LeaTTa: machine-checked MeTTa semantics in Lean 4

LeaTTa is a Lean 4 development for MeTTa, the MeTTaIL runtime path, a distributed atomspace model, and PoR-weighted Cordial Miners. The executable kernel ports Hyperon's minimal MeTTa interpreter, the small instruction set under the rest of MeTTa. The standard library is written in MeTTa over that kernel. The runnable interpreter is total, fuel-bounded, and Mathlib-free.

Version 1.0.8 adds the MORK/MM2 resource-readback proof surface on top of the 1.0.7 distributed atomspace release. The public branch has five checked surfaces:

• the minimal interpreter and standard library run Hyperon's oracle corpus: 270 passing assertions across 22 vendored files;
• Metatheory and Operational prove determinism, confluence of the deterministic fragment, type soundness, indexing soundness and completeness, query correspondence, observation records, MORK-backed query readback, MM2 lowering, ACT resource isolation, and the published four-register machine properties;
• Distributed models replica-local atoms, mutation events, vector clocks, fairness as a theorem parameter, ordered replay assumptions, and matching convergence;
• MeTTaIL, MeTTaILProofs, and MeTTaILTests formalize the determinate presentation pipeline and ship a small editable dialect that runs through the checked reducer;
• CordialMiners formalizes the PoR-weighted coarse consensus protocol, proves the top-level safety theorem, and hosts the protocol as a MeTTaIL runtime presentation.

The branch does not claim the full Hyperon module system, MeTTa on Rholang, unordered distributed convergence without replay assumptions, the full MeTTaIL-to-rho theorem, or the remaining denotational goals. Public theorem targets build with no sorry, admit, native_decide, partial, or unsafe. The longer Hyperon comparison is in the book appendix at mestto.github.io/LeaTTa.

HE evaluator repair batch

The he-conformance-repairs branch collects the complete evaluator repair batch from 26d05f0 on top of MesTTo's metatheory at 3885010. It keeps Lean 4.31.0 and incorporates the signature-selection proof corrections from the Lean 4.33.1 integration commit e5c5d36.

Repair area	Implementation and evidence
Equality-class reconciliation, transient aliases, malformed unify errors	Already included in the upstream base: aab8a11, 8614cd1, b052568
Idempotent alias replay and equality saturation	Core/Bindings.lean, Proofs/BindingLaws.lean; nested query regressions in he-conformance.metta
Cyclic matching and numeric type matching	Core/Matching.lean, Core/Atom.lean, Minimal/Interpreter.lean; matching regressions
Failed application-type matches, annotation freshening and intrinsic grounded types	Proofs/TypeSoundness.lean, Proofs/TypeInferenceFreshening.lean; typeops.metta
Ordered overload selection, argument masks, dependent return checks and complete diagnostics	Proofs/MultipleSignatureSelection.lean; typeops.metta
Capture avoidance in rules, queries and returned space atoms	Minimal/CaptureAvoidingFreshening.lean, Proofs/CaptureAvoidingFreshening.lean; freshness.metta
Live applicability bindings and caller-visible result bindings	The shared application plan in Minimal/Interpreter.lean; typeops.metta
First changed Empty or Error argument, quoted terminal data and operator evaluation	he-conformance.metta
Embedded metta expected type and selected space	typeops.metta, spaces.metta
Lossless collapsed binding payloads and conflict checks on resumption	Core/Bindings.lean, Proofs/BindingLaws.lean; control.metta
Grounded context-space handles, error-head classification and symbol meta-types	Core/Atom.lean, Proofs/Basic.lean; he-conformance.metta, typeops.metta

Implementation paths in this table are relative to MettaHyperonFull/; fixtures are under tests/regression/. The runnable kernel remains Mathlib-free. The original runtime repairs are preserved in full, with checked proofs and additional regression coverage.

Validation on 2026-10-13: the full default build passes, and the regression suite passes 101/101 plus the MeTTaIL runtime check. License-header, forbidden-placeholder and build-warning checks pass; NOTICE is unchanged. The complete vendored corpus passes 270/270 across all 22 files, using the original assertions, default fuel and existing timeout.

A separate check passes 67 assertions from typeops.metta and he-conformance.metta with the actual Hyperon 0.2.10 release:

python -m pip install hyperon==0.2.10
python scripts/run-he-conformance.py
# HE CONFORMANCE OK: 67/67 assertions

An earlier revision of this repair branch timed out in four corpus files. Live binding propagation caused repeated merges to accumulate duplicate aliases: one backchaining merge contained 7,201 relations but only six distinct constraints. Raw alias insertion is now idempotent, with proofs that it preserves the relation set and leaves the existing binding list unchanged when an edge is replayed. Compiler rewrite theorems also prove that stopping saturation at its fixed point and computing the closure once per ordered lookup return exactly the original results for every input and fuel. The repairs to binding propagation remain in place. Two small HE-checked nested-query regressions cover the tuple-head and sibling cases that exposed the slowdown.

Three alternating paired runs on the same machine give a median corpus runtime of 3.291 seconds for this branch and 2.028 seconds for an isolated upstream 3885010 build. Both measurements include process startup and use the same corpus and default fuel.

Previously hanging file	This branch, median	Upstream, median
b3_direct.metta	32.4 ms	10.2 ms
b2_backchain.metta	135.6 ms	20.8 ms
test_stdlib.metta	740.9 ms	1761.3 ms
b5_types_prelim.metta	299.6 ms	11.2 ms

The hangs are resolved; some per-file overhead remains.

Use a Python version supported by that Hyperon release. The existing freshness.metta fixture deliberately uses an internal # variable spelling; Hyperon's source parser rejects that spelling, so this fixture tests LeaTTa's internal collision handling separately.

This batch does not establish full equivalence to HE. For example, malformed arrow annotations still differ:

(: bad-nullary (->))
(: bad-nullary (-> R))
!(get-type (bad-nullary))

This branch returns R; HE 0.2.10 also returns ((->)). The branch filters the malformed candidate. Stateful evaluation order, module-loading coverage and the mixed tuple/function policy also remain outside the differential guarantee above. PLeaTTa's Prolog profile and the match-by-equality prototype belong to separate branches.

Start here

Run the public surfaces from a checkout:

lake build
lake build Distributed Metatheory Operational
./scripts/run-oracle.sh
./scripts/run-regression.sh
lake build MeTTaIL MeTTaILProofs MeTTaILTests
lake build CordialMiners CordialMiners.Runtime.Run

Expected results:

• lake build ends with Build completed successfully;
• Distributed, Metatheory, and Operational build directly from their public roots;
• ./scripts/run-oracle.sh reports ORACLE OK with 270 passing assertions;
• ./scripts/run-regression.sh reports REGRESSION OK;
• MeTTaIL and Cordial Miners build without forbidden placeholders.

Quick MeTTaIL runtime check

The executable can run an editable MeTTaIL dialect file. The fixture tests/mettail/bool.mettail declares:

sort Tm
term tt : Tm
term ff : Tm
term notOp : Tm -> Tm
rewrite notTt : (notOp tt) => ff
rewrite notFf : (notOp ff) => tt

Run it:

lake exe LeaTTa --mettail tests/mettail/bool.mettail --term '(notOp tt)'   # ff
lake exe LeaTTa --mettail tests/mettail/bool.mettail --term '(notOp ff)' --fuel 100   # tt
lake exe LeaTTa --mettail tests/mettail/bool.mettail --term '(notOp nope)' # (notOp nope)

Malformed input is rejected before reduction:

lake exe LeaTTa --mettail tests/mettail/bool.mettail --term '(notOp tt'
# stderr: MeTTaIL term did not parse

You can test a new dialect without changing tracked files:

tmp=$(mktemp -t mettail-readme.XXXXXX)
printf '%s\n' \
  'sort Tm' \
  'term a : Tm' \
  'term b : Tm' \
  'term step : Tm -> Tm' \
  'rewrite stepA : (step a) => b' > "$tmp"
lake exe LeaTTa --mettail "$tmp" --term '(step a)'   # b
rm -f "$tmp"

The regression suite includes that CLI path:

./scripts/run-regression.sh
# REGRESSION TOTAL: PASS=101 FAIL=0
# mettail-runtime: PASS
# REGRESSION OK

The Lean test target also checks the parser and runtime examples:

lake build MeTTaILTests
# Except.ok (some "ff")
# Except.ok (some "tt")
# Build completed successfully

Minimal MeTTa kernel

The executable kernel is in MettaHyperonFull/Minimal/.

Interpreter.lean is a faithful port of interpreter.rs: a continuation-passing nondeterministic stack machine with the thirteen minimal instructions eval, evalc, chain, unify, cons-atom, decons-atom, function, return, collapse-bind, superpose-bind, metta, metta-thread, capture, and context-space. One step is total. The driver is fuel-bounded because MeTTa programs can diverge.

Stdlib.lean implements the standard library in MeTTa over those instructions: if, let, let*, switch, case, map-atom, filter-atom, foldl-atom, set operations, assertions, match, and grounded operations.

The implemented runtime surface includes gradual get-type, application checking with (BadArgType ...), multi-type symbols, type-variable unification for parametric and dependent signatures, mixed integer and float arithmetic, named spaces, add-atom, remove-atom, get-atoms, match over &self and named spaces, conjunctive (, ...) match, state cells, bind! tokens, and import! as the only IO. Hyperon represents most of that state through Rc<RefCell> mutation. LeaTTa threads a pure World of spaces, state cells, and tokens through evaluation. Sequential file evaluation is preserved, so each ! query sees exactly the knowledge base before it.

Oracle validation

The oracle gate runs Hyperon's unmodified corpus vendored under tests/corpus/ (MIT, commit 3f76dc4). An assertion passes when it evaluates to (). The script builds the interpreter, runs every ! assertion, compares expected output, and exits nonzero on any mismatch or timeout. On this repair branch, all 270 assertions pass across the 22 included files.

./scripts/run-oracle.sh                                  # 270 / 270, ORACLE OK
lake exe LeaTTa --oracle tests/corpus/test_stdlib.metta    # single file

Hyperon test file	result
lib/tests/test_stdlib.metta	39 / 39
a1_symbols, a2_opencoggy, a3_twoside	7 / 7, 1 / 1, 4 / 4
b0_chaining_prelim, b1_equal_chain	5 / 5, 8 / 8
b2_backchain, b3_direct	6 / 6, 4 / 4
b4_nondeterm.metta	11 / 11
b5_types_prelim.metta	26 / 26
c1_grounded_basic.metta, c3_pln_stv.metta	21 / 21, 5 / 5
c2_spaces.metta	25 / 25
e1_kb_write, e2_states, e3_match_states	3 / 3, 14 / 14, 9 / 9
d1_gadt.metta, d2_higherfunc.metta	14 / 14, 25 / 25
d3_deptypes, d4_type_prop, d5_auto_types	8 / 8, 18 / 18, 7 / 7
g1_docs.metta	10 / 10

The passing tier includes GADTs, higher-order functions, dependent length arithmetic, types as propositions, auto type checking, get-doc, and help!.

f1_imports.metta is excluded because Hyperon marks it Python-mode-only. Its header says it does not work in no-Python mode because it assumes &self starts nearly empty with corelib and stdlib as separate modules. LeaTTa ships the prelude inside &self. The module behavior it would exercise, import! into named spaces and diamond-dependency deduplication, is covered by c2_spaces and g1_docs.

Proof surface

The kernel proof layer is MettaHyperonFull/Proofs/. It uses Mathlib and proves the replay properties that matter for on-chain MeTTa:

• the abstract machine is deterministic while nondeterminism stays in the result list;
• the deterministic fragment is confluent;
• first-argument rule indexing is sound and complete, including the same-head case where Hyperon's Space::visit undercounts;
• the gradual type checker is total and reports BadArgType only for a real argument type;
• alpha-equivalence is an equivalence relation;
• binding merge laws expose successful, conflicting, and unification-mediated cases;
• atomspace laws cover insert, remove, query, type assignment, equality rules, and threaded-world visibility;
• substitution-cycle audit facts state where direct-loop checks stop and recursive resolution needs an acyclicity condition;
• host-law records make grounded equality, matching, typing, execution, and display assumptions explicit;
• MORK-facing laws cover duplicate-preserving query backends, codecs, decoded binding rows, prepared query snapshots, sharded counts, named spaces, grounded host handles, MM2 lowering, and ACT resource isolation;
• observation records expose input, fuel, result atoms, error atoms, exhaustion status, and before/after worlds without a second evaluator;
• arrow-constructor laws pin Atom.mkArrow and TypeEnv.arrowParts? to the executable splitter.

MettaHyperonFull.Operational.* checks the published Meta-MeTTa operational semantics (arXiv 2305.17218): the four-register machine (i,k,w,o), barbed bisimulation, and a gas-bounded extension. Proofs/Correspondence.lean bridges the indexed kernel to that specification.

MettaHyperonFull.Distributed.* is a distributed atomspace proof target, not the Cordial Miners consensus proof. It models replica-local storage, add/remove events, vector clocks, local issue, remote delivery, fairness, ordered replay, and convergence. Checked claims include read-your-own-writes, mid-flight divergence before delivery, append-only logs, barrier extension, quiescent coverage, vector-clock max as a least upper bound, ordered atom-set convergence under OrderedReplayAssumptions, and matching convergence after ordered replay.

lake build Distributed

MORK and MM2 layer

The 1.0.8 proof addition models the MORK-facing boundary without importing a byte trie or native host runtime into the kernel. The core files are:

• Core/QueryBackend.lean and Proofs/QueryBackend.lean: duplicate-preserving query backends, reference backends, generated reference backends, and serial conjunctive-query laws;
• Core/MorkCodec.lean, Core/MorkCompactCodec.lean, and their proofs: logical and compact newVar/varRef codecs with decode-after-encode laws;
• Core/MorkEncodedSpace.lean, Core/MorkDecodedBindings.lean, and their proofs: encoded-space query refinement, repeated-variable coreference, decoded binding rows, and query/data separation;
• Core/MorkNamespace.lean, Core/MorkNamedSpaces.lean, Core/MorkPrepared.lean, Core/MorkSharded.lean, and their proofs: query-result namespace separation, named-space visibility, snapshot equivalence, prepared-query equivalence, and finite-sharding count laws;
• Core/MorkGroundedFilter.lean, Core/MorkGroundedRegistry.lean, and their proofs: mutable grounded rows filtered by stable live values and host handles;
• Core/MorkMM2.lean, Core/MorkMM2Lowering.lean, Core/MorkMM2Resources.lean, and their proofs: semantic MM2 exec consumption, lowering for I/, sources and O/, templates, equality and inequality source constraints, add/remove effects, priority-ordered exec selection, and a pure ACT resource boundary where resource names stay isolated from the main reference Space.

The layer intentionally does not claim byte-level path priority, mmap ACT files, host solver callbacks, WASM/native resources, lock scheduling, or the byte trie store.

MeTTaIL and rho-facing research

MeTTaIL/ formalizes the determinate part of F1R3FLY's Meta Type Talk Intermediate Language (https://github.com/F1R3FLY-io/MeTTaIL). A presentation declares sorts, constructors, equations, and rewrites for an object calculus. LeaTTa checks the presentation pipeline and runs the resulting rewrite system through a verified reducer.

Checked today:

• elaborate, desugar, type-lift, and monomorphize, pinned by decide against output captured from the Scala tool on Rholang.module;
• the GSLT reduction relation and soundness theorem;
• subject reduction and confluence for SKI and simply typed lambda calculus presentations;
• spice bounded reachability and mq-calculus Born-rule probability conservation;
• the MeTTa-to-GSLT bridge and presentation lattice laws;
• an executable reducer generated from a presentation;
• the AC-aware runtime fragment used by Cordial Miners;
• labelled transition systems, simulation, bisimulation, denotational kernels, context congruence, full abstraction, observational calibration, path-key RSpace branches, and costed transition systems;
• red/black reflective universes, quote/drop round trips, a generic final-coalgebra surface, and a bridge from final-behaviour models to FullyAbstractModel;
• the rset hereditarily finite core, colour automaton, atom opacity, red/black swap, finite support for atom renaming, extensional quotient, and union laws;
• the interacting trie-map surface RITM = ITM[RITM, 1 + RITM, RITM] and Jetta-style packed binding addresses comparable with path-key RSpace prefixes;
• rho syntax, quote/drop, ordinary and persistent send/receive COMM, | structural congruence, a one-channel RSpace boundary, and the K-shaped <In>/<Out> machine cases for the four persistence combinations;
• the first packet-level MeTTaIL-to-rho compiler bridge. A successful applyBaseRewrite is a real Reduces step, emits the encoded contractum at the source term location, and reifies the packet listener as a ready K communication step modulo | 0.

The public source references for the rho side are the F1R3FLY f1r3node reducer (https://github.com/F1R3FLY-io/f1r3node/blob/dylon/mettatron/rholang/src/rust/interpreter/reduce.rs) and the direct-rule compiler shape in mettail-rust branch GSLT2rho (https://github.com/F1R3FLY-io/mettail-rust/blob/GSLT2rho/gslt2rho/rho_compile/src/compile.rs). LeaTTa uses those as implementation guides, not as completed semantics.

Other MeTTaIL research surfaces are present as interfaces:

• RSet.lean covers the finite red/black set core from the rset paper and still leaves the recursive element quotient, algebraic-set-theory fixpoint, and final behaviour coalgebra open;
• Denotational.lean states labelled transition, bisimulation, full abstraction, path-key RSpace, and costed-system transfer theorems but does not construct the knotted topos, trie store, cut law, cost endofunctor, or language-specific calibration theorem;
• CostRoundTrip.lean states the continued-cost wrapper, starved-deadlock theorem, wrapped trace preservation, unit-token cost, and T o C image/fixed-point theorem;
• Native.lean, NativeGrammar.lean, and NativeTypes.lean state the native carrier, native grammatical-formalism, and OSLF native-type obligations. The grammar source is the public MeTTapedia GF/OSLF paper (https://github.com/zariuq/MeTTapedia/blob/main/papers/native-grammatical-formalism.pdf). These files define the interface a GF or MeTTaIL frontend has to instantiate. They are not a GF parser and do not import MeTTapedia;
• Hypercube.lean checks the finite modal/spatial center: sort slots, executable equation filters, explicit slot constraints, generated rule schemes, judgment footprints, and footprint containment. Generated typing rules and subject reduction are future work;
• KnottedUniverse.lean states reflective red/black universes, quote/drop equivalences, structure-preserving morphisms, coalgebras for a type endofunctor, and the terminal-coalgebra witness shape. It does not construct the knotted topos.

Build the layer with:

lake build MeTTaIL MeTTaILProofs MeTTaILTests

The axiom audit for this layer reports only the standard axioms used elsewhere in the project: propext, Classical.choice, and Quot.sound.

Verified spec-to-runtime path

The CLI file path goes from a small presentation file to a running reducer: parse declarations, monomorphize, run the term through the checked evaluator, and print the normal form. The one-step engine is sound and, for base rewriting, complete. It terminates under a checked measure and is confluent by Newman's lemma when termination and local confluence are supplied.

The AC layer gives a verified total order on terms, canonical forms that decide AC-equivalence, modulo-AC stepping sound for R/AC, and Church-Rosser modulo AC. ACMatch proves the scoped matcher used by Cordial Miners: one fixed subpattern plus one rest variable, following Eker's elementary AC matching and Dundua, Kutsia, and Marin's variadic equational matching.

The type layer gives head-sort preservation, all-subterms well-sortedness, substitution lemmas, matching lemmas, and subject reduction for one-step and many-step reduction. OSLF, OSLFRec, and OSLFCat formalize the first-order spatial and behavioural logic layer from "Logic as a Distributive Law": possibility as closure, necessity as interior, spatial composition as a bifunctor, and the order-theoretic Beck theorem for composing closures.

CriticalPairs proves the Critical Pair Theorem for a first-order term model with positions. Local peaks split into disjoint, variable-overlap, or critical-pair cases, yielding the Knuth-Bendix-Huet criterion: terminating, left-linear systems with joinable critical pairs are confluent.

MeTTaIL/Runtime/LanguageFile.lean is the checked CLI file format. It is not the full BNFC MeTTaIL surface. The remaining research items are in module headers and the proof-status appendix.

Cordial Miners

CordialMiners/ formalizes PoR-weighted Cordial Miners, a leaderless DAG-based BFT consensus protocol (arXiv 2205.09174), using the weighted Proof-of-Reputation variant from a Ben Goertzel blueprint.

The headline theorem is end_to_end_safety_of_finality_permanence: under a Byzantine-weight bound, honest non-equivocation, and finality permanence, the protocol never finalizes conflicting values and correct miners never publish conflicting positions. Prefix monotonicity is derived. The proof passes through weighted-overlap arithmetic, threshold finality, blocklace closure, equivocation detection, final-leader ratification, verified topological sorting, a coarse/fine simulation, lossless extraction to MeTTaIL atoms, executable demos, and the runtime bridge.

lake build CordialMiners
lake build CordialMiners.Runtime.Run

Expected runtime checks:

info: CordialMiners/Runtime/Run.lean:88:0: true
info: CordialMiners/Runtime/Run.lean:121:0: true
info: CordialMiners/Runtime/Run.lean:223:0: true

The runtime bridge encodes protocol facts and input events into MeTTaIL.AST, defines cmPresentation with six AC-marked coarse rewrite rules, and runs them with evalAC'. The main forward theorem is:

trec_step_forward_decode :
  TrecState.Step s s' ->
  ∃ events target,
    RewStepModAC acOpCM cmPresentation (encConfig eW eH events s) target ∧
    astToState dW dH target = s'

The backward theorem says a shaped runtime step modulo AC decodes either to one coarse step or to the same coarse state, provided field decoders respect ACEq acOpCM. The executable Nat/Nat instance discharges that condition as dNat_acEq and instantiates runtime_step_backward_nat.

Reader-facing details are in CordialMiners/README.md and CordialMiners/Runtime/README.md.

Documentation and book

The published site at mestto.github.io/LeaTTa contains:

• the Verso book at the site root;
• the doc-gen4 API reference at /api, with Mathlib references linked to official Mathlib documentation.

The book source is in book/. It covers the object language, interpreter, type system, metatheory, operational semantics, kernel correspondence, blockchain angle, distributed atomspace, MeTTaIL, Cordial Miners, future work, and limitations. Build it separately:

cd book
lake exe docs
python3 -m http.server 8137 --directory _out/html-multi

Open http://localhost:8137/ or book/_out/html-multi/index.html.

Install and run

The release binary is a single native LeaTTa executable. It links only against the standard C library, so a Linux x86_64 bundle does not require the Lean toolchain.

tar xzf leatta-1.0.8-linux-x86_64.tar.gz
cd leatta-1.0.8-linux-x86_64
./install.sh
LeaTTa --min '!(+ 1 (* 2 (- 10 4)))'   # [13]

Full install and source-build notes are in INSTALL.md.

Build from source:

lake build
lake exe LeaTTa --min '!(+ 1 (* 2 (- 10 4)))'                 # [13]
lake exe LeaTTa --min '!(map-atom (1 2 3) $x (* $x $x))'      # [(1 4 9)]
lake exe LeaTTa --min '!(case (+ 1 1) ((1 one) (2 two)))'     # [two]
make release
./scripts/run-oracle.sh

Where LeaTTa improves the current interpreter

The minimal interpreter in hyperon-experimental is marked provisional at several decision points. LeaTTa replaces those points with total checked constructs:

• the mutable is_evaluated() bit, commented as a hack in interpreter.rs, becomes static return-type gating from the declared function type;
• the is_variable_op hotfix becomes a total isVariableHeaded guard;
• Rc<RefCell> mutation becomes a pure immutable stack;
• the global make_unique counter becomes a threaded pure gensym;
• the unbounded loop becomes a fuel-bounded driver with a checked termination measure;
• linear rule lookup becomes explicit first-argument indexing.

The full comparison table is in the book appendix.

Layout and scope

• Active kernel: MettaHyperonFull/Core, MettaHyperonFull/Runtime.Parser, MettaHyperonFull/Minimal.Interpreter, and MettaHyperonFull/Minimal.Stdlib.
• Active proofs: MettaHyperonFull/Proofs, MettaHyperonFull/Operational, MettaHyperonFull/Distributed, MORK/MM2 proof modules, MeTTaIL proof modules, and Cordial Miners.
• Runtime presentations: MeTTaIL, MeTTaILTests, CordialMiners.Runtime.
• Archived and not built: earlier exploratory four-register runtime, categorical metagraph rewriting, Ruliad sketches, and an earlier approximate standard library under archive/.
• In scope: minimal interpreter, standard library, runtime type system, mixed arithmetic, mutable spaces and state, conjunctive match, metatheory, published operational semantics, distributed atomspace, MORK/MM2 query/resource proof interfaces, MeTTaIL runtime path, and Cordial Miners safety.
• Open: full Hyperon module system, unordered distributed convergence without replay assumptions, full MeTTaIL-to-rho theorem, full Rholang/RSpace runtime semantics, trie storage, cut law, cost endofunctor, and language-specific observational calibration.