Formal Vocabulary, Notation, and Claim Classes

  • Version: 0.1.35
  • Status: Normative for all new specification, plan, proof-model, and certificate text.
  • Source: docs/hologram_formal_analysis_direction.pdf §§1, 7, 13; tracker #122; issue #123.
  • Related: terminology for existing graph-compiler concepts lives in docs/transformerless/GLOSSARY.md; this document governs claim language and mathematical notation. Where the two overlap, this document wins for claim classification and the glossary wins for structural terms. The route-scope decision is ADR-0004, and current execution follows the native geometric AI policy. Geometric Intelligence Evaluation retains historical experiment protocols. A preserved #973 terminal contract is R4PredictiveBlockDeltaPromptCapacityV5.

Current authority (2026-09-05). This document governs claim language and typed mathematical distinctions. The project plan owns the goal, and current-state.md owns the implemented model and next work. Dated results and table cells naming a "current" reference, a parked operation or an unimplemented boundary retain their original artifact scope; they do not select a new task or prohibit a versioned native successor. Preserve their measurements without importing a reference model's capability or runtime guarantees into the active model.

This spec separates the statement classes that the research notes previously mixed — architectural definitions, compiler optimization objectives, empirical certification claims, and structural runtime guarantees — and fixes the symbols that recur across the compiler, artifact, certificate, and proof model.

1. Claim classes (role of a statement)

Every equation and every normative mathematical statement in repository documents MUST carry exactly one of these role labels, written inline in bold (**Definition**, **Objective**, **Guarantee**, **Assumption**, **Empirical Criterion**):

LabelMeaningWhere it may appear
DefinitionIntroduces an architectural object or mathematical term. True by convention; never "proven".Anywhere
ObjectiveA quantity the offline compiler attempts to optimize. Never a runtime invariant, never a theorem.Compiler-side docs/code only
GuaranteeA structural property of the compiled artifact or runtime. Requires a claim status from §2 and a proof artifact, witness, test, or explicit Unproven record.Artifact spec, runtime docs, proof model
AssumptionA condition a proof or certificate requires but the implementation does not itself establish.Proofs, certificates
Empirical CriterionA measured property with a declared distribution, protocol, sample count, and uncertainty. Never stated as a proof.Certificates, evaluation docs

Design aspirations are not theorems: an optimization target (arg max …) is an Objective, and a measured approximation (P_θ(a|c) ≈ P_θ(a|Z(c))) is an Empirical Criterion until the protocol behind it is declared.

2. Claim status (evidence behind a statement)

Every Guarantee and every Empirical Criterion MUST also carry one of these statuses. The statuses are the machine-readable vocabulary of the proof-status matrix (crates/uor-r4-proof-model/src/proof_matrix.rs); the mapping is normative.

StatusMeaningProofStatus mapping
StructuralEstablished by construction or by a named machine-checked test, fuzz target, or formal proof artifact that has actually run for the declared evidence.Verified
WitnessedEstablished per execution by a bounded, replayable witness an independent verifier replays without the teacher.Verified (witness path)
EmpiricalMeasured on a pinned corpus with protocol, confidence intervals, and provenance identifiers, per docs/transformerless/EQUIVALENCE_AND_EMPIRICAL_PROTOCOL.md.DifferentialPass
AssumedRequired by a proof or certificate but not established by the implementation.recorded as an assumption entry
UnprovenAsserted as a goal but currently without evidence.Unverified (blocks verify_all)

ExecutableSpec in the proof matrix means the obligation has a runnable specification but no activated binding check yet; documents must label such claims Unproven (or Assumed where applicable), never Structural.

2.1 Wording rule (automated check, dormant by default)

The prohibited phrases are "machine-verified", "machine verified", "exact equivalence", "exact teacher equivalence", and "provably equivalent". They are prohibited in normative documents (docs/**, crate READMEs) unless the same line links a proof artifact or certificate, or the phrase is explicitly disavowed on that line. The uor-r4-claim-wording gate enforces this when explicitly activated (see §6). The project does not claim exact teacher equivalence (disallowed per this rule), does not claim human-level reasoning, and does not treat plausible language output as evidence of coherent internal state transitions.

3. Core notation

Symbols are listed with their claim role and their concrete Rust binding. Entries marked compiler/reference-only have no deployed-runtime representation yet; the runtime path for them lands in the issues noted.

SymbolRoleMeaningRust / artifact binding
x_t^ℓDefinitionHidden/residual state at token position t and decoder layer ℓ.Active geometric-decoder/source-control state; exact binding lands under #950/#952.
q_t^ℓ, k_i^ℓ, v_i^ℓDefinitionLearned query, causal prefix key, and value contribution. These predictive roles are distinct from immutable R4G1 route codes, addresses, and CIDs.The #950/#951 G0/G1 mixer and DirectCausalGeometricAttentionR4V1 are retained comparators. The current offline reference is HELM-D-R4 under ADR-0005; its first parity arm leaves every learned Q/K/V and W_o unchanged.
p_t, e_t, N_tDefinitionRegistered prime route atom; canonical unordered semiprime transition expert e_t = p_(t-1)*p_t; and bounded ordered n-let. If adjacent atoms repeat, e_t = p_t^2 is retained. The square-free case is the subset p_(t-1) != p_t. N_t retains order and a multiplicity-preserving sorted factor multiset rather than relying on an overflowing numeric product.Schema-2 source-free manifest records under #958. This binding establishes representation and deterministic recall substrate only.
Gamma = (gamma_0,...,gamma_(m-1)), theta_j(p), delta_theta_jDefinitionGamma is a finite ordered, revisioned zeta-ordinate grid. theta_j(p) = wrap(gamma_j*log(p)); delta_theta_j is the declared local phase difference.Compiler/reference math and bounded compiled signatures under #958. Critical-line use is an Assumption, not a proof of RH; no serving or capability claim follows.
s_t in S3 subset R4, h(s_t) in S2 subset R3, tau_t in S1Definitions_t is the full normalized local spin state, h is its many-to-one Hopf observation, and tau_t is retained quantized fiber/transport phase (“torsion”). h(s_t) alone is not a rebuild identity.Quantized source-free route records and experimental bounded lookup under #958; no physical or quantum interpretation is claimed.
theta=atan2(sin(theta),cos(theta)), a(theta)=sin(theta)^2, chi(theta)=sign(sin(theta)), pol(theta)=sign(cos(theta))DefinitionS3/R3 trigonometric transition contract. (sin=+/-1,cos=0) has activation 1; (sin=0,cos=+1) is continuous-null activation 0. Chirality and, where needed, cosine polarity preserve orientation/antipodes. tan(theta) is a local chart only.At a tangent pole or declared null boundary, the architecture switches angle/cotangent chart and records a signed quarter-turn phase plus torsion shift instead of dividing by zero or terminating. Semantic value is Unproven.
m_E=sqrt(2), m_C=2i, m_R in [0,2]DefinitionTyped least-cost adapter markers: Euclidean orthogonal-unit chord, complex/discrete antipodal displacement, and declared normalized Riemannian/chord score interval. They are not literal equalities of Euclidean, complex, discrete, and Riemannian domains.A target-specific cost profile binds chart, units, orientation, quantization, error bounds, canonical tie-break, and conversion witness before choosing the cheapest faithful operation.
r_t = a_t + b_t*phi in Z[phi]DefinitionDirection-preserving golden radial shell. phi:(a,b)->(b,a+b) and phi^-1:(a,b)->(b-a,a). Repeated steps have Fibonacci recurrence, which does not establish semantic value.Exact coefficient updates and manifest chart profile under #958; product use remains separately qualified.
I1, I2, ISDefinitionBounded causal indexes keyed by last route, ordered previous-plus-last route, and ordered sentence-route identity.Incremental identity, exact rows, and an off-serving geometric-attention experiment exist under #958. Exact hits are recall; attention and product claims require the evaluation policy.
e^(i*pi)+pi^0 =_bridge 0^0, b_0DefinitionTyped zero/identity transition. ContinuousNull preserves the complex cancellation as 0; DiscreteEmptyProduct phase-shifts/retypes the seam into the discrete identity 1. =_bridge is a domain-transition operator, not ordinary numerical equality. The selected tag is part of route identity and deliberately exposes the -1 / 0 / +1 landmarks without assigning two simultaneous values in one untyped calculation.ZeroPowerBridge and schema-2 manifest binding under #958 select the boundary value; the full phase-shift transition remains an architectural contract for the route hierarchy.
C_lexDefinitionDeterministic lexical codec from pinned input bytes/tokenization to registered route atoms and back to output bytes where decoding is defined. Normalization, unknown-unit behavior, tokenizer CID, registry, and codec version are part of its identity.Product binding remains Unproven until a provider-free serving path exercises the same codec end to end.
B_ico : Lambda_E8 ~=_Z I -> (x,x') in R4 ⊕ R4, Phi_E8 = H4 ⊕ φH4AssumptionProject shorthand: E8 = H4 × H4. The load-bearing implementation contract realizes that conceptual identity through the chosen icosian presentation: the E8 lattice and icosian ring are identified as Z-modules, one state is represented by golden/Galois-coupled R4 points, and the declared 600-cell folding is H4 ⊕ φH4. This is a typed construction, not a claim that R4 = E8.Basis, glue/parity rule, conjugation, scale, orientation, root order, and inverse witness must be kappa-bound. Architectural load-bearing is assumed and is scheduled for test under D20 §4, which bounds how long the assumption may stand unchallenged; held-out advantage remains unproven. Structural priority and measured advantage stay distinct claims.
kappa(X)DefinitionCanonical content identity of the schema/provenance-bound byte envelope X. Equality identifies equal canonical envelopes under the named digest contract; digest distance has no routing meaning.Manifest/artifact CIDs. Kappa is integrity and provenance, not a semantic code.
R_t^q, q in {local,sentence,paragraph,conversation,global}DefinitionIdentity-scoped, causal route accumulator at hierarchy scope q. A parent scope commits to ordered child identities and bounded geometric summaries; it does not rescan all descendant tokens at query time.ADR-0004 contract; scopes beyond the current local/sentence substrate are Unproven.
T_t^q = (v_session, w_window, E_proj, F_shared, c_res, alpha_Hopf, B_ico)DefinitionBounded transported trajectory/harmonic summary for scope q: session hypersphere vector, winding/window state, projection energy, shared-prime factors, cosine resonance, accumulated Hopf phase, and paired-H4/E8 coordinate. It is updated incrementally from the full transported path, not reconstructed from only the last route.Ancestor routing evidence motivates these fields; current held-out locality when exact hierarchy keys miss is an Empirical Criterion, status Unproven until reproduced.
W_covDefinitionBounded coverage witness recording codec/address coverage, hierarchy rows read and hit/miss, candidates before/after admission, controls, selection or abstention, and artifact identities.Coverage establishes reachability only; capability status comes from a separate empirical record.
A_recall, A_geoDefinitionA_recall retrieves a stored continuation by exact/backoff route identity. A_geo selects or orders admitted causal support using declared geometric terms and must be load-bearing against matched controls on anti-recall inputs.Exact-row recall is implemented substrate. Geometric-attention advantage is an Empirical Criterion, currently not implied by implementation presence.
F_i^b, P_(j -> i)^bDefinitionF_i^b is the exact cumulative Spin/H4 orthogonal model-frame basis for R4 head block b; local encoding is transpose(F_i^b) x, and P_(j -> i)^b = transpose(F_i^b) F_j^b moves both locally encoded causal K_j^b and V_j^b into query frame i. The current vector action is a compiler-side f64/f32 oracle, not exact deployed arithmetic.HELM-D-R4 uses this declared orthogonal gauge transport before comparison and aggregation. It is not called Levi-Civita or shortest-geodesic transport without a separate proof.
B_c(g), theta_x^rDefinitionConnectionGaugeCovarianceV4 represents each predictive role by the same three local coefficients theta in a declared tangent frame B_c(g) and transports them by B_c(d) transpose(B_c(s)).V4 passed construction covariance, all 120 frames/14,400 ordered pairs, and finite-difference gradients, but its protected held-out reveal scored 13/24 for every main arm with insufficient destructive-control separation. It is frozen negative evidence.
alpha_(i,j), r_iDefinitionIn the HELM-D-R4 parity arm, alpha_(i,j) is the unchanged stable causal-softmax normalization of the donor's ordinary compatibility score after K transport, and r_i = F_i sum_(j<=i) alpha_(i,j) P_(j -> i) transpose(F_j) v_j is mapped back before unchanged W_o.This is the accepted gauge-equivalent reference-attention baseline. Its bounded held-out full-decoder parity result and native R4SoftmaxReferenceGeneratorV1 gate are Pass. The dedicated opt-in, loopback-only HTTP endpoint additionally passes exact eight-token CLI token/decision-CID/state-CID parity with all 30 layers audited and zero future reads. Dashboard wiring/static native-readiness and WASM-isolation checks pass; browser interaction/E2E is NOT_RUN. These are numerical/behavioral source-donor/reference results, not geometric advantage or a deployed transformerless-runtime mechanism. The generator uses UOR's pinned SmolLM2 HuggingFaceLlamaOracle decoder path and remains Transformer-compatible and source-weight backed. HELM-D is an MIT architectural reference only at 7501deca8f413848bfef804be64ce874b72a3cd7; no HELM checkpoint or code executed and no upstream result is inherited. See helm_d_r4_softmax_decoder_973.md, r4_softmax_reference_generation_973.md, r4_softmax_reference_generation_attempt_01_result_973.json, and r4_softmax_reference_http_bridge_973.md.
R4SoftmaxTeacherTraceV1, R4SoftmaxTraceCompilerV1DefinitionConstruction-only trace schema and compiler. The trace binds causal layerwise token, Q/K/V, attention-support, value-aggregate, decoded-output, and logit states from the exact qualified reference. The first compiler emits matched teacher-distilled, observed-count, and document-permuted Q16 suffix arms without granting future or target inputs at inference.Pass at the bounded source-free-distillation scope: context-bearing covered teacher CE was 2.660721 versus count 9.678894 and permuted 4.342019; teacher top-1 was 3/9 versus 2/9 and 1/9; actual-next top-1 was 2/9, tied with count. Artifact/reveal CIDs and replay were exact, causal audits passed, and student inference made zero source calls. Decoded continuation collapsed into a , Scotland cycle, and the student does not use geometric trace state, so this is not geometry advantage, coherent generation, reasoning, or a deployed transformerless runtime. See r4_softmax_trace_student_973.md.
R4SoftmaxTraceStateStudentV1DefinitionBounded source-free state transition and next-token readout over the captured query-gauge Q, transported K/V support, weighted aggregate, and decoded model-frame trace, compared with the established suffix student, an equal-budget non-geometric recurrent state, and a transport-permuted control.Completed negative at the frozen #1011 gate. On the same nine context positions and 422,875 Q16 teacher mass, covered CE / teacher top-1 / actual-next top-1 were suffix 2.660721032, 3/9, 2/9; plain 2.660770919, 3/9, 2/9; geometric 2.660705367, 3/9, 2/9; and transport-permuted 2.660729215, 3/9, 2/9. The geometric gains of 0.000015665 nats over suffix and 0.000023848 over permutation were below the frozen 0.10 threshold; no top-1 decision changed, the control lost none, and all arms retained the period-two , Scotland cycle. Exact replay and causal/runtime audits passed. State/freeze/seal/result CIDs are respectively blake3:b617fc38e7bef1cdea76991f6e5e7cc653118451d63bcbd595f8ffd7e247ae7b, blake3:67cf67bb46b94cf5644b8dde286e89adb7e49159b3749790dffb500d8047fedb, blake3:64587526f7883ab046e884a28b6af7e9e89818c9ead2039f8c995de7fb483060, and blake3:dc04a8a8b21750799db2d451c8237d1e62cf90ffa74561fb54272b1e9704c824. Terminal: STOP_R4_SOFTMAX_TRACE_STATE_STUDENT_REPAIR_OR_RETIRE_REPRESENTATION. This falsifies the current 4D signed-reduction/token-derived state cell, not ordinary R4/Spin softmax attention. See r4_softmax_trace_state_student_1011.md.
R4GroupAddressedRetentionLMV1, M_t^b(h)Definition#973's fixed-size source-free state: four banks hold an R4-block field over 120 addresses, updated through one frozen group action and read candidate-relatively with separate learned query/value tables. Exact-H4 is compared with equal-size cyclic-120 and destructive scrambled-H4 actions; a state-off intervention tests whether retention changes logits.Construction terminal: UNAVAILABLE_FRAME_POPULATION_OR_LOCAL_BUDGET. Geometry, population, reachability, gradients, memory, equal work, and held-out sealing passed; timing and disposable learning smoke failed. Main optimization and held-out model criteria are NOT_RUN; there is no attention or H4-advantage verdict. The exact cell is not retried or tuned. See r4_group_addressed_retention_973.md.
R4GroupAddressedRetentionDecoderV1CpuRecoveryEmpirical CriterionIndependently frozen complete construction run of the group-addressed retained read/write law inside an exact 3.17M-parameter, two-block decoder recipe. A state-off intervention measures whether retained state is load-bearing; exact-H4 versus scrambled transport measures geometry-specific separation.Completed: state-off on the disjoint construction-validation partition lost 0.967227 nats and 182 top-1 hits, qualifying a bounded causal retained-attention component. Aggregate validation CE moved 8.371911 -> 8.976155, so the exact data/dose/parameter recipe did not satisfy its frozen full-decoder generalization criterion. Scrambled transport was 0.033049 nats better, so no H4-specific advantage follows. Formal H4 specificity is NOT_EVALUATED. Result CID blake3:68355ad2f61d02dc73dbf22de4c24834815a23069ed5735630dc365081cf91db; see r4_group_addressed_retention_decoder_cpu_recovery_973.md.
R4RetainedLanguagePathV1DefinitionA 252,160-parameter source-free language path using the frozen group-addressed retained-state law, compared under equal parameter budget with an ordinary causal-softmax decoder and under a state-off intervention.Qualified at its declared scope: retained NLL 3.899862, ordinary-control NLL 3.903394, and state-off NLL 4.234849; state removal lost 16,660 correct next-token decisions. A separate retained-only smoke produced five valid, exactly replayed local continuations without forbidden/provider/source reads, but every output drifted from its prompt subject or scene. This establishes load-bearing retained attention and autonomous local decoding, not prompt-conditioned coherence, H4 superiority, reasoning, or exact lowering. See r4_retained_language_path_v1_973.md.
R4PairedH4LanguagePathV1DefinitionThe one frozen V1-capacity successor that holds training data, seed, schedule, parameter count, and work fixed while replacing the shared per-layer H4 token address with a deterministic paired-H4 address.The address construction reduced repeated joint addresses by 97.5477%. On fresh held-out language it scored NLL 3.8832293739 and top-1 29.780171%, versus V1 3.8901151940 and 29.706357%. This is construction/general-language evidence only; it does not establish prompt capacity or H4 superiority.
g_d, G_promptEmpirical CriterionFor each of 512 bidirectional matched-prompt contrasts sharing the final four prompt tokens, `g_d = (log P(y_down prompt) - log P(y_d
R4LayerwiseNormalizedRetainedReadoutLanguagePathV1Definition#973's final zero-parameter readout variant: E @ [N(h) + (g/sqrt(2))*(N(a1)+N(a2))], fixed candidate g=1 versus matched V1 g=0, with recurrence, state, parameters, data/order, optimizer, 2,730-step dose, and tied vocabulary projection held fixed.Completed partial: V3 candidate G_prompt = 0.0286980210 versus V1 0.0073316237, incremental 0.0213663973, with 339/512 wins. Every fresh-language and mechanics gate passed, but the candidate missed the absolute 0.0433216988 and incremental 0.0253415693 prompt-gain floors. Terminal LAYERWISE_NORMALIZED_RETAINED_READOUT_PROMPT_CAPACITY_PARTIAL; result CID blake3:35396bd6e64fc2c0bc7d86a84cc9e212ed913ce28e5353f5f2b8212b4cf2c532; independent verification CID blake3:3f316541dbab8061ed5ba891bf6a47ef22c55bca21fba01f6f97dbb3cb8497aa. Generation, reasoning, and exact/geometry-native lowering are NOT_RUN.
R4LearnedCandidateLeafAssociativeReadoutV1, zG[t,c]Definition#973's completed learned readout reads V1's strict-prior transported retained value at candidate c's canonical exact-H4 leaf with an independently learned [2,4096,12,4] candidate-query table: zG[t,c] = <E[c],N(h[t])> + (1/(2*sqrt(48)))*sum_l <qG[l,c],R(V[l,t,lambda(c)])>. Qualified V1 is immutable. The equal-parameter address-blind arm has an unshared, byte-identically initialized query table and reads the occupied-address mean; a fixed-leaf derangement reuses the geometric table while destroying candidate/address binding.Completed negative at the frozen prompt-capacity scope: geometric gain 0.00637679, 299/512 wins, and own NLL 3.71038302 failed the capacity floors. Pooled gain 0.01026323, 324/512 wins, and own NLL 3.68289051 was partial but also below both gain floors. Both fresh-language and state-load-bearing gates passed. Terminal LEARNED_ASSOCIATIVE_READOUT_NO_CAPACITY; result CID blake3:cedba37738ee249457bb589f716ee75afb16a0c4937c2a22ae9f917dd3eb97c1; verification CID blake3:443d711ce9a228e26e2eb2eebb55c582848424e2677c3473d41deaf8afd69ec7. See r4_learned_associative_readout_prompt_capacity_973.md.
R4LearnedAssociativeReadoutPromptCapacityV1Empirical CriterionEach learned arm must pass the frozen V4 absolute and incremental prompt-gain floors, 308/512 wins, own-NLL nonregression, fresh-language nonregression, load-bearing state, mechanics, replay, work, and causal audits. GEOMETRY_ATTRIBUTED separately requires the geometric arm to exceed both pooled and fixed-leaf-deranged controls by the frozen incremental effect and win/NLL rules.Empirical terminal: neither arm passed capacity. Geometry attribution failed: geometric-minus-pooled gain -0.00388645 with 209/512 paired improvements and geometric-minus-deranged gain -0.00028887 with 251/512, below +0.02534157 and 308/512. All ten mechanics gates and independent replay passed; post-reveal optimizer steps were zero. No retry or generation is authorized. Its then-next separately frozen architecture had to alter the retained value write/binding law and retain pooled plus geometry-destroying controls; R4PredictiveBlockDeltaPromptCapacityV5 below records that successor. #973 remains open and #954 blocked.
R4PredictiveBlockDeltaBindingV1, S_t^(b,l), r_tDefinition#973's completed predictive write/binding law stores four bounded banks of 12 independent 4 x 4 R4 block memories. In the current frame, each bank transports prior state, applies either the full delta update S_t = rho*P*S_(t-1)*P^-1 + eta*(v_t - rho*P*S_(t-1)*P^-1*k_bar) outer k_bar or the declared additive control, and reads r_t = sum_b softmax(alpha)_b*S_t*q_t. The observed value and candidate scorer are anchored to immutable token H4 leaves.Compiler-side f32 experiment only. The full-delta geometric, identity/plain, and additive arms each used 9,228 trainable values and identical 2,730-step construction dose. This definition grants no runtime legality, geometry advantage, attention-capacity, generation, or lowering claim. See r4_predictive_block_delta_binding_prompt_capacity_973.md.
R4PredictiveBlockDeltaPromptCapacityV5Empirical CriterionOn the sealed 512-direction V5 population, terminal capacity requires geometric gain at least 0.04332169878499658, at least 308/512 wins, the frozen incremental and own-NLL rules against immutable V1 and pooled, fresh-language nonregression, load-bearing state, and complete integrity. Geometry and delta overwrite are separate attributions against independently fitted plain plus transport-permuted, and independently fitted additive, using the frozen 0.025341569256760274, paired-win, and own-NLL rules.Empirical terminal: PREDICTIVE_BINDING_NO_TERMINAL_CAPACITY. Geometric gain 0.03896945868086732 with 375/512 wins missed the absolute floor. Geometric-minus-plain gain 0.023929811749894725 and worse own NLL failed geometry attribution even though the transport-permuted gates passed. Geometric-minus-additive was -0.006512463228773413 with 234/512, so delta superiority was not established. Fresh-language and integrity passed. Result CID blake3:6c67544d675eafcb8eb9c0dabb93617e3f6c3295af812e8acbb687107c010a74; scoring CID blake3:44f8941d24a99fc230710fd700e7a7b13cee87587bfbe4e13bf7b095222e2ee6; recovery CID blake3:7b76e36e44798bebf184ece08fdd8a2065bdd370106b5d64d5fae4c59dc6d88b; exact-replay verification CID blake3:567cf336eb05c3ec562aef7135f6fb35b580d02c758b0e79f2508cae57065f5d. Action STOP_WITHOUT_GENERATION; only this law is retired.
O_trace^(0..3)ObjectiveConstruction-only, leave-one-document-out observability ladder measuring the same teacher-relative candidate loss at four frozen boundaries: full ordered final-layer Q/K/V trace blocks; fixed 576-to-4 signed reduction; token-derived role maps plus recurrent state features; and fitted residual readout/logit scale.Completed in #1012 at INSUFFICIENT_SUPPORT_COVERAGE: aggregate primary coverage was 0.6202622204224402, but the minimum fold was 0.3469116829611222, below the frozen 50% floor, so no boundary attribution is licensed. On the covered rows, full Q/K/V CE was 2.215410922655504 versus suffix 2.215064603216862 (suffix - full = -0.0003463194386417179, direction 0/4); the fixed label control separated by 1.3807454322642605 nats in 4/4. Exact replay and zero source/document-13 reads passed. The ladder will not be expanded or repeated. #1014 subsequently completed the direct-learning pivot; see Delta_attn-off below.
d_R4, mu_R4ObjectiveIf reactivated, a separately trained intrinsic arm may score transported manifold-valued Q/K pairs and aggregate transported values with an artifact-bound equivariant geometric centroid mu_R4.V1 stopped UNAVAILABLE before D3; learned-manifold V2 failed donor retention and matched Euclidean parity; the 8/8-contract attempt stopped at its two-document preflight and rejected tangent readout with pooled normalized audit-MSE ratio 1.0643688804269025. Intrinsic attention remains unestablished and this objective is Parked.
A_M, K_hat, N_t, Z_tObjectiveA finite fiber-aware resonance amplitude, its nonnegative normalized kernel K_hat = floor + abs(A_M)^2, and the recurrent transported value-numerator and normalization-denominator mode sums that replace dense all-pairs weighting.ADR-0005 and the #973 multi-resonance reuse audit retain the conditional contract. The normalized sieve, recurrent factorization, and H4/Q29/integer lowering are Parked, NOT_IMPLEMENTED, and NOT_RUN.
I_geoDefinitionCausal inference step mapping observed prefix plus bounded hierarchy state to next-token scores/selection and updated state, followed by lexical decoding.A support trace alone is not inference or coherent generation.
Corr(D), Abst(D)Empirical CriterionCorrectness and abstention on declared distribution D: correctness uses an independent oracle/constraint/source and is reported both conditional on answered cases and over all cases; abstention is a typed outcome, never silently scored correct.Protocol and denominators are required by geometric_intelligence_evaluation.md.
Reason(D)Empirical CriterionOn anti-recall tasks from D, typed intermediate route transitions preserve constraints, compare alternatives or counterfactuals, and reach an independently checkable conclusion.Fluent text, recalled answers, teacher agreement, or non-zero geometric activity is insufficient evidence.
Serve_pfDefinitionProvider-free serving: the evaluated process performs no runtime call to Ollama, a cloud model, teacher endpoint, or other generative provider; all required artifacts and decoding are local and pinned.Does not imply transformerless, geometry-only, multiplication-free, correct, or production-ready.
M_R4^ℓDefinitionDeclared bounded causal geometric mixing operator at layer ℓ; its metric, chart, support rule, transport, and aggregation must be frozen by the owning issue.ADR-0005 accepts ordinary R4/Spin softmax attention as the reference baseline. Source-backed generation/native HTTP and the first source-free suffix-distillation compiler pass at their declared scopes, but the suffix student decoded a short repetition. R4SoftmaxTraceStateStudentV1 then completed negative without selection-bearing separation from its transport-permuted control. The construction-only O_trace^(0..3) ladder completed with insufficient per-fold support and no boundary attribution. #1014 then established ordinary causal attention as load-bearing at the directly learned R4/Spin scope through Delta_attn-off = 2.6773925609275944 nats and two-arm Rust parity; its complete quality DoD failed at enabled NLL 2.127407277216677 and subject/scene retention 3/5. Intrinsic attention replacement, resonance, reasoning, WASM promotion, and recurrent/exact lowering remain parked. A coherent geometry-native deployed realization remains NOT_YET_IMPLEMENTED.
Delta_attn-offEmpirical CriterionOn one frozen checkpoint and one identical sealed population, enabled causal attention NLL is subtracted from NLL after zeroing every attention output after W_o and before residual addition. The owning issue must preserve all other weights, inputs, decoding, and work.#1014 measured 4.804799838144271 - 2.127407277216677 = 2.6773925609275944 nats/token against a frozen >= 0.10 rule. Both policies matched Python/Rust top-1 within 0.005; all six layers had exact causal/R4/output-policy audits and zero future reads. This establishes load-bearing ordinary causal attention for that learned R4/Spin checkpoint, not geometry advantage, coherent quality, transformerlessness, or exact deployment. See r4_softmax_end_to_end_attention_1014.md.
d_R4^ℓDefinitionDeclared compatibility or least-energy score over factor overlap, prime-gap phase delta, R4/S3 transport, Hopf observation, and torsion.Normative G1R objective under #958; the score and its selected-support trace are NOT_YET_IMPLEMENTED. The learned d_R4^ℓ(q,k) from #950/#951 remains a negative comparator.
N_t^ℓDefinitionBounded causal neighborhood of prior route and memory states selected before value aggregation. Future positions are excluded.Exact-row reachability exists in the source-free substrate; the intervention-qualified geometric neighborhood and deployed trace/census are NOT_YET_IMPLEMENTED.
P_mem^ℓDefinitionDeterministic, tokenizer-CID-bound projection from an identity-scoped memory span into ordered prime-route and spin/torsion state.G0/G1 key/value adapter is historical; the factorable memory projection and product integration are NOT_YET_IMPLEMENTED under #958/#953.
Δ_geo-nullEmpirical CriterionDifference on the predeclared held-out metric between real geometry/memory and its equal-budget disabled or permuted control. A nonzero support trace alone is reachability, not advantage.Bounded one-layer report under #951.
SDefinitionSemantic state space; every observation projects into a state s ∈ S. Regions are subsets R_i ⊆ S; beliefs are predicates over S; goals are desired subsets of S.Compiler/reference-only abstraction (issue #124). The deployed runtime carries a fixed-capacity approximation: the runtime state (frontier, rolling context code, token shortlist) in crates/uor-r4-core.
G = (V, E)DefinitionCompiled semantic graph: V packed semantic states, E typed transitions.crates/uor-r4-graph-format GraphView over the R4G1 NODE/EDGE sections (docs/transformerless/R4G1.md).
H(x)DefinitionHolographic encoding of observation x: a family of overlapping projections {h_0, …, h_k} with partial recoverability, distributed evidence, progressive fidelity. The deployed path today is still the single compiled Boolean semantic code; the certifier additionally defines and measures overlapping projection families for issue #126.Runtime binding: sign-bit signature path in crates/uor-r4-core (see "Semantic code H(x)" in the glossary). Measurement contract binding: crates/uor-r4-graph-certify/src/holographic_encoding.rs and its deterministic fixture tests.
T : S × A → SDefinitionTyped graph dynamics: transition function over states and actions/semantic operators A.Compiler/reference-only (issue #124). Deployed precursor: forward transition edges E_f / R4G1 ROUT section.
RDefinitionReconstruction / behavioral-recovery operator; R(H(x)) ≈ x read behaviorally as the divergence condition below.Compiler/certifier-only; exercised through the fidelity-certification harness (score.rs, Gate C).
C : Θ → GDefinitionThe compiler as a map from teacher parameter space Θ to the space of compiled artifacts. Compilation is lossy semantic compression: parameters → behavioral probing → latent graph induction → Boolean synthesis → packed immutable artifact.crates/uor-r4-core compiler pipeline; graph generalization in crates/uor-r4-graph-compiler.
P_θ(·‖c)DefinitionTeacher distribution over next tokens for context c, pinned HF revision, deterministic mode.TeacherOracle next-token surface (crates/uor-r4-core).
P_G(·‖H(x))DefinitionRuntime/graph distribution produced by the compiled artifact.Graph scorer (crates/uor-r4-core score.rs; R4G1 adapter src/r4g1.rs).
D(·, ·), εDefinitionDeclared divergence measure and empirical tolerance for the behavioral reconstruction condition (below).docs/transformerless/EQUIVALENCE_AND_EMPIRICAL_PROTOCOL.md.

Behavioral reconstruction condition (the testable form of R(H(x)) ≈ x):

Empirical Criterion. D(P_θ(· | x), P_G(· | H(x))) ≤ ε for a declared divergence D and tolerance ε, measured on a pinned held-out distribution with confidence intervals. Status: Empirical; never a structural claim.

4. Objectives versus runtime invariants

These are Objectives — quantities the offline compiler optimizes. They must never be stated as runtime properties:

QuantityRoleBinding
J = L_teacher + λ·C_runtime + μ·C_artifactObjectiveCompiler cost model: teacher behavioral loss, inference cost, artifact size/complexity (issue #129).
min_Z I(Z;X) − β·I(Z;Y_future)ObjectiveInformation-bottleneck compression target: discard surface detail, keep future-relevant information (issue #127).
H(A | R), H(S_future | R)ObjectivePredictive-entropy criteria for splitting, merging, or removing regions (issue #127).
π* = arg max_π [ V(G | T(B,π)) − P(C,π) − R(U,π) ]ObjectivePlan-ranking target for bounded future-state optimization (issue #131). Not a theorem.

These are Guarantees — structural properties of artifact and runtime, each with a proof-model entry. Their current statuses are the proof-matrix records; a document citing one MUST cite the same status:

GuaranteeStatusEvidence
Allocation freedom on the prediction hot pathStructuralallocation_proof counting-allocator harness; allocation_census.rs (proof matrix: PDF §16)
Bounded packed rangesStructuralrange_bounds_proof (Theorem 8)
Deterministic top-K (canonical tie-break)Structuraldeterministic_topk_proof (PDF §23)
Forward/reverse index consistencyStructuraltheorem7_proof (Theorem 7)
Score arithmetic safety (no overflow/panic)StructuralKani-1 harness (kani_proofs.rs)
Fixed-capacity container invariantsStructuralKani-2 harness (kani_proofs.rs)
Inference operation-set conformanceWitnessed (Structural after machine-code audit)INFERENCE_OPERATION_CONTRACT.md + P-4 source scans (transformerless/mod.rs)
Termination, bounded frontier width, valid references, canonical serialization, provenance completenessper proof matrixR4G1 two-stage validation + proof-model entries; anything lacking an executed evidence artifact is Unproven

5. Term discipline (overloaded words)

Avoid (unqualified)Use insteadRule
"intent"Future-state optimization: belief = estimated current state, goal = desired future-state subset G ⊆ S, constraint = forbidden subset F ⊆ S, action = transition operator, plan = bounded trajectory π = (a_0, …, a_n) with T^π(s_0) ∈ G and T^π_i(s_0) ∉ F for all intermediate i (Definition, PDF §12).Unqualified "intent" is informal prose; it must not appear in a labeled statement.
"semantic atom"Semantic region — multiresolution, overlapping, proof-addressable; explicitly not an atom (glossary)."Semantic atom" is prohibited in normative text.
"equivalence"Qualified forms only: byte reproducibility (identical pinned inputs ⇒ identical artifact bytes) or behavioral equivalence (an Empirical Criterion, valid only on the declared distribution, per the equivalence protocol).Unqualified "equivalence" is informal; "exact equivalence" hits the §2.1 wording rule.
"reasoning"Precise mechanisms: typed state transitions, graph navigation, bounded planning (trajectory evaluation over T).Bare "reasoning" is informal; plausible language output is never evidence of it (§2.1).
"transformerless"Zero source-attention calls and no dense full-prefix Q·K matrix/softmax kernel on the exact promoted decoder path; the replacement uses bounded geometric support shown load-bearing by disabled/permuted interventions.Does not imply multiplication-free, integer-only, allocation-free, coherent, or production-ready. Name retained source components and the checkpoint/scope.
"geometric intelligence"Name the measured mechanism: geometric support change, memory-causal logit effect, student-prefix rollout, or another declared empirical result.Geometry being present or visually structured is not evidence of intelligence.

Existing documents predate this convention. Their already-qualified uses ("behavioral equivalence", "not semantic atoms", "equivalence-tested") conform; unqualified uses are hereby marked informal and are migrated opportunistically, not by wholesale rewrite.

6. Enforcement

  • The uor-r4-claim-wording crate (Rust port of the retired scripts/check_claim_wording.py) scans docs/**/*.md and crate README.md files and fails on §2.1 violations. Run locally: cargo run -q --locked -p uor-r4-claim-wording -- ..
  • Current PR and merge-group CI run the focused native checks described by the execution policy. Four historical required statuses acknowledge compatibility only. Run the claim-wording script when changing claims; broader legacy release checks remain separately activated. Neither status names nor unrun jobs are evidence.
  • The proof-status matrix (proof_matrix.rs) is the machine-readable registry for §2 statuses; verify_all fails on any Unverified entry.

Changelog

  • 0.1.35 (2026-09-01) — Recorded the independently verified R4PredictiveBlockDeltaPromptCapacityV5 terminal PREDICTIVE_BINDING_NO_TERMINAL_CAPACITY. Geometric gain 0.03896945868086732 with 375/512 wins missed the frozen absolute 0.04332169878499658 capacity floor. Geometry versus independently fitted plain missed the gain and own-NLL rules even though transport-permuted gates passed; full delta also failed attribution versus independently fitted additive at -0.006512463228773413 and 234/512. Fresh-language and integrity gates passed. A scoring-only recovery preserved the frozen arms, created no optimizer, and was independently replayed exactly. The binding action is STOP_WITHOUT_GENERATION; retire only this predictive block-delta law, without revising ordinary-softmax or retained-attention evidence. Result/scoring/recovery/verification CIDs are respectively blake3:6c67544d675eafcb8eb9c0dabb93617e3f6c3295af812e8acbb687107c010a74, blake3:44f8941d24a99fc230710fd700e7a7b13cee87587bfbe4e13bf7b095222e2ee6, blake3:7b76e36e44798bebf184ece08fdd8a2065bdd370106b5d64d5fae4c59dc6d88b, and blake3:567cf336eb05c3ec562aef7135f6fb35b580d02c758b0e79f2508cae57065f5d.
  • 0.1.34 (2026-09-01) — Recorded the independently verified R4LearnedAssociativeReadoutPromptCapacityV1 terminal LEARNED_ASSOCIATIVE_READOUT_NO_CAPACITY. Geometric gain 0.00637679 and pooled gain 0.01026323 both missed the frozen absolute 0.04332170 and incremental 0.02534157 capacity effects; geometry also lost the pooled and fixed-leaf-deranged attribution comparisons. Both learned arms passed every fresh-language and state-load-bearing gate, so the pooled improvement is retained as a non-geometric control signal. All mechanics and independent replay passed. No retry or generation is authorized; the next separately frozen #973 architecture must change the retained value write/binding law. Result CID blake3:cedba37738ee249457bb589f716ee75afb16a0c4937c2a22ae9f917dd3eb97c1; verification CID blake3:443d711ce9a228e26e2eb2eebb55c582848424e2677c3473d41deaf8afd69ec7.
  • 0.1.33 (2026-09-01) — Froze R4LearnedCandidateLeafAssociativeReadoutV1 and its independent R4LearnedAssociativeReadoutPromptCapacityV1 campaign before implementation, V4 population creation, fitting, or outcome access. Bound the exact-leaf, address-blind pooled, fixed-leaf-deranged, head-off, and state-off arms plus the fresh-data, create-once reveal, capacity, geometry-attribution, compute, and divergent-action contracts. All empirical outcomes remain NOT_RUN; #973 remains open and #954 blocked.
  • 0.1.32 (2026-09-01) — Recorded R4LayerwiseNormalizedRetainedReadoutLanguagePathV1 as terminal LAYERWISE_NORMALIZED_RETAINED_READOUT_PROMPT_CAPACITY_PARTIAL. All fresh-language, state-off, causal, replay, and independent-verification gates passed, while prompt gain 0.0286980210 and incremental gain 0.0213663973 missed their frozen floors. Ended the parameter-free readout ladder and made a freshly frozen learned associative binding/readout the sole #973 successor. Candidate generation, reasoning, and exact/geometry-native lowering remain NOT_RUN; #954 remains blocked.
  • 0.1.31 (2026-09-01) — Recorded R4DirectRetainedReadoutLanguagePathV1 as terminal DIRECT_RETAINED_READOUT_PROMPT_CAPACITY_PARTIAL. The zero-parameter readout increased prompt gain and fresh-language quality while passing state-off, causal, replay, and independent-verification checks, but missed both frozen prompt-gain floors. Generation and lowering remain NOT_RUN. Froze exactly one layerwise-normalized, zero-parameter successor and a stop/pivot falsifier; no coherence, reasoning, H4-superiority, exact-runtime, browser, or release claim follows.
  • 0.1.30 (2026-09-01) — Recorded qualified R4RetainedLanguagePathV1 and its terminal paired-H4 capacity successor. Paired addressing reduced construction repeats 97.5477% and slightly improved fresh-language NLL/top-1, but prompt gain fell from 0.0063672952 to 0.0062477543, with 282/512 wins. Terminal: PAIRED_H4_PROMPT_CAPACITY_FAIL; result CID blake3:508a4ff352f1e533d669d9616f65b972b0f13e8efe35867b7b095281ad940274. Preserved V1, rejected candidate generation, and directed #973 to an independently frozen prompt-state-to-logit readout seam. The canonical prompt population CID is blake3:c11a7c935139ca169460b90c01392d7c9e0929e4c10710e76e6c8f74cbdf0340; it replaced a provisional pre-freeze scan that omitted whitespace normalization and overlapped training data.
  • 0.1.29 (2026-09-01) — Recorded #973's completed CPU-recovery boundary. Disabling retained state on the disjoint construction-validation partition lost 0.967227 nats and 182 top-1 hits, qualifying a bounded causal retained-attention component. The exact 3.17M-parameter, two-block, data/dose recipe did not satisfy the frozen full-decoder generalization criterion; scrambled transport was 0.033049 nats better, so no H4-specific advantage follows. Directed #973 to a data-supported language-path decoder with an ordinary matched non-geometric control. Also made representative-step backend/thread/worker calibration the operating rule for substantial offline jobs; deployed runtime remains CPU/table-native and CUDA remains explicit-scope only.
  • 0.1.28 (2026-09-01) — Recorded the terminal #973 construction result UNAVAILABLE_FRAME_POPULATION_OR_LOCAL_BUDGET. Mechanical audits passed; timing and disposable learning smoke failed, so main and held-out model work stayed NOT_RUN with no attention or H4 verdict. Retired retries/tuning of the exact cell and directed #973 to select a fuller source-free decoder block.
  • 0.1.27 (2026-09-01) — Defined the independently re-scoped #973 R4GroupAddressedRetentionLMV1 and recorded only its observed construction boundary: geometry and population PASS; training and held-out model evidence NOT_RUN. Retained the older gated-delta, trace-state, intrinsic/readout, and resonance lanes as negative or parked history, with #954 still blocked and no C1-SB6 authorization.
  • 0.1.26 (2026-08-31) — Recorded the one bounded #1019 local fast-path result: an isolated exact-shape MPS test with 10 warmup plus 40 measured steps combined fused AdamW and deferred logging, measuring 4.485223 s/step versus the signed 3.491307 s/step. Because it was slower, fused=True was removed immediately. This is a bounded implementation negative, not a model or attention result. #1019 tuning/full-run work stops and remains optional/ paused; product work and #973 no longer wait for it. The active next step is using and productizing the bounded #1017 generator through r4 generate.
  • 0.1.25 (2026-08-31) — Corrected #1019's active execution scope without rewriting its frozen contract or measured preflight: UOR's deployed architecture/runtime remains CPU-native; Apple Accelerate/BLAS and MPS are permitted only for local offline training, compilation, and bounded tests; CUDA and external GPU execution are out of scope. The observed UNAVAILABLE_HARDWARE_BUDGET terminal applies only to the frozen eight-hour offline implementation. Reuse the passed population/smoke/parity artifacts and proceed prototype-first on the local M1: build, demonstrate, then harden one working efficiency mechanism without recurring broad research gates. #1019 is an optional quality-capacity improvement and no longer blocks use or productization of the bounded #1017 generator through r4 generate.
  • 0.1.24 (2026-08-31) — Bound #1019 as the sole parameter-capacity successor: twelve layers, exactly 13,130,784 parameters, seed 1019, 16,800 steps, and 275,251,200 tokens over the unchanged causal-softmax R4/Spin and Rust evidence path. Recorded every execution gate as NOT_RUN, required the cheap preflights and an eight-hour hardware projection, and prohibited paid external execution without explicit owner approval. This is a frozen Objective/Empirical Criterion, not a result claim.
  • 0.1.23 (2026-08-31) — Recorded #1017's frozen continuation at 149,995,520 cumulative tokens: enabled Rust parity, all mechanical gates, prompt retention 5/5, and normalized replay 5/5 passed; sealed NLL 1.5727521962806827 failed the strict <1.50 gate. The result is negative solely on NLL. Preserved #1014's bounded attention conclusion, prohibited further 7.15M exposure/LR tuning, and advanced only one predeclared parameter-capacity increase.
  • 0.1.22 (2026-08-30) — Added Delta_attn-off and recorded the completed #1014 split verdict: ordinary causal attention established at the learned R4/Spin intervention/parity scope; full language-quality DoD negative at enabled NLL 2.127407277216677 and prompt retention 3/5. The exact campaign closes without rerun or diagnostic expansion. Its #1017 exposure continuation then closed negative solely on sealed NLL 1.5727521962806827, with retention 5/5 and all other gates passing. One predeclared parameter-capacity increase over the unchanged mechanism is next.
  • 0.1.21 (2026-08-30) — Recorded #1012 INSUFFICIENT_SUPPORT_COVERAGE, forbade boundary attribution and further support/localization loops, and advanced #1014, direct end-to-end causal-softmax R4 attention training on a fresh untouched split with autonomous decoding. This stops the bounded current-step trace-distillation path, not all trace distillation or attention.
  • 0.1.20 (2026-08-30) — Bound the completed negative R4SoftmaxTraceStateStudentV1 result, exact metrics and artifact identities; narrowed its falsifier to the tested 4D signed-reduction/token-derived state cell; and defined the construction-only leave-one-document-out O_trace^(0..3) observability ladder in #1012, the native child/blocker of #973, as the next gate.
  • 0.1.19 (2026-08-30) — Promoted R4SoftmaxTeacherTraceV1 and R4SoftmaxTraceCompilerV1 from proposed objectives to a bounded empirical source-free-distillation result, retained the incoherent decoded-cycle boundary, and defined R4SoftmaxTraceStateStudentV1 as the next geometric state-compilation objective.
  • 0.1.18 (2026-08-30) — Recorded the exact opt-in loopback HTTP/CLI reference parity boundary and defined the proposed, not-yet-implemented R4SoftmaxTeacherTraceV1 / R4SoftmaxTraceCompilerV1 source-free trace-compilation objective.
  • 0.1.15 (2026-08-26) — Defined the geometric-intelligence route hierarchy and its claim boundaries: typed zero/identity and trigonometric chart bridges, lexical/prime/semiprime/n-let identities, zeta/Hopf/torsion/golden state, conceptual E8 = H4 × H4 realized by the witnessed icosian H4 ⊕ φH4 construction, hierarchical kappa and trajectory summaries, attention versus recall, inference, correctness, reasoning, and provider-free serving.
  • 0.1.14 (2026-08-20) — Added the issue-#830 register execution-scope, serving-reachability, and empirical-verdict vocabulary: a per-row execution scope (reference-only / offline-compiler / certifier-instrument / dormant-portable-runtime / normative-runtime / deployed-production) and a serving reachability (deployed-serving / off-serving-path / dormant-gated), plus a three-value empirical status (PASS / FAIL / UNAVAILABLE) kept as a separate axis from the harness-built build level — an absent fixture is UNAVAILABLE, never PASS (crates/repo-model/src/registry.rs, crates/repo-model/src/empirical.rs; rendered into CONFORMANCE.md).
  • 0.1.13 (2026-07-25) — Added issue-#175 compiler parallelism benchmarks and scaling certificate definitions (Compiler Parallelism Scaling Report, Multicore Thread Sweep Matrix, Stage Scaling Classification Taxonomy, Byte-Equality Scaling Premise in docs/compiler_scaling_certificate.md and uor-r4-graph-certify::compiler_scaling).
  • 0.1.12 (2026-07-25) — Added issue-#174 CPU-only compiler dependency and feature audit definitions (Compiler Dependency Denylist Gate, Default Feature Unification Rule, Teacher-Backend Isolation Invariant, CPU-Only Runner Compliance in docs/compiler_dependency_audit.md and uor-r4-graph-compiler::dependency_audit).
  • 0.1.11 (2026-07-24) — Added issue-#170 parallel observation, trace, and evaluation processing definitions (Content-Addressed Shard Partitioning, Ordered Deterministic Reduction, Shard ID Determinism, Teacher-Backend Boundary in docs/parallel_observation_shards.md and uor-r4-graph-compiler::observation_shards).
  • 0.1.10 (2026-07-24) — Added issue-#169 compiler memory-budget and backpressure model definitions (Concurrency-Aware Memory Formula, Per-Stage Memory Estimate, In-Flight Backpressure Limiter, Constrained-Memory Determinism in docs/compiler_memory_budget.md and uor-r4-graph-compiler::memory_budget).
  • 0.1.9 (2026-07-24) — Added issue-#168 compiler thread-pool and jobs configuration definitions (Compiler Concurrency Control, Jobs Precedence Resolution, Dedicated Thread-Pool Ownership, Oversubscription Policy in docs/compiler_concurrency_config.md and uor-r4-graph-compiler::jobs_config).
  • 0.1.8 (2026-07-24) — Added issue-#167 normative reproducibility & canonical byte-equality definitions (Normative Reproducibility Invariant, Parallel Reproducibility Harness, Thread-Count Invariance, Deterministic Reduction Policy in docs/reproducibility.md and uor-r4-graph-compiler::reproducibility).
  • 0.1.7 (2026-07-24) — Added issue-#166 compiler stage ownership and parallelization DAG definitions (Compiler Stage DAG, Parallel-Safe Stage, Deterministic Merge Stage, Bounded Parallel Stage, Sequential Canonical Finalization Spine in docs/compiler_stage_dag.md and uor-r4-graph-compiler::stage_dag).
  • 0.1.6 (2026-07-24) — Added issue-#165 deterministic compiler executor definitions (Compiler Executor Abstraction, Sequential Reference Executor, Rayon Parallel Executor in uor-r4-graph-compiler::executor).
  • 0.1.5 (2026-07-24) — Added issue-#161 runtime operation, allocation, and CPU portability certificate definitions (Runtime Performance Certificate, Evidentiary Class Schema, Declared-Zero Evidence Link, CPU Portability Record in uor-r4-graph-certify::performance_certificate).
  • 0.1.4 (2026-07-24) — Added issue-#160 machine-code, allocator, and dependency CI audit definitions (Machine-Code Disassembly Audit, Counting Allocator Witness, Dependency Denylist Gate in uor-r4-proof-model::inference_audit).
  • 0.1.3 (2026-07-24) — Added the issue-#158 normative scoring semantics definitions (Fixed-Point Scoring Semantics, Residual Taxonomy, Overlap Residualization, Deterministic Tie-Breaking in docs/scoring_semantics.md and uor-r4-graph-format::scoring_semantics).
  • 0.1.2 (2026-07-24) — Added the issue-#157 normative inference contract definitions (Normative Inference Contract, Permitted Operation Class, Zero-Allocation Steady State, CPU-Only Target Contract in docs/inference_contract.md and uor-r4-graph-format::inference_contract).
  • 0.1.1 (2026-07-24) — Added the issue-#126 measurement-contract binding for H(x) (projection family schema, ablation semantics, and deterministic partial recovery/progressive-fidelity fixture in uor-r4-graph-certify).
  • 0.1.0 (2026-07-24) — Initial version. Claim classes, claim statuses, core notation (S, G=(V,E), H(x), T, R, C, P_θ, P_G), objectives vs. runtime invariants, term discipline, and the CI wording rule. (Issue #123.)