Skip to main content

66  Mathematical and Search Substrates

66.1 Chapter status

Field	Value
Chapter ID	mathematical-and-search-substrates
Part	Part III - Routing, Compression, Representation, and Substrates
Status	conceptual
Manuscript maturity	v0.3 manuscript draft
Last updated	2026-08-08
Primary source records	genesiscode, temporal_coil_research, cognitive_compilation, treellm, simulation_scaling, coilmoecot, circle_calculus_core, circle_ai_architectures, proof_carrying_circular_computation, theseus_circle_transfer, ext_mamba_2023, ext_universal_transformer_2019, ext_recurrent_transformer_2026, ext_v_jepa_2_2025
Claim label	Design rationale
Evidence level	argument
Source queue	primary: genesiscode, temporal_coil_research; supporting: cognitive_compilation, treellm, simulation_scaling, circle_calculus_core, circle_ai_architectures, proof_carrying_circular_computation, theseus_circle_transfer; variants/external literature: ext_mamba_2023, ext_universal_transformer_2019, ext_recurrent_transformer_2026, ext_v_jepa_2_2025; connector/recovery: coilmoecot
Source loading state	source notes: genesiscode, temporal_coil_research, cognitive_compilation, treellm, simulation_scaling, coilmoecot, circle_calculus_core, circle_ai_architectures, proof_carrying_circular_computation, theseus_circle_transfer, ext_mamba_2023, ext_universal_transformer_2019, ext_recurrent_transformer_2026, ext_v_jepa_2_2025; raw cache: genesiscode, temporal_coil_research, cognitive_compilation, treellm, simulation_scaling, coilmoecot
Test state	substrate_adoption_record.valid.json passes repository-level protocol fixture validation for baseline obligations, consumer gate/policy, axis ledger, routing-permission effect, fallback substrate, retirement/supersession path, support-state effect, and non-claims; python3 scripts/validate_substrate_adoption_trace.py independently derives four synthetic adoption states and eight expected-invalid controls while preserving no-promotion boundaries; AsiStackProofs.SearchSubstrates implements nineteen declarations spanning finite record countermodels, a reachable trace classifier, twelve exact route witnesses, and route-permission algebra; baseline comparison, representation-efficiency, sequence-substrate A/B, and real falsification tests remain planned.

66.2 Drafting guardrail

An elegant substrate has no automatic standing in the architecture. Coils, calculi, geometric search, cyclic compute, and state-space sequence backbones remain optional specialist candidates until a bounded adoption record names their task, comparator, falsifier, consumer, and retirement path. No model- quality, compression, search, or runtime benefit follows without those baselines and artifacts.

Simulation claims are bounded by scope, fidelity, and resources. Mathematical substrates need the same discipline: an elegant representation is not a right to become architecture. A substrate enters the ASI Stack through an adoption record that states intended use, baseline, falsification condition, proof boundary, and evidence gate before it can become a route, representation, or compression dependency.

The adoption record is a routing permission, not a trophy. It says where a substrate may be tried, what it must be compared against, and what evidence would retire, narrow, or promote it.

66.3 Human Reading Path

Concrete lens. The simpler baseline treats a theorem or interesting architecture as general adoption evidence. The chapter keeps the ordinary substrate active and grants only the exact consumer axis that has passed its own baseline and control.

After bounding simulations, the stack turns to optional mathematical and search substrates. Elegant calculi, geometric search spaces, coils, cyclic compute, and state-space structures may be useful, but elegance alone does not make them part of the stack.

The governing rule is simple: a substrate enters through an interface, not through fascination. It needs a task boundary, baseline, negative control, evidence packet, failure mode, and reason why it improves the governed system rather than merely adding a beautiful internal story. The value is disciplined optionality: keeping promising ideas available without letting them silently become architecture.

Optionality has value only while the default path remains honest and available. The stack can explore unusual mathematics without letting novelty become permission, but each elegant substrate still has to answer what work it improves, what risk it adds, and what result would stop adoption. Search becomes architectural when its advantage is bounded by a task. Promotion comes through comparison, not aesthetic force, and comparison must leave failure visible enough for the next candidate to learn from it.

66.4 Problem

The source corpus contains many potentially useful substrates: semantic IR, small deterministic kernels, temporal coils, cyclic address laws, proof-carrying receipts, TreeLLM-style semantic graphs, and routed specialist lanes. They need an evaluation lane, not a permission slip for overclaiming.

Adoption discipline is the central substrate constraint. A substrate can be mathematically clean and still useless for a workload. A proof can establish address safety while saying nothing about speed. An experiment plan can be well designed while producing no positive result. The stack must keep exploratory substrates available without making them core until evidence warrants it.

That discipline matters because substrate ideas are seductive. A new calculus, search structure, graph representation, or cyclic mechanism can make the architecture feel deeper before it makes any task better. Candidate substrates need a comparative holding area where they can be constrained, rejected, or carried forward as optional specialists without becoming hidden commitments.

66.5 Why existing approaches are insufficient

Novel substrates can become authority theater if they are not tied to baselines, adoption gates, and falsification criteria. Opaque math can be treated as proof. Single-seed experiments can be treated as discovery. A structural theorem can be laundered into a performance claim. A specialist lane can become load-bearing before its readiness gates exist.

GenesisCode supplies the small-kernel and obligation discipline. Temporal Coil Research supplies manifest-driven A/B testing. Circle Calculus supplies proof-status boundaries. Mamba supplies an external example of a sequence-substrate efficiency claim that still needs task-specific adoption evidence. The Theseus transfer lane supplies the rule that deterministic structural fixtures are not model-quality evidence.

The adoption filter is where attractive mechanisms wait until the book can say what they are for, what ordinary method they replace, what negative control would embarrass them, and what residual would remain even after a favorable result. That filter keeps the stack open to unusual mathematics without letting novelty become authority.

The filter also prevents cross-axis leakage. A substrate might improve a representation task while making routing worse. It might reduce active work while increasing verification burden. It might have a proof-relevant finite structure while offering no downstream quality benefit. The adoption record keeps those axes separate so the book can use a substrate in one narrow lane without implying it should become a general backbone.

Real mathematics can make substrate leakage harder to notice. A correct theorem can make the surrounding architecture feel more justified than it is. The theorem should travel only as far as its statement, model, assumptions, and proof artifact permit.

66.5.1 Strongest objection

Optional-substrate framing can become a parking lot for unfalsifiable novelty: every unusual mechanism survives because it is never made load-bearing enough to fail. The adoption gate answers only if it forces a named baseline, a task, a budget, a negative control, and a retirement condition; structural elegance alone is not a reason to keep a substrate.

66.6 Core Claim

Reader claim. Elegant mathematics earns exactly the authority its evidence supports. A structural proof may justify a diagnostic lane while leaving search quality, downstream task quality, cost, and adoption entirely open.

Operational rule. Give every substrate a named ordinary baseline, workload, consumer, requested claim axis, negative control, falsifier, fallback, and retirement path. Only measured passing axes may change routing permission; an unmeasured axis, failed control, or theorem outside its statement blocks or retires adoption.

[mathematical-and-search-substrates.core, label: Design rationale, support: argument] Mathematical and Search Substrates owns a consumer-, use-, workload-, claim-axis-, implementation-, baseline-, resource-, and time-specific Substrate Adoption Lease: an unusual calculus, representation, recurrence, search procedure, latent world model, or sequence backbone may affect only the axes and consumers that pass matched ordinary and current baselines, negative controls, complete cost and rights accounting, falsification, fallback, independent reproduction, and transfer; structural elegance, a theorem, a source-reported benchmark, synthetic fixture validity, or one favorable axis alone confers no general quality, efficiency, safety, support, deployment, or SOTA authority.

The distinct owner is the adoption boundary between a candidate substrate and a named consumer. That boundary does not own the consumer objective, plan semantics, context fidelity, route decision, decoding latency, deliberation budget, resource ledger, theorem truth, security or rights decision, runtime binding, readiness state, incident response, release, or public claim. It owns the least-authority lease that says which exact substrate version a consumer may use for which evidenced axis, under which workload, implementation, baseline, budget, expiry, fallback, and non-claims.

Substrate source mappings ground the discussion of candidate calculi, proof boundaries, manifests, A/B variants, sequence-substrate context, and adoption discipline. They do not promote any substrate to core architecture.

66.6.1 Claim-source mapping status

Appendix C records exact passage-reviewed mappings for all fourteen assigned sources. The mappings support optional substrate candidates, proof boundaries, manifests, A/B discipline, semantic IR, graph and latent-prediction substrates, cyclic contracts, transfer boundaries, state-space and recurrent sequence context, and adoption gates. They do not promote a core substrate, model-quality result, performance result, local Circle proof ownership claim, Theseus transfer result, or ASI Stack reproduction.

Source	What it supports	Limit
genesiscode	Tiny deterministic calculus, effect boundaries, semantic patches, replayable logs, obligation artifacts, provenance hashes, and capability-mediated execution.	No GenesisCode implementation, proof artifact, replay checker, benchmark, or security audit exists here.
temporal_coil_research	Manifest-driven A/B discipline plus a source-reported 11-variant, three-seed, six-round campaign whose small differences were dominated by a collapse composite while task-relevant lanes stayed flat.	Inconclusive source-reported result only; no local reproduction, independent evaluator, clean causal component effect, transfer, or general benefit/failure conclusion.
cognitive_compilation	Substrates as compiler/IR options with source plans, semantic atoms, dependency analysis, validation requirements, target lowering, localized repair, and evaluator-backed artifacts.	No working cognitive compiler, trace suite, or empirical ablation exists here.
treellm	Semantic-graph and path-token substrates as optional representation/search candidates with traversal, residual attributes, editable graph updates, and shared semantic OS framing.	No local TreeLLM implementation, measured compression ratio, reasoning benchmark, or verified token format exists.
simulation_scaling	Resource-bounded substrate adoption through scope, fidelity, temporal semantics, demand, capacity, and bottleneck assumptions.	No simulation benchmark, feasibility calculator, or independent literature audit was run.
coilmoecot	Routed cyclic/specialist lanes, Graph/Trace-first placement, explicit insertion points, removable shadow mode, ledger-derived diagnostics, and anti-experts as bounded negative-signal evidence.	Design/spec source only; no CoilMoECOT benchmark, route run, ablation, canary, rollback, or performance evidence has been reproduced.
circle_calculus_core	Proof-boundary discipline for cyclic addressing, phase, recurrence, coverage, and contract receipts through theorem manifests and evidence layers.	A separate external Circle rope receipt slice is recorded in docs/circle_external_receipt_slice.md, but this row remains proof-boundary context, not local proof ownership, substrate adoption evidence, or model-quality evidence.
circle_ai_architectures	Cyclic-substrate adoption only where phase, recurrence, rotation, sparse cyclic mixing, circular memory, harmonic transforms, or geometry-aware structure is real and baselines/negative controls exist.	A separate external Circle rope receipt slice is recorded in docs/circle_external_receipt_slice.md, but no cyclic-substrate sidecar tests, MLX experiments, benchmark fixtures, or model-quality results were run.
proof_carrying_circular_computation	Cyclic compute as a proof-plus-benchmark program: cyclic address primitives, stride coverage, rewrite/address transformations, backend selection, and baseline benchmarking.	No Circle sidecar examples, CoilIR backend, or backend benchmarks were run; the separate external rope receipt slice does not validate cyclic-compute backend adoption.
theseus_circle_transfer	Transfer boundaries for deterministic Circle fixtures into private Theseus benchmark design, with workload, baseline, negative-control, metric, script, report, and non-claim requirements.	No Circle-to-Theseus consumer, smoke workload, proxy benchmark, inference run, or private-result import was executed.
ext_mamba_2023	Selective state-space and recurrent sequence models as optional sequence substrates whose efficiency claims must be separated from decoding tricks, verifier adequacy, and downstream task quality.	No Mamba model was trained, served, benchmarked, routed through governance gates, or compared against local baselines here.
ext_universal_transformer_2019	Historical comparator for shared-weight depth recurrence, parallel self-attention, and per-position adaptive halting across structural, compute, stopping, and utility axes.	No model, checkpoint, benchmark, stability, efficiency, scaling, or local reproduction result is imported.
ext_recurrent_transformer_2026	Current comparator for layerwise recurrent KV memory, exact tiling, effective-depth/width tradeoffs, and autoregressive decoding cost.	No checkpoint, cache implementation, kernel, hardware, quality, scale, production, or local reproduction result exists here.
ext_v_jepa_2_2025	Latent-prediction comparator separating action-free representation learning, action-conditioned prediction, planning use, and downstream evidence axes.	Source-reported video and robot results do not establish local quality, causal understanding, safe control, transfer, deployment, or an ASI Stack result.

66.7 Mechanism

66.7.1 Worked adoption trace: one coil, four honest dispositions

The local substrate trace follows substrate://temporal-coil-demo against baseline://ordinary-dense-or-learned. With only a structural boundary and a planned search test, the candidate is exploratory and receives planning-only permission. When a consumer asks only for the represented structural axis, the same candidate becomes structural-only and diagnostic-only. When a router asks for unmeasured routing quality, adoption is blocked and the ordinary fallback stays active. Finally, when the declared wrong-period/nonperiodic control erases the search advantage, the search axis becomes measured-negative and the candidate is retired.

These are four synthetic dispositions, not four substrate experiments. Their value is that the axis ledger changes the decision without changing the candidate’s glamour. The validator also rejects eight nearby shortcuts: missing baseline or falsifier, theorem spillover, routing on an unmeasured axis, promotion after a failed control, missing fallback, support promotion, and a missing non-claim boundary. The trace therefore shows how the stack can remain open to coils, calculi, state-space models, and new search methods while making “interesting” strictly weaker than “adopted.”

Substrate adoption is a promotion protocol, not a catalog entry. GenesisCode keeps proposals behind a small-kernel, capability, replay, and obligation boundary. Temporal Coil Research says cyclic mechanisms need manifest-driven A/B variants and multiseed readouts before they are believed. Cognitive Compilation and TreeLLM show how compiler IR and semantic graphs can become candidate substrates without becoming authority sources. Simulation Scaling and Mamba keep resource and sequence-backbone claims separate from downstream quality, while Circle and CoilMoECOT keep cyclic specialists optional until proof boundaries and benchmark gates are real.

The Temporal Coil record also contains a narrow, inconclusive experiment rather than only a plan. Across 11 variants, three seeds, and six rounds per variant, the reported winner split among coil-off, progressive-FFT, and progressive conditions. The largest reported mean deltas were small—adaptive +0.002935, no-hints +0.002754, and reward-only +0.002527—while pass, reward, and holdout lanes stayed flat and most separation came from a collapse composite. One threshold-tuned adaptive seed reached +0.007057, but that is neither a multiseed effect nor a clean causal comparison. Smoke runs showed wiring, not capability. The correct conclusion is therefore not “coils work” or “coils fail.” The workload did not distinguish mechanism, placement, hinting, reward, and scoring effects well enough. A replacement experiment needs a trace-native task, placement ablations, stronger simple baselines, matched compute, more seeds, and an independent evaluator.

A Substrate Adoption Record turns that discipline into a routing object. It asks what the substrate is for, what baseline it must beat, what would falsify it, what proof or structural facts are actually available, which consumers may read it, which claim axes remain unmeasured, what routing permission follows, which ordinary substrate remains the fallback, and whether the correct state is exploratory, structural-only, canary, qualified-for-scope, retired, superseded, or blocked.

The important move is from fascination to accountability. A coil, calculus, state-space layer, graph substrate, compiler IR, or cyclic backend is not evaluated by how suggestive it feels. It is evaluated by a declared workload and by the failure cases that would stop adoption. If the substrate only has a structural proof, the record says structural-only. If it has a benchmark plan but no run, the axis ledger says planned or unmeasured. If it fails a baseline, the failure becomes part of the research memory instead of disappearing into the next promising variant.

CoilMoECOT makes placement part of the substrate hypothesis. Its safest order is Graph/Trace first, Concept second, and Token last: earlier trace layers have cleaner events, cheaper baselines, and more intelligible failure attribution. The coil remains a removable lane, starts in shadow mode, and can enter as a pre-planner risk prior, post-plan risk scan, or post-run distillation feature. Negative-signal specialists must emit inspectable penalties with confidence, affected candidates, and override policy; they do not receive a hidden veto. No weight update may escape the declared slice, and the non-coil path remains the recovery route.

flowchart LR
  A["Candidate substrate"] --> B["Expected advantage"]
  B --> C["Ordinary baselines"]
  C --> D["Falsification condition"]
  D --> E["Experiment or proof boundary"]
  E --> F["Axis ledger and consumer gate"]
  F --> G{"Evidence sufficient for this consumer?"}
  G -- "yes" --> H["Canary specialist route"]
  G -- "no" --> I["Exploratory, structural-only, or residual"]
  H --> J["Readiness gate"]
  I --> J
  J --> K["Research backlog or canary plan"]

What the substrate adoption gate shows: Substrate adoption starts with an expected advantage, baselines, falsification condition, and proof or experiment boundary. The consumer gate decides whether evidence is sufficient for a canary route or whether the substrate remains exploratory, structural-only, or residual.

A substrate adoption record separates:

• Substrate identity and intended use.
• Expected advantage.
• Required baselines and negative controls.
• Baseline obligations such as shared workload, metric definitions, data conditions, hardware notes, and fallback policy.
• Proof or structural facts, if any.
• Workload, metric, script, and report requirements.
• Consumer gate, consumer policy, and routing-permission effect.
• Axis ledger for structural, speed, memory, routing, compression, search, verifier, and downstream-quality claims.
• Falsification condition.
• Adoption decision, fallback substrate, retirement or supersession path, support-state effect, and residuals.
• Explicit non-claims.

This lets the architecture keep a library of promising substrates without turning the library into a mythology.

TreeLLM supplies a useful matched substrate family because its own drafts move among hard DAG traversal, HLSH coordinates, ANN soft-linking, learned routing, recurrent navigation, adaptive short-id caching, and multi-hop speculative prefetch. The adoption record must not bundle them into one “graph intelligence” treatment. Exact graph lookup, ordinary knowledge-graph retrieval, dense vector search, hybrid retrieval, GraphRAG-style assembly, learned navigation, HLSH/path tokens, and speculative traversal are separate candidates with matched information and I/O budgets. Hard-edge, no-residual, random-coordinate, fixed-versus-evolving-anchor, no-cache/conventional-cache, no-speculation, and oracle-route controls isolate the claimed mechanism.

The causal signature must match the story. If HLSH matters, semantic neighborhoods should survive held-out senses and graph updates better than random or ordinary identifiers. If structured tokens matter, a matched model should improve consumer utility or reduce total burden after first-use, residual, cache, verification, and fallback costs. If soft-linking matters, it should recover missing edges without amplifying false neighbors. If speculation matters, useful tail latency should improve after wrong-path reads and privacy/epoch checks. If routing matters, learned choices should beat a strong default and approach an oracle without collapsing to one lane. None of these results exists locally; the complete lineage remains an exploratory candidate with explicit retirement conditions.

Optionality is a positive design state. It means the stack can route experiments toward a substrate while keeping default behavior on ordinary baselines. It also means a future implementation can test a substrate without changing the book’s core architecture claim. The substrate becomes load-bearing only after evidence, not before.

Adoption states should therefore be operational: exploratory, structural-only, canary, qualified-for-scope, blocked, retired, or superseded. A state that does not change routing, testing, consumer access, or evidence obligations is a label, not governance. The schema uses the same states as machine-readable values such as structural_only and qualified_for_scope.

66.7.2 Adoption packet lanes

The adoption packet has four lanes:

Lane	Required question	Promotion blocker
Structural lane	What fact, proof receipt, or deterministic fixture exists?	Missing theorem, unresolved receipt, or model mismatch.
Empirical lane	What workload, metric, baseline, negative control, and report exist?	Missing workload evidence or failed controls.
Consumer lane	Which downstream use may rely on the record?	Consumer asks for an axis the record does not support.
Retirement lane	What result narrows, supersedes, or retires the substrate?	No falsification condition or hidden negative result.

66.7.3 Eighteen-stage adoption lifecycle

The lease is a lifecycle rather than a one-time benchmark:

1. Register the substrate, version, implementation, provenance, rights, mathematical object, and executable surface without route authority.
2. Bind named consumers, intended uses, natural workloads, environments, and separately typed claim axes.
3. Freeze ordinary, strongest-current, simple-ablation, and embarrassing controls under matched information and resources.
4. Preregister the advantage, causal signature, falsifier, narrowing rule, retirement trigger, and evidence horizon.
5. Package structural results with statements, assumptions, proof receipts, implementation relations, consumer bounds, and non-claims.
6. Build an inspectable candidate and executable ordinary fallback with reproducible parameters, kernels, caches, and search budgets.
7. Run overclaim mutations for missing baselines, theorem spillover, unmeasured axes, failed controls, hidden budgets, unavailable fallback, rights, and support state.
8. Execute matched natural workloads across seeds, difficulty, lengths, modalities, and adversarial cases.
9. Preserve structural validity, representation fidelity/rate, search, routing, reasoning/task utility, latency, memory, energy, verifier, safety, and rights as separate axes.
10. Account jointly for training, inference, search, compilation, verification, storage, transfer, operator, debugging, fallback, recovery, and opportunity burdens.
11. Evaluate attacks, pathological structures, shift, privacy, licensing, inaccessible states, and evaluator dependence.
12. Issue the least-authority consumer gate for the exact evidenced axis, workload, environment, budget, expiry, and prohibited inference.
13. Canary beside the executable ordinary route with bounded traffic, live comparators, regression alarms, and no irreversible dependency.
14. Monitor drift, tails, subgroup regressions, verifier disagreement, displaced resources, and downstream residuals.
15. Fallback, quarantine, rollback, narrow, supersede, refute, or retire on failed evidence, controls, budgets, rights, or recovery.
16. Reproduce with an independently implemented substrate and evaluator that do not share the decisive defect.
17. Transfer across consumers, workloads, models, scales, hardware, data regimes, organizations, rights regimes, attacks, and time.
18. Expire on material change and assign every theorem-model gap, hidden cost, failed gate, transfer failure, and residual to an owner and reopening condition.

This order makes “optional” falsifiable. A candidate without a dated next run, narrowing decision, retirement decision, or blocked-after-full-attempt record is not a research option; it is an indefinitely protected story.

66.8 Interfaces

Substrate adoption moves through the Substrate Adoption Record.

Minimum fields:

• substrate_id
• substrate_kind
• intended_use
• expected_advantage
• baseline_refs
• baseline_obligations
• negative_controls
• proof_boundary
• experiment_requirements
• consumer_gate
• consumer_policy
• axis_ledger
• falsification_condition
• adoption_state
• routing_permission_effect
• fallback_substrate
• retirement_or_supersession_path
• support_state_effect
• residuals
• evidence_refs
• non_claims

Routing treats a candidate substrate as a specialist. Compression tests representation efficiency. Evidence compares against baselines. Proof envelopes preserve theorem status. Fast-generation routes may consume substrate evidence, but adoption remains gated by substrate-specific A/B records. Readiness gates decide whether a substrate is exploratory, canary, qualified, or retired.

The Substrate Adoption Record carries an axis ledger for structure, speed, memory, routing quality, compression quality, search quality, verifier burden, and downstream task quality. A substrate can pass one axis and fail another without contradiction. A future route may consume only the axes that are both evidenced and allowed by the consumer gate.

Twelve owner interfaces keep the lease narrow. Intent names objectives and prohibited inferences. Planning and Cognitive Compilation retain plan and IR semantics. Context and Representation retain source fidelity, reconstruction, taint, and semantic loss. Routing consumes the lease and preserves fallback. Fast Generation retains accelerated-route latency and decoding claims. Deliberation retains extra-inference and stopping authority. Resource Economics retains complete cost, queues, capacity, and opportunity burden. Verification and Proof Envelopes retain theorem adequacy, runtime refinement, and evaluator independence. Security, Privacy, and Rights retain attacks, data, licenses, provenance, and access. Runtime and Artifact Graphs retain executable bindings, hashes, replay, fallback artifacts, and rollback closure. Readiness, Incident, and Release retain qualification, quarantine, recovery, and public authority. Claim Ledgers and Evidence retain atom status, negative results, reproduction, transfer, expiry, and residual lineage.

66.9 Invariants

• Every lease is specific to consumer, use, workload, claim axis, implementation, baseline, resources, environment, and time.
• Exploratory, structural-only, empirical-canary, qualified-for-scope, blocked, retired, refuted, and superseded are distinct operational states.
• A theorem authorizes only its exact statement under its assumptions and implementation relation; it supplies no empirical axis implicitly.
• Structural, representation, search, routing, reasoning, task, resource, verifier, safety, rights, deployment, and SOTA axes remain separate.
• Ordinary, strongest-current, simple-ablation, and embarrassing controls are frozen before outcome inspection.
• Information, data, compute, search budget, hardware, tuning, verifier access, retries, and evaluator opportunity are matched or residualized.
• Failed hypotheses, nulls, regressions, refusals, timeouts, OOMs, fallbacks, repairs, and exclusions remain in complete denominators.
• A favorable mean cannot hide tails, subgroups, pathological structures, verifier disagreement, or downstream regressions.
• Training, inference, compilation, search, memory, storage, verification, human, fallback, recovery, and opportunity burdens remain in one comparison.
• No consumer may rely on an unmeasured, blocked, expired, non-transferable, or rights-incompatible axis.
• Paper results, schema validity, fixtures, theorem counts, and green validators remain comparator or record evidence, not local benefit.
• Every canary retains a tested, provisioned ordinary fallback with bounded reversal time.
• Qualification narrows or expires on material substrate, implementation, theorem, workload, baseline, model, hardware, data, evaluator, policy, rights, threat, consumer, or time change.
• Independent reproduction cannot share the decisive implementation, evaluator, leakage, or benchmark defect.
• Transfer requires heterogeneous settings and preserves failures rather than averaging incompatible regimes.
• Support, deployment, safety, efficiency, generality, or SOTA movement requires an accepted transition for that exact claim and scope.
• Every live optional candidate has a dated run, narrowing, retirement, or blocked-after-full-attempt condition.
• Every failed gate, hidden cost, theorem-model gap, fallback defect, transfer failure, and residual has an owner and reopening condition.

The adoption invariant is intentionally conservative: the default state is optional. A substrate earns broader use through bounded evidence, not through elegance.

Negative-control memory forces substrate adoption to keep its counterevidence attached. The ordinary baseline, simple ablation, and embarrassing control remain attached to the substrate record after a favorable result. Promotion without the controls is not promotion; it is wishful selection.

66.10 Failure modes

• A notation, geometry, recurrence, or latent space is treated as evidence of usefulness.
• A structural theorem is laundered into speed, quality, safety, compression, search, reasoning, or deployment authority.
• A weak, outdated, undertuned, or resource-starved baseline makes the candidate appear useful.
• Outcome-driven search over substrates, seeds, tasks, metrics, budgets, or plots is reported as prospective.
• One favorable axis or average hides worse fidelity, tails, subgroups, verifier burden, safety, rights, or utility.
• Backbone throughput, linear scaling, cache savings, or active parameters are treated as single-request quality or economic evidence.
• Search breadth, extra inference, retrieval, or evaluator calls disappear from the resource bill.
• Failed, refused, timed-out, OOM, fallback, repaired, or unscorable cases disappear from denominators.
• A fixture, schema check, theorem count, or deterministic trace is called a model, kernel, search, or workload result.
• A consumer reads a structural-only or unmeasured axis because the substrate was admitted elsewhere.
• The fallback is stale, unavailable, slower than the recovery window, or silently depends on the candidate.
• The candidate becomes central before regression alarms, quarantine, rollback, and retirement work.
• Leakage, contamination, evaluator coupling, or shared implementation defects create false agreement.
• Privacy, license, provenance, accessibility, or dual-use constraints are omitted because the object is mathematical.
• A narrow result is generalized across workloads, scales, models, hardware, organizations, or time without transfer.
• Negative results are relabeled exploratory indefinitely, turning optionality into an unfalsifiable parking lot.
• Source-reported Mamba, Universal Transformer, Recurrent Transformer, V-JEPA, Circle, TreeLLM, coil, or calculus results are imported as ASI Stack results.
• Governance burden exceeds benefit, but sunk cost or narrative commitment prevents retirement.

Substrate failures produce downgrades, residuals, or blocked promotions. Negative results are useful; hidden negative results are architectural debt.

Theorem spillover is the search-substrate overclaim. A proof of an address, phase, recurrence, or structural property is used to imply speed, search quality, compression, or reasoning quality that the theorem never measured.

Substrate capture is the organizational version of the same mistake. A mathematically elegant candidate becomes so central to the story that ordinary baselines, simpler alternatives, and negative controls are treated as distractions. The adoption process should resist that by requiring workload fit, falsification conditions, comparable baselines, and explicit retirement triggers. Interesting structure earns a place in the stack only by surviving contact with the tasks it claims to help.

66.11 Minimum Viable Implementation

Optional substrates need a substrate adoption record before they can become part of the stack. The repository fixture records the candidate, intended use, advantage, baselines, baseline obligations, negative controls, proof boundary, experiment requirements, consumer gate and policy, axis ledger, falsification condition, adoption state, routing-permission effect, fallback substrate, retirement/supersession path, support-state effect, residuals, evidence references, and non-claims.

The substrate-adoption fixture does not show substrate usefulness. Its value is narrower: it prevents the book from describing adoption without the fields that adoption requires.

The Substrate adoption trace is now the next minimum executable form. It records valid_exploratory_registration, valid_structural_only_receipt, valid_consumer_axis_blocked, and valid_negative_control_retirement: exploratory registration, a structural-only proof receipt, a blocked consumer-axis request, and a negative-control retirement/refutation path. Its expected-invalid controls reject missing baselines, missing falsification conditions, theorem spillover into route permission, unmeasured-axis routing, failed-control promotion, missing fallback, support-promotion overclaim, and missing non-claim boundaries.

The trace makes optionality executable before any substrate becomes load-bearing. It still does not run a substrate A/B test, representation-efficiency test, search-quality test, routing-quality test, compression-quality test, Circle sidecar, CoilMoECOT run, Mamba comparison, TreeLLM implementation, or Theseus transfer consumer.

The exact minimum therefore comprises one schema-valid record, four valid synthetic states, eight expected-invalid controls, and nineteen live theorem declarations in AsiStackProofs.SearchSubstrates. Fourteen declarations belong to a reachable classifier and its route witnesses; five retain earlier finite record countermodels. It contains no implemented candidate substrate, learned model, kernel comparison, natural workload, measured benefit, independent reproduction, transfer result, or accepted chapter-core transition. The first real evaluation must include a candidate that remains interesting but unadopted when its negative control fails.

66.12 Mature Research Target

Mathematical and search substrates need an adoption exchange before they can affect ordinary routes. It keeps unusual mathematics available for narrow specialist routes while making every adoption decision baseline-bound, falsifiable, consumer-gated, and reversible.

Mathematical substrates remain optional specialists until baseline-symmetric workloads show where they improve search, routing, compression, or representation without hiding tradeoffs. Adoption records carry identity, intended use, expected advantage, ordinary baselines, baseline obligations, negative controls, proof boundary, workload, metric, script and report requirements, axis ledger, consumer policy, falsification condition, adoption state, fallback substrate, retirement path, residuals, and non-claims.

Structural facts, empirical workload evidence, consumer permissions, resource costs, and retirement or falsification behavior stay in separate lanes. A substrate can be structural-only, exploratory, canary, qualified-for-scope, blocked, retired, or superseded without contradiction. Routing treats substrates as specialists, compression and semantic layers test representation efficiency, proof envelopes carry structural facts, fast-generation routes consume only supported axes, and readiness gates decide whether the substrate remains optional, canary, or qualified for one scope.

Favorable A/B results open canary routes, failed negative controls narrow adoption, hidden controls become residuals, unmeasured axes remain blocked, and theorem-only records stay structural until workload evidence exists. Performance overclaiming, opaque mathematics as authority, adoption without regression tests, single-seed or confounded experiments, theorem spillover, and sequence-throughput claims used for unrelated quality axes become blocked promotion, residual records, fallback to ordinary substrates, or retirement decisions.

66.12.1 Argument-exit campaign

A competent full attempt begins with a small set of actually implementable candidates, not every idea in the source archive: one cyclic or geometric search operator, one graph or semantic representation, and one recurrent, state-space, or latent-prediction substrate. Each candidate receives a bounded natural workload that exercises its claimed mechanism and a consumer that can use the result without making it general infrastructure. Ordinary Transformer, state-space, recurrent, literal search, standard retrieval, graph, compiler, and no-special-substrate routes serve where applicable as ordinary, strongest-current, simple-ablation, and embarrassing controls.

The preregistration freezes data, splits, contamination probes, seeds, difficulty strata, sequence lengths, modalities, hardware, kernels, caches, search budget, tuning budget, retries, stopping, verifier access, evaluators, failure states, and promotion rule. The campaign measures representation fidelity and rate, verified search success and regret, routing calibration, reasoning and downstream utility, latency and tails, memory, energy, training and inference work, verifier and operator burden, safety, privacy, license and provenance compliance, fallback success, recovery time, and residuals in the same report. Failed, refused, timed-out, OOM, fallback, repaired, and unscorable cases remain in the denominator.

Causal ablations remove or randomize the proposed cyclic relation, graph topology, recurrence, selective state, latent predictor, adaptive halting, or search heuristic while matching parameter, token, compute, memory, tuning, and evaluator opportunity. The predicted mechanism-specific signature must vanish or change accordingly. An independently implemented candidate and an independently implemented evaluator then reproduce the result before transfer across at least two consumers and heterogeneous models, scales, hardware, data regimes, organizations, rights conditions, attacks, and time. The terminal outcome may be qualification, narrowing, null, refutation, retirement, or blocked-after-full-attempt; novelty is not owed a positive result.

This substrate-adoption gate remains an architectural proposal. The core stays at argument until a consumer-specific natural campaign clears the matched baseline, joint-burden, causal, fallback, independent-reproduction, and heterogeneous-transfer gates. A successful axis moves only that axis and scope. Useful, efficient, safe, deployable, general, transferable, or SOTA language requires separate accepted transitions.

66.13 Codex test plan

Test	Purpose	Status
Substrate adoption record fixture validation	Validate that a candidate substrate records intended use, expected advantage, baselines, baseline obligations, negative controls, proof boundary, experiment requirements, consumer gate/policy, axis ledger, falsification condition, adoption state, routing-permission effect, fallback substrate, retirement or supersession path, support-state effect, residuals, evidence references, and non-claims.	implemented; passing via python3 scripts/validate_protocol_examples.py
Required-field negative case	Prove that a substrate adoption record missing baseline refs, measured target, or falsification criterion rejects the finite adoption-field predicate.	implemented in AsiStackProofs.SearchSubstrates.substrate_adoption_record_missing_required_field_rejected; no substrate-quality claim
Unproven qualified-state negative case	Prove that a record simultaneously marked qualified and lacking passing evidence contradicts the finite rule that unproven records remain in a non-core state.	implemented in AsiStackProofs.SearchSubstrates.unproven_qualified_record_contradicts_noncore_invariant; this rejects an inconsistent authored record, not the substrate itself and not a core-substrate promotion
Qualified-without-evidence negative case	Prove that a qualified substrate with false passing-evidence rejects the finite core-adoption predicate.	implemented in AsiStackProofs.SearchSubstrates.qualified_substrate_without_passing_evidence_rejected; no empirical substrate result
Consumer-axis negative case	Prove that consumer reliance on an unmeasured or blocked substrate axis rejects the finite consumer-axis predicate.	implemented in AsiStackProofs.SearchSubstrates.consumer_axis_reliance_without_measurement_or_unblocked_axis_rejected; no routing, compression, or reasoning-quality claim
Canary evidence-packet negative case	Prove that canary substrate promotion missing workload, baseline, negative-control, or result-report fields rejects the finite canary-evidence predicate.	implemented in AsiStackProofs.SearchSubstrates.canary_substrate_without_complete_evidence_packet_rejected; no canary benchmark claim
Substrate adoption trace	Validate exploratory registration, structural-only receipt boundaries, consumer-axis blocking, negative-control retirement/refutation, fallback preservation, and no-promotion boundaries with expected-invalid controls for missing baseline, missing falsification, theorem spillover, unmeasured-axis routing, failed-control promotion, missing fallback, support overclaim, and missing non-claims.	executable trace and reachable Lean classifier implemented; independently derives 4 accepted states and 8 exact rejections with 19 total module declarations; no substrate A/B test, substrate adoption, model-quality result, runtime result, support-state transition, or chapter-core promotion
Baseline comparison test	Check that adoption records name ordinary baselines before quality claims.	planned; not run
Representation efficiency test	Check whether a substrate improves a bounded representation task against a baseline.	planned; not run
Falsification review	Check that the record names what would count against adoption.	planned; not run
Sequence-substrate A/B comparison test	Check whether a state-space or recurrent sequence substrate improves a declared workload against a transformer baseline with verifier adequacy and fallback recorded.	planned; not run

The implemented rows validate fixture/schema consistency and finite-record negative cases only. The remaining rows require concrete workloads, baselines, negative controls, and run artifacts; they are not reported substrate results. When a substrate test is implemented, this chapter should link to the command, fixture, environment notes, workload, and result summary, and Appendix E should be regenerated or updated accordingly.

66.13.1 Formalization hooks

Tag	Module	Target	Status
lean:substrates.search.operational_invariant	AsiStackProofs.SearchSubstrates	A substrate adoption record missing a baseline reference, measured target, or falsification criterion fails the finite adoption-fields predicate.	implemented
lean:substrates.search.failure_blocks_promotion	AsiStackProofs.SearchSubstrates	A substrate record marked qualified without passing evidence fails the finite core-adoption predicate.	implemented
lean:substrates.search.adoption_trace_bridge	AsiStackProofs.SearchSubstrates	A reachable formal substrate-adoption classifier derives four accepted non-promoting states and eight exact rejecting controls from concrete trace inputs; consumer permission is reserved to measured-positive routes, while failed controls, theorem spillover, missing fallback, support promotion, and incomplete boundaries reject.	implemented

The first two Lean hooks are finite-record rejection results over substrate-adoption and promotion-review fields. The first rejects an adoption record when at least one required field is explicitly false. The second rejects a qualified promotion record when passing evidence is explicitly false. The module also proves bounded negative cases named unproven_qualified_record_contradicts_noncore_invariant, consumer_axis_reliance_without_measurement_or_unblocked_axis_rejected, and canary_substrate_without_complete_evidence_packet_rejected. The third hook replaces the retired authored trace summary with classifyAdoptionTrace, which derives all four accepted non-promoting routes and all eight expected rejecting controls from concrete fields. Route algebra limits consumer permission to the two measured constructors and excludes every rejection route. These cases reject malformed finite records; they do not prove that any substrate improves search, routing, compression, representation, long-context behavior, runtime, verifier burden, or model quality.

The module now contains nineteen theorem declarations. Five are the retained finite countermodels over missing fields, unproven qualification, qualified-without-evidence, unsupported consumer axes, and incomplete canary packets. Fourteen belong to the reachable classifier: two constrain route permission, four derive accepted non-promoting states, and eight derive exact rejections. The hand-authored trace-summary witness remains retired after the seventh C6 rationalization tranche removed five pass-through or summary projections. The independently encoded validator reproduces the twelve closed decisions, while Lean additionally requires measured permission to match the canary or qualified adoption state. This is not a proof-quality or support promotion. None proves a semantic correspondence between a mathematical substrate and an executable implementation, search completeness or optimality, representation fidelity, model quality, speed, memory efficiency, safe fallback, liveness, recovery, independent reproduction, transfer, or SOTA. Any declaration without a named claim consumer should be retired or replaced by transition semantics, countermodels, liveness, or implementation refinement rather than retained for the count.

66.14 JEPA and latent world models use the same adoption gate

Joint-embedding prediction is not a free-standing layer in this spine. It is a substrate candidate whose consumers are planning, data engines, and the reference architecture. ext_v_jepa_2_2025 demonstrates the relevant shape: predict future latent representations, add an action-conditioned predictor, and use representation-space costs for model-predictive control. That changes the adoption packet by adding predictive-state identity, observation and action provenance, forecast horizon, calibration, intervention tests, prediction-error ledgers, and sim-to-real residuals.

The structural claim—prediction occurs in representation space—must remain separate from benchmark quality, causal understanding, controller quality, and safe transfer. A JEPA or energy-based candidate reaches qualified_for_scope only after matched pixel/token/latent baselines, measured search and latency cost, counterfactual or intervention probes, downstream task evidence, and a fallback. A paper architecture or identifiability result cannot by itself authorize adoption.

66.15 Source crosswalk

Source ID	Title	Layer	Planned use	Readiness
genesiscode	GenesisCode	executable_specification	Tiny pure calculus + obligations + provenance for auditable AI-symbiotic programming.	source note available; local raw cache available
temporal_coil_research	Temporal Coil Research	mathematical_search_substrate	Manifest-driven A/B process for temporal-coil integration in language model training.	source note available; local raw cache available
cognitive_compilation	Cognitive Compilation	planning_semantic_ir	Compiler framing for LLM-centered planning, semantic IR, target compilation, incremental repair.	source note available; local raw cache available
treellm	TreeLLM correction lineage	semantic_representation	Matched candidate family for hard graph, HLSH/path tokens, ANN fallback, learned and recurrent navigation, caching, and speculative traversal, with mechanism-specific ablations and retirement tests.	source note available; local raw cache available
simulation_scaling	Simulation Scaling Law	compute_fidelity_constraints	Resource constraints on scope, clockspeed, and fidelity in simulations.	source note available; local raw cache available
coilmoecot	CoilMoECOT Whitepaper v2.0	mathematical_search_substrate	Prime-temporal coil networks as governed specialist cores in MoECOT runtime.	source note available; local raw cache available
circle_calculus_core	Circle Calculus	proof_carrying_mathematical_substrate	Proof-carrying finite cyclic mathematics project with Lean proofs, Python reference models, Rust utilities, theorem manifests, papers, and Quarto living book.	source note available
circle_ai_architectures	Circle AI Architectures	cyclic_ai_architecture	Disciplined Circle AI thesis: use phase, recurrence, rotation, sparse cyclic mixing, circular memory, harmonic transforms, or geometry-aware structure only where the structure is real and baselines support it.	source note available
proof_carrying_circular_computation	Proof-Carrying Circular Computation	proof_carrying_compute_substrate	CoilIR-style path from circle/coil expressions to dictionary-recognized cyclic structure, Lean-proved rewrite/address transformations, backend selection, and benchmark validation.	source note available
theseus_circle_transfer	Theseus Circle Calculus Transfer Lane	proof_contract_transfer	Report-only bridge from Circle finite fixtures into private Theseus benchmark design with explicit quality/runtime/memory/transfer/failure-case claim boundaries.	source note available
ext_mamba_2023	Mamba: Linear-Time Sequence Modeling with Selective State Spaces	sequence_substrates	Selective state-space models as an optional sequence-substrate comparison point, not a promoted ASI substrate.	source note available
ext_universal_transformer_2019, ext_recurrent_transformer_2026	Universal Transformers and The Recurrent Transformer	recurrent_transformer_substrates	External comparators for shared-weight depth recurrence, adaptive halting, recurrent KV state, and effective-depth tradeoffs.	source notes available; theoretical, scale, kernel, and local-reproduction limits retained

The crosswalk keeps optional substrates optional. genesiscode supplies small-kernel and obligation discipline, temporal_coil_research supplies manifest-driven A/B discipline, coilmoecot stays a readable design source rather than benchmark evidence, the Circle sources supply proof-boundary language, and ext_mamba_2023 keeps sequence-backbone efficiency separate from substrate adoption. None of the assigned sources promotes a substrate without baselines, controls, and adoption evidence.

66.15.1 Manifest source assignment reconciliation

These rows keep Mathematical and Search Substrates’s manifest assignments visible at their recorded review boundary. Passage review does not establish local reproduction, performance, safety, deployment, or support-state movement.

Source	Intake role	Boundary
ext_v_jepa_2_2025	Passage-reviewed comparator: V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning. Adds a concrete latent-prediction architecture family to the substrate adoption gate: action-free representation learning, an action-conditioned predictor, and explicit separation between structural design and capability evidence.	The reported video and robot results do not establish local model quality, causal understanding, safe control, transfer, deployment, or an ASI Stack result. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row.

66.16 Summary

Mathematical substrates are valuable when they become disciplined specialists. They should carry baselines, proof boundaries, falsification conditions, evidence requirements, and readiness states.

The stack can explore coils, calculi, semantic graphs, cyclic compute, and sequence-substrate alternatives without making them core prematurely. Optionality is not weakness; it is how the architecture keeps evidence in charge. The adoption policy then narrows into proof-carrying contracts: how a finite structural fact can travel without becoming a quality claim.

A substrate should become easier to reject as it becomes more interesting. If the record cannot say what would falsify adoption, the mechanism is not ready to influence architecture. This stance protects both ambition and rigor. The stack can keep unusual mathematics, search spaces, cyclic structures, and semantic representations in view without giving them authority they have not earned. A candidate substrate should name the workloads it might help, the baselines it must beat, the proof facts it can honestly carry, the empirical facts it still lacks, and the fallback path if it fails. That is how exploratory mathematics becomes an engineering option instead of a belief system.

66.17 Evidence reconciliation (2026-07-16)

The invariant protocol, field meanings, and inference limits are stated once in Living Book Methodology. This packet contains only the chapter-specific projection; its authoritative per-atom rows are the mathematical-and-search-substrates slice of experiments/claim_family_terminal_coverage/results/result.json.

The core remains blocked after full attempt at argument support. The strongest family attempt was KERC canonical-language and hierarchical-residual campaign. Its exact boundary is: The historical broad-efficiency transition is N1: the frozen implementation was inadequate, so broader KERC remains untested; two narrow finite observations survive, with no semantic, multilingual, production, energy, or core claim. Across 74 atoms, the terminal ledger records 74 blocked_after_full_attempt.

Chapter-specific field	Value
Family / atom denominator	CF-06 / 74 atoms
Terminal dispositions	74 blocked_after_full_attempt
Core	mathematical-and-search-substrates.core: blocked_after_full_attempt at argument
Core attempted / missing lanes	causal, empirical, executable, formal, source-synthesis / normative, transfer
Attempted local lanes	causal, empirical, executable, formal, source-synthesis
Missing or unproved lanes	normative, transfer
Strongest family bundle	KERC canonical-language and hierarchical-residual campaign (natural_work_and_end_to_end): A 192-record bilingual templated compiler/runtime study with 64 held-out records, five seeds, eight baseline families, 13 ablations, and 20 attacks.
Negative controls	surface and kernel-native baselines; 13 ablations; 20 attacks; ten laundering mutations.
Accepted transitions	none
Maximum inference	The historical broad-efficiency transition is N1: the frozen implementation was inadequate, so broader KERC remains untested; two narrow finite observations survive, with no semantic, multilingual, production, energy, or core claim.
Reproduction / next burden	Replay scripts/validate_p4_m8_kerc_campaign.py and scripts/validate_claim_family_terminal_program.py; fill the named atom-specific lanes under a new prospective protocol.

66.18 Handoff

Exploratory substrates become usable only when their structural claims can travel without inflating into performance claims. Circle Calculus and Proof-Carrying AI Contracts gives that travel path. It packages finite facts as theorem references, receipts, fingerprints, deterministic fields, validator state, consumer gates, and non-claims so downstream systems can inspect structure without treating it as model quality or deployment evidence.