two-hub architecture (P, S → χ via Hub 1; T, E → mountain/rope via Hub 2), with Hub 2 capturing more between-slice variance than Hub 1 under most metrics. Specification is not a single move but a request for coordinated readouts across functionally distinct subsystems, and the architecture of those subsystems shapes what kinds of disagreement specification can resolve.
• Type A drift has a cross-artifact dimension, with two distinct subclasses. The apparatus’s Axiom 2 notation drift between v6.11’s specification and the implementation’s code is a Type A failure in the strict sense — frame treated as fixed while operationally mutating — but it occurs between artifacts within a project rather than across reasoning stages within a single agent’s chain. The repair operation (frame-fix by reconciliation) is unchanged; the substrate is novel. A separate subclass — apparent Type A drift over numerical or versioned claims about continuously-extending artifacts — is properly Type C with a schema-layer time-index repair, not Type A with reconciliation (§4.1). Distinguishing the two prevents over-applying expensive governance to cheap problems.
• Type B axiomatic consistency has a scope-design dimension. The apparatus’s site-stability claim survives with a precise qualification (scope ≤ global at the analytical observer position) rather than as universal, and the qualification was invisible until tested on a site that violated the design constraint. The repair is domain specification rather than axiom revision; the framing falls structurally between Type B (about the formal system) and Type C (about specification).

These are not three new failure types. They are internal architecture within the existing three. Type C’s functional decomposition is still indexical underspecification’s repair domain — the move is still “specify the index” — but the move has texture that the abstract framing didn’t capture, and the texture matters operationally for what specification can claim to have accomplished. Type A’s cross-artifact drift is still drift — the move is still “frame-fix by reconciliation” — but the substrate (papers vs code) is one the original examples didn’t make visible. Type B’s scope-design dimension is still about the formal system and its consistency, but the original framing assumed axioms either hold or fail to hold without recognizing that “hold on the design domain” is a meaningful third possibility.

The synthesis is that the trifurcation’s repair operations are richer than the diagnostic taxonomy made apparent, and the architectural framing (Asymmetry) is a step toward seeing that richness. The audit chain disclosed the next layer: each repair operation has internal structure that becomes visible only when the apparatus is mechanical and the artifacts are multiple. The taxonomy holds; the asymmetry holds; the granularity is what implementing both produces.

A concrete consequence: a system that claims to operationalize the trifurcation cannot rely on the abstract repair operations alone. It needs to know which subsystems its indexing exercises (Type C granularity), to maintain reconciliation between its specification artifacts and its implementation artifacts (Type A granularity), and to specify the design domain on which its axiomatic claims produce well-defined output (Type B granularity). These are three additional architectural commitments beyond what the diagnostic and architectural framings demanded. The DR apparatus partially makes them: the two-hub architecture is implemented, the Axiom 2 reconciliation is now in place, and the scope-design constraint is articulated though not yet programmatically enforced. A successor apparatus would benefit from making them explicit at the design layer rather than discovering them by audit.

—

7. What the Audits Did Not Extend

Three things the audit chain explicitly did not modify in the trifurcation framework, surfaced for clarity.

First, the diagnostic order is preserved. The order Type C first, then Type A, then Type B — established in Debugging Philosophy and refined in Asymmetry of Failure Types — is unchanged by the audit findings. Specification before frame-fixing before structural check remains the right sequencing because each move’s input requires the prior move’s output. The granularity findings refine what each move consists of without reordering them.

Second, the trifurcation is not made finer-grained at the type level. There are still three failure modes, not five or seven. What the audits found is internal architecture within each type, not new types. A failure that exhibits cross-artifact drift is still Type A. A failure that exhibits hub-decomposition disagreement is still Type C. A failure that exhibits scope-design violation is structurally Type-B-adjacent but its repair operation is closer to Type C’s specification move; the synthesis paper notes this without proposing a separate Type B’ or Type BC.

Third, the trifurcation’s relationship to Metrics as Routing is not changed. Framework metrics remain governance stands rather than truth measurements. The two-hub architecture’s Hub 1 and Hub 2 are routing apparatus, not measurement layers — they decide which classification path to take, not what is objectively true about a constraint. The site-stability qualification doesn’t make σ values more truth-bearing; it specifies the design domain on which the routing produces stable outputs. Type B’s domain-specification repair is still a governance move (declaring where the system’s outputs are reliable) rather than a truth-measurement move.

The audits sharpened the trifurcation’s repair operations without changing what the trifurcation is.

—

8. Forward

The synthesis paper closes the audit chain that began with BC coupling. The findings now sit in three documents — coupling\_structure\_evidence.md, metric\_audit\_writeup.md, and the sheaf audit writeup — with the v6.11 framework paper updated for the Axiom 2 notation drift and the construct-distinction note. This paper composes those findings with the prior trifurcation literature (debugging\_philosophy.md, asymmetry\_of\_failure\_types.md) and the Type C extension (when\_splitting\_isnt\_solving.md).

What this paper does not do is propose a successor framework, a major revision of v6.11, or new axioms for the apparatus. The audits established that v6.11’s empirical commitments survive with refinements, and the synthesis paper articulates what the refinements consist of. A reference document orienting a new model to the apparatus — what it does, how to think about it, what its findings depend on — is the natural next step but is operational documentation rather than intellectual contribution.

Two engine extensions are warranted by the audit chain and remain unimplemented. A scope-design validator on site\_contexts/N predicates would catch the failure mode the sheaf audit surfaced before any future site predicate is added that violates it. MaxEnt parameterization for arbitrary sites would allow Arakelov fragility to be computed on alternate sites and unlock the deferred sub-question from the sheaf audit. Neither adds new findings; both harden the apparatus against rediscovering what the audits already established.

The audits’ larger finding, the one that runs underneath all three granularity claims, is that operationalizing a diagnostic taxonomy discloses architecture the diagnostic taxonomy doesn’t see — and that the architecture matters for what the diagnostic can claim. The trifurcation as a diagnostic move (specify the index, fix the frame, revise the axioms) is correct as far as it goes. The asymmetry framing extends it to architecture (detection, prevention, governance). The granularity framing extends it to internal structure within each repair operation. Each layer is a necessary refinement of what came before, and each was discovered by building rather than by deduction. The discipline is to extend the framework as the apparatus reveals what the framework needs to say, rather than predicting in advance what implementation will require.

—

References

Internal to the project:

• Debugging Philosophy: A Trifurcation Framework for Paradox Classification. Establishes the original Type A/B/C taxonomy and the diagnostic-test structure this paper extends.
• The Asymmetry of Failure Types. Develops the trifurcation architecturally: detection for Type B, prevention for Type C, governance for Type A. The synthesis paper composes with this rather than re-derives it.
• When Splitting Isn’t Solving: Sheaves, Presheaves, and the Structure of Indexical Disagreement. The Type C extension: specification produces clean splits (sheaves) or structured splits (presheaves), and the cohomological obstruction H¹ measures the difference.
• Deferential Realism: A Presheaf Framework for Observer-Dependent Classification (v6.11). The framework paper. Source for the canonical site, the dimensional hierarchy claim, the product-site expansion, and the binary site-stability claim. Updated for the Axiom 2 notation drift and the construct-distinction note based on audit findings.
• Metrics as Routing: Why Thresholds Are Governance Stands, Not Truth Claims. The foundation document on framework metrics. The synthesis paper preserves rather than modifies the metrics-as-routing claim.
• Coupling Structure and Position-Space Geometry: An Evidence Document. Audit infrastructure and findings on BC coupling and position-space metric sensitivity. §4.4 contains the implementation-adjacent findings the synthesis paper draws on.
• When the Site Changes the Boundary: Scope Modifier Mechanics and the Limits of Site-Stability (sheaf audit writeup). Source for the §5 scope-design finding.
• Metric Audit (metric\_audit\_writeup.md, plus metric\_audit\_recon.md, metric\_audit\_proposal.md, metric\_audit\_results.{md,json}, audit3\_te\_robustness.{md,json}). Source for the §3 two-hub architecture finding and the within-rope-group T-axis effect.
• When Metrics Aren’t Measurement: Cluster-Space Architecture and the Limits of Signature-Pathway Detection. Companion paper on the metric layer’s architecture; cited as background.
• Observers, Not Humans: Deferential Realism Across Observer Classes (v5). Background for the universality-class framing of structural claims.

—

Working paper. The audit chain that produced the synthesis was conducted in dialogue with Claude (Anthropic) instances across multiple conversation contexts. The synthesis itself emerged from the recognition that the audit findings were no longer producing new architectural results and were instead refining the trifurcation’s repair operations in ways that had a coherent shape across all three types.