AletheiaGraph · Testbed & Qualification

Maturity is evidence, not elapsed time.

AletheiaGraph treats maturity as the durable accumulation of evidence about what the engine can and cannot be trusted to do. Calendar age and installed base matter because they create unpredictable field exposure, but they are proxies for experience—not the definition of maturity itself.

Passing a functional test shows that an operation can work. Qualification asks what must remain true, how deliberately we tried to break it, and whether the evidence survives recovery, repetition and regression.

Why this page exists

Software has traditionally acquired much of its confidence through elapsed production time. A broad installed base supplies accidental chaos experiments: unusual workloads, timing combinations, hardware failures, operator mistakes and interactions its designers did not anticipate. That evidence is real. But field age is only one way a system accumulates experience. A modern qualification program can deliberately manufacture many classes of adversity, preserve every discovered failure as an executable regression, and rerun that accumulated experience against every subsequent change.

Open-world evidence

Field exposure

Production diversity reveals unknown interactions that no finite laboratory program can guarantee it has anticipated.

Engineered evidence

Qualification depth

Defined invariants are attacked through deterministic failure injection, concurrency, recovery, scaling, semantic equivalence and boundary cases.

Durable experience

Failure memory

A discovered defect becomes a reproducible test and permanent regression gate, so the system does not have to relearn the same lesson through institutional memory.

The relevant maturity question is therefore not simply “How old is the software?” It is “What has the software been required to prove, under what conditions, and what evidence is retained?”

The development testbed

AletheiaGraph's current hardening work is being performed on the FreeBSD family, presently using GhostBSD for development convenience, with SBCL as the Common Lisp runtime. Qualification runners are designed to make test setup explicit and repeatable. Exact hardware, filesystem and revision metadata should accompany published benchmark results rather than being hidden behind generalized performance claims.

Qualification loop
Define invariantwhat must remain true?
→
Create adversaryfailure / scale / stale state
→
Executerepeatable runner
→
Reopen / compareauthoritative verification

Current qualification snapshot

The testbed is broader than a unit-test counter. We publish the smallest defensible numbers we can reconstruct from surviving qualification output, and keep larger durability, scaling, recovery and multi-backend runs separate rather than pretending every run is equivalent to one assertion.

370+
enumerated low-level tests

From the ordered-index baseline alone, before later qualification stages.

35+
additional qualification stages

Named durability, index, planner, reasoning and integrated qualification runs.

2
physical index backends

Skip List and B+ tree paths are both exercised.

100k
record-scale qualification

B+ tree scaling work extends beyond toy graphs and micro-fixtures.

Pass count is not a maturity metric by itself. A deterministic killpoint recovery run, full-GC lifecycle test, or 100,000-record scaling qualification should not be presented as equivalent to one tiny unit assertion. We publish counts together with test intent and failure model.

Qualification ledger by wall

Status terminology: Qualification complete means the currently defined AletheiaGraph qualification campaign for that subsystem has passed. It does not imply formal certification, decades of production field history, or that future regressions are impossible.
AreaWhat was exercisedCount we can safely publishStatus
Authoritative write-path, concurrency & recovery — Wall ATransaction correctness, WAL durability, deterministic crash/killpoint recovery, shared-node concurrency, OCC conflict semantics, write intent/reservation, publication ordering, checkpoint correctness, multi-key overlap, dirty-read prevention, after-apply hooks and recovery conflict semantics.21 canonical qualification stages: 20 subordinate runners + 1 current-semantics multi-key overlap gate. Includes a 16,000-transaction shared-node lost-update workload and 200 multi-key transactions / 400 node updates.Qualification complete
Indexing, access paths & index integrity — Wall BOrdered and single-field indexes, exact/range lookup, ASC/DESC traversal, LIMIT, keyset pagination, composite and directional composite indexes, statistics, verify/repair, reopen/recovery integrity, transactional maintenance, tombstones, lifecycle/GC, cold/warm reverse lookup, traversal indexes, backend parity and scaling.370 enumerated ordered-index baseline tests + approximately 13 additional named qualification stages.Qualification complete
Native query, planning & execution — Wall CGraph-native query IR/algebra, explicit scan and index access paths, statistics and cost estimation, plan selection, missing/unbound-field semantics, ordered execution, pagination, projection, composite indexes, joins, aggregation, EXPLAIN/plan introspection, execution integration and fallback behavior.Exact individual case count not yet reconstructed; approximately 16 named qualification stages across 3A–3H plus P10/P11/P12 closure.Qualification complete
Reasoning Engine Safety & Correctness — Wall DCore semantics, recursion, functors, graph predicates, mutation, concurrency, repeated-query isolation, inference budgets, cycle protection, timeouts and spatial reasoning.4 explicitly identified hardening cases + 6 major suite groups.Qualification complete
Reasoning Durability & Recovery — Wall ECommitted-fact visibility, reopen consistency, durable retract, denied-write behavior, deterministic re-query and inference limits after reopen.8 integrated qualification checks.Qualification complete
Epistemic Classification — Wall F1:stored, :inferred and :proposed semantics, multi-graph schema, graph-local persistence, validation, transaction provenance and reopen durability.6 multi-graph qualification checks.Qualification complete
Proof Topology & Evidence Provenance — Wall F2Vertex evidence, edge evidence, multi-support conclusions, reusable support records, source ID + commit-epoch, graph isolation, reopen durability and authority preservation.10 qualification gates.Qualification complete
Evidence Validity & Invalidation — Wall F3Classifies cited evidence as :current, :stale, :retracted or :missing; preserves immutable historical citations; keeps validity support-local, read-only, graph-isolated and reopen-durable.10 qualification gates.Qualification complete
Authority Transition / Promotion — Wall F4Inferred/proposed claims remain non-authoritative until explicit governed promotion; current evidence is required; promotion is durable and preserves epistemic state, claim version and historical proof.12 qualification gates.Qualification complete
Advanced Epistemic / Truth / Decision Qualification — Wall F5 through F6HDurable proof recovery; truthiness, confidence and provenance semantics; canonical logical graph state; SHA-256 state hashing; physical-layout and recovery reproducibility; same-pin epistemic binding; pinned provenance reconstruction; and durable Decision Records.Qualification has progressed through F6H-D; the durable Decision Record line closed with 12/12 qualification gates passing.Qualification complete through F6H-D
Chandra governed CU workloadReal Chandra append, lookup, spoke-tail, chain traversal, chain verification, reopen and contention paths executed against AletheiaGraph rather than synthetic CRUD alone.Each canonical sweep executes 1,200 governed CU commits: 240 commits at each of 1, 2, 4, 8 and 16 writers. A sweep exercises approximately 21,600 persistent secondary-index insertions and 8,400 removals, with exact authoritative-chain verification and persistent-backend correctness checked at every concurrency point.Baseline qualified / hardening
Durable commit-path optimizationOptimize and requalify the Aletheia durable commit/publication path under governed Chandra workloads without weakening ordering, durability, recovery or attribution semantics.1T qualified. Delta-aware secondary-index maintenance skips physical mutation when an update preserves the same canonical indexed key. The canonical 1T semantic closure completed 449/449 checks across memory, skip-list and B+ tree backends, update transitions, canonical-equivalent keys, delete, reopen and authoritative verification. In the governed CU benchmark, P16 secondary-index time fell from 845.8 μs to 516.7 μs, total Apply fell 24.8%, total commit fell 13.9%, and throughput increased 13.0% versus the preceding 1S baseline, without weakening durable ordering or recovery semantics.Qualification complete — 1T
Cross-stack memory / GC / lifecycleSustained allocation behavior, retained heap, repeated full GC, close/reopen cycles, resource release and long-running workload stability across query, reasoning and governed writes.Qualified. The integrated lifecycle gate completed 160/160 checks over 12 repeated mutation/query/reasoning/reopen cycles. Authoritative reads, maintained indexes, native query and Prolog remained in agreement before and after reopen; read-only Prolog policy continued to deny mutation; alternating snapshot/non-snapshot closes passed; and post-GC live heap remained essentially flat at approximately 122 MiB.Qualification complete
Sustained mixed-workload qualificationRepeated committed mutation, indexed native query, direct index verification, Prolog reasoning, policy enforcement, close/reopen activity and full-GC lifecycle checks in one sustained workload rather than isolated component tests.Qualified. The 16-cycle / 512-background-row mixed workload completed 245/245 checks. Every pre- and post-reopen native equality query used a scan-verified :INDEX path, direct index lookup matched authoritative state, Prolog agreed with the same truth set, secondary indexes verified clean, denied writes remained denied, alternating reopen modes passed and retained heap stayed bounded at approximately 121–122.5 MiB.Qualification complete
Native scalar-query planner integrationEnsure a one-predicate native :QUERY reaches the qualified scalar planner instead of being misrouted through composite-only planning, while preserving authoritative scan verification and fallback behavior.Qualification exposed a production integration defect: semantically correct native queries were silently falling back to :SCAN with :NO-ELIGIBLE-COMPOSITE-INDEX even though a valid single-field index existed. Dispatch was corrected so one predicate uses the qualified scalar planner and two-or-more predicates retain the conjunctive/composite planner. Focused qualification passed 5/5, including indexed execution after reopen and safe scan fallback for an unindexed predicate.Qualification complete
Production Prolog / reasoning integrationProduction integration of the qualified reasoning substrate with graph-backed predicates, recursion/cycle controls, inference budgets, deadlines, provenance and reopen-safe behavior.Reasoning safety, durability and cross-stack lifecycle behavior are already qualified. The next v1 frontier is production integration: turning the qualified reasoning substrate into the production execution path without losing bounded execution, provenance, policy enforcement or reopen determinism.Current v1 frontier
Decision Record → Chandra evidence bindingBind Aletheia durable Decision Records into Chandra CU evidence so evaluated state, provenance and disposition become part of the governed evidentiary path without making Aletheia the owner of CU semantics.Aletheia Decision Records are qualified; Chandra evidence binding remains an integration step.Required before freeze
Form-defined composite index optimizationComposite indexes created from user-designed Form Designer fields must support left-prefix reuse, correct key-order semantics, composite-prefix statistics, redundant-index detection, reusable native query IR templates, authoritative-scan-equivalent fallback, reopen/rebuild integrity and characterized write/storage amplification.Composite index and planner substrate exists; this gate qualifies the application-facing optimization path without making automatic index recommendation/tuning a v1 requirement.Required before freeze
SPARQL compatibility subsetRestricted SPARQL lowering into the native query IR.Implementation remains a compatibility-layer task rather than a prerequisite for the native planner.Planned for v1
End-to-end Chandra / Aletheia qualificationIntegrated governed-write, query, reasoning, provenance, Decision Record, reopen/recovery and authority-boundary qualification across the combined stack.Individual substrate and workload gates are passing; the final integrated qualification run remains.Required before freeze
Networked query / Marshaller boundaryFederated query routing, pushdown, partial-result handling, source attribution and CRC-preserving combination across independently governed graph instances.Architecture is defined sufficiently to preserve authority boundaries, but distributed/surface-aware execution is not a v1 freeze requirement.Post-v1 / later hardening
CRC federation reference harnessRun semantically equivalent distributed graph workloads against independent reference engines (RDF4J FedX, Apache Jena ARQ/Fuseki and GraphDB) and later AletheiaGraph/Marshaller; capture source selection, remote requests, boundary exchanges, transferred data, graph work, traversal depth and planned-versus-actual exposure.This is the external validation path for CRC reachability/federation claims. It is intentionally independent of AletheiaGraph's own implementation so the standard is not validated only against itself.Planned / post-v1 qualification
Full v1 regression / release freezeCross-wall regression, Skip List/B+ tree backend parity, governed-workload regression, memory/GC and sustained-performance checks, reasoning integration, reopen/recovery, source provenance and warning baseline.Final consolidated freeze run has not yet been executed.Final gate

Wall A in detail

Wall A is broader than a transaction smoke test. It qualifies the authoritative write and recovery path under contention, interruption and restart. The canonical v6 runner executes 20 subordinate qualification runners and one additional current-semantics multi-key overlap gate, for 21 qualification stages. Four superseded historical runners are explicitly excluded from that count.

Wall A subdomainWhat it covers
Atomicity & concurrencyCommit behavior, shared-node contention, lost-update prevention, multi-thread correctness and multi-key overlap.
Durability & recoveryWAL persistence, committed-state survival, deterministic killpoints, reopen/recovery and sidecar/checkpoint behavior.
Conflict & visibility semanticsOCC conflicts, authoritative last-writer behavior, read-only validation, dirty-read prevention and recovery conflict semantics.
Publication & checkpoint correctnessOrdered publication, successor wakeup, waiter behavior, checkpoint durability and highest-transaction markers.
Write reservation / intentReservation acquisition, synchronous and asynchronous write-intent paths, cancellation and cleanup.
Apply-hook & provenance integrityAfter-apply behavior, one-time hook consumption and qualification against the intended current source tree.
Counting note: 21 is the number of canonical Wall A qualification stages run by the v6 gate, not the number of internal assertions inside those runners. The concurrency workload is 32 threads × 500 increments = 16,000 transactional increments, with the exact final count and no lost updates required. The multi-key overlap workload is 200 transactions producing 400 node updates, with final expected counts A=100, B=200 and C=100.

Wall B in detail

Wall B qualifies the entire index substrate as a derived acceleration layer over the authoritative graph. The objective is not merely to prove that an index can return values; it is to characterize correctness, ordering, pagination, lifecycle, recovery, memory behavior, planner statistics, traversal support and backend behavior under realistic scale.

Wall B domain

Correctness & access paths

Exact/range lookup, ASC/DESC traversal, single-field and composite structures, LIMIT, keyset pagination and agreement with authoritative scans.

Wall B domain

Integrity & recovery

Verify/repair, reopen integrity, transactional maintenance, recovery behavior, regeneration from authoritative graph state and tombstone semantics.

Wall B domain

Planner substrate & traversal

Cardinality/distinct/selectivity statistics, traversal-oriented and reverse indexes, and the physical information consumed later by Wall C.

Wall B domain

Scale, lifecycle & backends

Scaling, build/rebuild cost, write amplification, memory/GC lifecycle, cold/warm behavior and parity across Skip List and B+ tree backends.

Wall B capability coverage

Wall B capabilityWhat was qualified
Ordered index correctnessExact lookup, range lookup, ascending/descending traversal and ordered retrieval.
Index vs scan agreementIndexed results must match authoritative full-scan semantics rather than becoming a second source of truth.
Single-field indexesCore property/value index construction, maintenance and lookup behavior.
Composite indexesMulti-field index structures and compound lookup semantics.
Directional composite indexesComposite ordering capable of preserving ASC/DESC planning behavior across multiple properties.
LIMIT supportBounded ordered-index execution without unnecessarily consuming the full qualifying result set.
Keyset paginationStable continuation through ordered index space without offset scanning.
Range semanticsInclusive/exclusive boundaries and bounded range traversal.
Statistics generationEntry counts, distinct counts and selectivity/cardinality summaries.
Planner statistics substrateIndex statistics designed for direct consumption by Wall C cost estimation and plan selection.
Verify / consistency checkingDetection of divergence between derived index state and the authoritative graph.
Repair / regenerationReconstruction of indexes from authoritative graph state rather than trusting corrupted derived state.
Reopen integrityPersisted indexes remain semantically correct after close/reopen.
Recovery behaviorIndex state remains consistent through transaction and recovery lifecycle events.
Transactional maintenanceIndex changes remain synchronized with committed graph mutations and do not expose partial authoritative state.
Soft-delete / tombstone behaviorDeleted graph objects cease appearing through indexes while lifecycle/tombstone metadata remains correct.
GC / lifecycle behaviorIndex caches, waiter state and reverse-index references do not indefinitely retain closed graphs.
Cold vs warm behaviorReverse/index paths behave correctly after cold reopen as well as with warmed caches.
Traversal-oriented indexesAdjacency/reverse lookup support required for efficient graph traversal.
Wrong-graph isolationIndex lookup resolves against the graph explicitly passed to the operation rather than accidentally consulting ambient *graph*.
Backend parityBoth Skip List and B+ tree persistence/index paths are exercised.
ScalingLarger cardinalities/selectivities, build/rebuild cost, write amplification, memory cost and query latency.
Core Wall B progression: v3B.1a established the 370-case ordered-index baseline; later stages qualified LIMIT, keyset pagination, verify, repair, reopen integrity, statistics, scaling, traversal, cold/warm reverse lookup, GC/index lifecycle, composite indexes, soft-delete/tombstone behavior, and Skip List/B+ tree backend paths.
Indexes remain derived acceleration structures. Verify and repair qualification deliberately compare them against the authoritative graph so performance structures cannot silently redefine truth.

Qualification found and removed real defects

Wall B was also an optimization and lifecycle hardening campaign. Qualification exposed defects that would not have been visible in a feature checklist, including an O(n²)-style secondary-index apply path that was reduced to a fast exact-lookup path, graph retention through write-reservation waiter state, and graph retention through reverse-index cache state. The important result is not only broader feature coverage, but improved correctness, recoverability, memory safety and scaling behavior of the index substrate.

Wall C in detail

Wall C qualifies the full native query decision path: from graph-native logical representation, through statistics and physical access-path choice, into real execution. The goal is not merely to make queries faster; it is to prove that optimized execution preserves the semantics of the authoritative graph.

Wall C domain

Query semantics & algebra

Graph-native logical representation and operators: filtering, projection, ordering, joins, aggregation, pagination and bounded execution.

Wall C domain

Statistics & cost estimation

Cardinality, distinct counts, selectivity, range estimates, index-cost estimates and behavior when statistics are stale, absent or incomplete.

Wall C domain

Plan selection & access paths

Choosing among scans, equality/range indexes, ordered access and directional composite indexes without changing query meaning.

Wall C domain

Execution correctness & advanced operators

Real executor integration, authoritative-result agreement, joins, aggregation, EXPLAIN and safe fallback behavior.

Wall C capability coverage

Wall C capabilityWhat was qualified
Native query IR / algebraInternal graph-native representation for filtering, projection, ordering, joins, aggregation, pagination and bounded execution.
Explicit access pathsAuthoritative graph scan, index equality, range, ordered-index paths and explicit-path execution.
Cost estimationCardinality/selectivity estimates, range estimates, index-cost estimates and planner confidence in available statistics.
Statistics substrateDistinct counts, cardinality and selectivity summaries, including stale-, absent- and incomplete-statistics behavior.
Plan selectionChoosing among scan and indexed alternatives while preserving the logical query's semantics.
Authoritative-result equivalencePlanned and indexed results must agree with authoritative scan execution rather than becoming a second source of truth.
Missing / unbound fieldsQueries remain correct when a field is absent; the planner cannot assume indexed values represent every possible match.
Ordered executionASC/DESC execution, ordered indexes, LIMIT interaction and directional composite-index behavior.
PaginationLIMIT and keyset-style paging behavior, including use of qualified ordered access paths rather than full rescans where applicable.
ProjectionReturning selected fields/results without altering filtering, ordering or underlying query semantics.
Composite indexesMulti-field access paths and directional composite planning for queries whose useful physical ordering spans more than one property.
JoinsJoin execution plus cost-based join planning across the native query substrate.
AggregationAggregate execution and aggregate planning as native operators rather than application-side post-processing.
EXPLAIN / plan introspectionSelected access paths and physical plans are observable so planner decisions can be qualified and debugged.
Execution integrationSelected plans are run through the real executor; qualification is not limited to planner objects or estimated costs.
Fallback behaviorCorrect execution when statistics are stale/absent, an index is unusable, or no optimized path is appropriate.
Core Wall C progression: 3A/3B established query IR and explicit access paths; 3C added estimator/statistics; later qualification covered planned-result correctness, native query algebra, directional composite indexes, costed directional planning, joins and join planning, aggregation and aggregate planning, with P10/P11/P12 closing remaining integration issues. Exact internal assertion counts are not yet reconstructed, so the public site reports named qualification stages rather than an inflated grand total.
The optimizer is not allowed to become authoritative merely because an index exists. Indexed and planned execution must agree with authoritative graph semantics, including missing-field and stale-statistics cases.

Wall D in detail

Wall D qualifies the reasoning engine itself: whether Prolog and graph-backed reasoning execute correctly, recursively, concurrently and within explicit resource bounds. This is the in-memory/runtime safety wall for reasoning, distinct from Wall E's persistence and recovery guarantees.

Wall D domain

Semantics & graph integration

Core Prolog behavior, recursion, structured terms/functors, compiled rule behavior and graph-backed predicates over AletheiaGraph state.

Wall D domain

Mutation & concurrency

Reasoning-side mutation/retract behavior, concurrent reasoning operations and isolation of bindings and transient query state.

Wall D domain

Resource bounds

Explicit inference budgets, cyclic-recursion protection, spin-loop interruption and wall-clock deadlines.

Wall D domain

Failure containment

Runaway or resource-bound reasoning fails gracefully without crashing or contaminating the graph engine.

Wall D capability coverage

Wall D capabilityWhat was qualified
Core Prolog semanticsBasic rule/query behavior and recursive reasoning.
Structured terms / functorsFunctors, structured term handling and compiled rule behavior.
Graph-backed predicatesReasoning directly over graph vertices and edges rather than a separate logical truth store.
Mutation semanticsProlog-triggered retract and mutation paths.
Concurrent reasoningMultiple reasoning operations execute without cross-query state contamination.
Query-state isolationRepeated and recursive queries do not leak bindings or transient execution state.
Inference budgetsRecursive reasoning is bounded by an explicit inference limit.
Cycle protectionCyclic recursion aborts gracefully instead of running forever.
Wall-clock timeoutRunaway reasoning can be terminated by deadline.
Spatial / graph reasoning integrationProlog integrates with graph and spatial predicates.
Failure containmentResource-bound failures do not crash the engine or corrupt surrounding reasoning state.
Deterministic control behaviorBudget and timeout enforcement behave consistently enough to be tested and relied upon.
Explicit Wall D hardening checks: inference-budget-bounds-recursion, cyclic-recursion-fails-gracefully, inference-budget-breaks-spin-loop, and query-timeout-breaks-spin-loop. Larger qualification groups covered D1–D3 core semantics/algorithms, D2 functors, D4 mutation, D5 concurrency, D6–D9 stress/isolation/resource bounds and D10 spatial-Prolog integration.
Wall D proves that AletheiaGraph's reasoning engine is not only expressive, but bounded, isolated, concurrency-safe and resistant to runaway recursion.

Wall E in detail

Wall E moves the reasoning system across the persistence boundary. It asks whether facts visible to reasoning, reasoning-side retractions, denied writes, deterministic answers and resource controls remain correct after commit, close/reopen and recovery.

Wall E capabilityWhat was qualified
Committed fact visibilityProlog sees facts only after they are durably committed.
Reopen consistencyThe same committed facts remain reasoning-visible after close/reopen.
Retract durabilityA committed Prolog retract remains effective after reopen.
Denied-write safetyRejected reasoning-side writes have no side effect.
Denied-write durabilityThe no-effect guarantee for rejected writes still holds after reopen.
Deterministic result recoveryOrdinary Prolog query results remain the same across reopen.
Inference-budget persistenceInference limits still work after the graph is reopened.
Timeout persistenceDeadline enforcement still works after reopen.
Eight Wall E checks: E1 committed facts visible; E2 committed facts survive reopen; E3 committed retract visible; E4 retract survives reopen; E5 denied write has no effect; E6 denied-write no-effect survives reopen; E7 deterministic query result after reopen; E8 inference budget and timeout still work after reopen.
Wall D asks whether reasoning works safely. Wall E asks whether reasoning remains correct after persistence and recovery boundaries. A reasoning engine can pass D and still fail E; AletheiaGraph qualifies both layers separately.

What skepticism does the test answer?

A raw count is less useful than the engineering question behind it. These are the failure modes the qualification program is intended to answer directly.

Enterprise questionEvidence exercised in the testbed
Does it survive crashes and restarts?Wall A v6: 21 canonical stages spanning WAL/recovery, deterministic killpoints, checkpoint durability, OCC conflict semantics, dirty-read prevention, publication ordering and close/reopen verification.
Can indexes be trusted instead of scans?370-case ordered-index baseline plus verify/repair, reopen integrity and backend qualification.
Does it scale beyond toy graphs?100k-record B+ tree qualification, scaling-matrix work and memory/GC testing.
Can the optimizer change the answer?Wall C qualification spans native query algebra, statistics and cost estimation, plan selection, joins, ordering, aggregation and execution-level agreement with authoritative graph semantics—including missing/unbound fields and stale-statistics fallback.
Can agents reason recursively without taking down the engine?Inference budgets, cyclic-recursion guards, infinite-loop protection and wall-clock deadlines.
Does reasoning survive database lifecycle events?Wall E durable reasoning and reopen qualification.
Can the engine distinguish facts, inferences and proposed actions?Wall F1 graph-native epistemic typing and persistence qualification.
Can a conclusion retain proof of exactly which committed evidence supported it?Wall F2 qualifies durable proof relationships from conclusions to version-specific vertex or edge evidence using source identity plus commit-epoch, with graph isolation and reopen durability.
Are multiple physical backends exercised?Both Skip List and B+ tree index backends are included in qualification.
What remains unfinished?Wall F has progressed through F6H-D. Remaining v1 work is systems hardening and integration: durable commit-path optimization under governed CU workloads, integrated memory/GC and sustained mixed-workload qualification, Form Designer composite-index optimization, production reasoning integration, Decision Record → Chandra evidence binding, the restricted SPARQL subset, end-to-end Chandra/Aletheia qualification and the final release freeze.

Wall F in detail

Wall F covers proof, provenance, truth semantics and durable decision evidence. The work did not stop at F4: F5 closed durable proof recovery, and F6 was deliberately decomposed into a longer qualification sequence so confidence, disposition, truthiness, ingestion confidence, divergence and finally pinned decision-state evidence could be qualified separately rather than hidden behind one “truth score.”

Current Wall F result: F1 through F5 are qualification complete. F6 progressed through the semantic sequence and then through F6H-D, where the durable Decision Record line closed with 12/12 gates passing. The F6H stages also qualified canonical logical state, SHA-256 state hashing, layout invariance, recovery reproducibility, same-pin epistemic binding and pinned provenance reconstruction.
Wall F now spans the complete evidentiary path: classify the epistemic object, preserve its proof, determine evidence validity, govern authority transition, survive recovery, evaluate confidence/truth/disposition, pin the evaluated graph state and persist a reconstructible Decision Record.
Wall F progression:
F1 — Epistemic classification
F2 — Proof topology & evidence provenance
F3 — Evidence validity & invalidation
F4 — Authority transition / promotion
F5 — Durable proof recovery
F6A — Reference determination
F6B — Actionability reference
F6C — Confidence reference
F6CG1–CG9 — Confidence / provenance extensions
F6D — Disposition
F6E — Truthiness
F6F — Ingestion confidence
F6G — Client divergence
F6H — Epistemically pinned decision-state and durable Decision Records
StagePublic labelWhat it provesStatus
F1Epistemic ClassificationDurable distinction among :stored, :inferred and :proposed; graph-local system schema; commit provenance; reopen durability.Qualification complete
F2Proof Topology & Evidence ProvenanceVertex evidence, edge evidence, multi-support conclusions, reusable support records, source ID + commit-epoch, graph isolation, reopen durability and authority preservation.Qualification complete — 10 gates
F3Evidence Validity & InvalidationClassify cited evidence as :current, :stale, :retracted or :missing; preserve historical proof; keep validity support-local, read-only and graph-local; preserve classifications across reopen.Qualification complete — 10 gates
F4Authority Transition / PromotionKeep inferred/proposed claims non-authoritative until explicit promotion; require current evidence; create durable authority-transition state; preserve epistemic kind, exact claim version and historical proof.Qualification complete — 12 gates
F5Durable Proof RecoveryProof DAGs, support state, invalidation state and authority state remain correct through close/reopen, recovery and deterministic failure boundaries.Qualification complete
F6AReference DeterminationEstablish the reference determination semantics used by the later truth/confidence decision path.Qualification complete
F6BActionability ReferenceEstablish actionability semantics independently from later confidence and disposition decisions.Qualification complete
F6CConfidence ReferenceEstablish confidence semantics on the qualified evidence/provenance substrate.Qualification complete
F6CG1–CG9Confidence / Provenance ExtensionsExtend and qualify confidence behavior across provenance, temporal, evidence-strength, staleness, conflict, domain/density and measured-confidence concerns.Qualification complete
F6DDispositionKeep policy/disposition evidence distinct from execution authority and preserve the basis used to classify an outcome.Qualification complete
F6ETruthinessEvaluate truthiness on the qualified proof/confidence substrate without creating a parallel authoritative truth store.Qualification complete
F6FIngestion ConfidencePreserve ingestion confidence and its contribution to effective confidence as explicit evidence.Qualification complete
F6GClient DivergencePreserve client-side divergence/evaluation differences as explicit decision evidence rather than silently collapsing them.Qualification complete
F6HPinned Decision-State & Durable Decision RecordsBind evaluation to canonical logical graph state, reconstruct provenance at the same pin and persist durable, replay-resistant Decision Records.Qualification complete through F6H-D

F6H qualified sequence

StageQualified behaviorStatus
F6H-A0Snapshot enumeration over the logical graph state.12/12 PASS
F6H-A0H1Generic / type-0 edge coverage.4/4 PASS
F6H-A0H2Hard-purge exclusion from canonical logical state.12/12 PASS
F6H-A1Canonical logical state representation.12/12 PASS
F6H-A2SHA-256 logical graph-state hash.12/12 PASS
F6H-A3Physical/index-layout invariance of the logical state hash.12/12 PASS
F6H-A4Recovery reproducibility of the canonical logical state.12/12 PASS
F6H-BSame-pin epistemic binding.12/12 PASS
F6H-CPinned provenance reconstruction.12/12 PASS
F6H-DDurable Decision Record preserving evaluated state, evidence, provenance, confidence and disposition across commit/reopen boundaries.12/12 PASS

F3 qualified behaviors

GateQualified behavior
F3.1Fresh vertex evidence classifies as :current.
F3.2Fresh edge evidence classifies as :current.
F3.3–F3.4Updating cited vertex or edge evidence makes the historical support :stale without rewriting the citation.
F3.5Deleted cited evidence classifies as :retracted.
F3.6Unresolvable cited evidence classifies as :missing.
F3.7Validity remains local to the support being evaluated.
F3.8Historical evidence citations remain immutable.
F3.9Validity classifications survive close/reopen.
F3.10Classification is read-only and graph-local; wrong-graph use is rejected.

F4 qualified behaviors

Qualification focusQualified behavior
Authority by originStored state is authoritative by origin; inferred and proposed state is not.
Eligibility vs promotionCurrent supporting evidence can establish promotion eligibility without implicitly promoting the conclusion.
Explicit promotionAuthority transition requires an explicit governed promotion and creates durable transition state.
Epistemic preservationPromotion preserves the claim's epistemic kind, proposition, exact claim version and historical proof rather than rewriting its history.
Invalid evidence rejectionStale, retracted or missing evidence blocks promotion.
Boundary rejectionUnsupported, repeated and wrong-graph promotion attempts are rejected.
Reopen durabilityAuthority-transition state remains correct after close/reopen.
Proof preservationPromotion does not mutate the underlying proof relationship.

Epistemic state

Stored, inferred and proposed objects remain distinct and durable.

Evidence & proof topology

Conclusions retain explicit, reusable and version-specific support relationships.

Validity & authority

Evidence can become stale without erasing history, and conclusions cannot silently promote themselves into authoritative state.

Truth scoring & contradiction

Competing evidence is interpreted only after proof and provenance semantics are established.

Formal-wall boundary: Walls A through E are closed, and Wall F has progressed through the F6H durable Decision Record line. Remaining v1 work is now concentrated less in adding another formal wall and more in production integration. Governed Chandra CU workload qualification, durable commit-path optimization, cross-stack memory/GC/lifecycle qualification and the sustained mixed-workload gate are closed at the current baseline. The next frontier is production Prolog/reasoning integration, followed by Decision Record → Chandra evidence binding, the restricted SPARQL compatibility subset, end-to-end qualification and the final release freeze.

Operational-path hardening

Agent-first qualification is not only about database features in isolation. Autonomous clients repeatedly stress authoritative writes, repeated reasoning, concurrent mutation, recovery and long-lived memory behavior. AletheiaGraph therefore hardens the operational paths those clients actually drive.

Authoritative write path

Governed mutations & CU append

Ordinary graph create/update/delete behavior remains transactional, while Chandra CUs follow append-only governed publication semantics with spoke-tail advancement, lineage verification, publication ordering and durable recovery.

Throughput path

Chandra CU workload optimization

Canonical sweeps exercise 1,200 governed commits across five contention levels. Remove-path qualification reduced persistent skip-list traversal from three searches per removal to one; insertion-path hardening now focuses on measured persistent pointer-mutation cost while preserving ordered visibility, durability, lineage, attribution and recovery semantics.

Lifecycle

Memory / GC behavior — qualified

The integrated 12-cycle lifecycle gate held authoritative/index/native-query/Prolog agreement across repeated reopen while full-GC live heap remained essentially flat around 122 MiB.

Read path

Lean indexed execution

Lean B+ tree descent, partial decode, in-place edits, equality/range/ordered LIMIT and planner reuse of qualified indexes.

Form-defined access path

Composite index optimization

User-designed forms may define composite indexes and reusable query shapes. v1 qualification covers prefix reuse, key ordering, prefix statistics, overlap detection, safe fallback and reopen/rebuild behavior without allowing form metadata or derived indexes to become authoritative.

Reasoning safety

Repeated autonomous reasoning

Recursion, graph-backed predicates, concurrency, inference budgets, cycle termination, infinite-loop breaking and timeout enforcement.

Reasoning lifecycle

Durable re-query

Committed facts, retract durability, denied writes, deterministic results and resource controls after graph reopen.

We test not only database features, but the exact operational paths an autonomous agent will stress: rapid authoritative writes, repeated reasoning, failure and recovery, concurrent mutation and provenance-preserving decisions.
Why integrated qualification matters: the sustained-workload campaign found a defect that ordinary correctness testing would have missed. Native equality queries returned the right rows, but the native wrapper silently routed a single predicate through composite-only planning and fell back to a full scan. The direct index was correct; the integration path was wrong. The fix was separately qualified, then the full mixed-workload gate was rerun to 245/245. Correct answers alone were not enough.

What we test

FamilyWhat the qualification is trying to proveStatus
Durability / killpointsInterrupted execution cannot silently publish ambiguous authoritative state; reopen recovers to a valid boundary.Foundation qualification complete / regression
Transaction publicationGraph and derived structures cross publication boundaries consistently; partial work does not masquerade as committed work.Hardening
Concurrent writersHandoff, cancellation, stale state and contention have defined behavior rather than accidental ordering.Qualified / regression
Index lookup / rangeIndexed paths return the same graph objects and semantics as authoritative scans.Qualification complete
Wrong-graph isolationIndex operations resolve against the graph argument supplied to the operation rather than leaking through ambient graph state.Qualification complete
Index reopen / rebuildPersistent indexes survive reopen or can be reconstructed and checked against authoritative graph state.Qualification complete
Statistics edge casesEmpty indexes, zero distinct values, missing or stale statistics do not produce arithmetic failures or semantically invalid plans.Qualification complete
Planner equivalenceEquality, range, ordered, join and aggregate plans agree with authoritative semantics before production selection is trusted.Qualification complete / hardening
Constant-work scalingOperations intended to be constant-work remain so as graph cardinality grows; write amplification, allocation and GC costs are measured.Qualified baseline / continuing regression
Reasoning durabilityDerived conclusions remain deterministic, attributable and recoverable; dependencies and invalidation remain explicit across reopen.Qualification complete
Governed Chandra CU workloadAppend, retrieve, traverse, verify, reopen and same-spoke contention exercise the real governed application path through AletheiaGraph.Baseline qualified / hardening
Sustained integrated lifecycleMixed writes, queries, reasoning, memory/GC, checkpoint/reopen and long-run resource stability are exercised together.Current v1 hardening

A representative index invariant

index-lookup / index-range
≡
authoritative graph scan

Performance structures are tested as derived access paths. The engine should be able to prove that an indexed answer agrees with the underlying graph, including boundary cases where fields are missing or statistics are absent.

Failure is part of the workload

Database correctness is often exposed by what happens between ordinary operations. AletheiaGraph's durability program uses deterministic interruption points and reopen verification so failures can be reproduced at specific publication stages rather than approximated by occasional random process termination.

Begin mutation

Establish the transaction and intended graph/index changes.

Interrupt at a known boundary

Terminate at a deterministic internal stage.

Reopen

Recover the persistent graph rather than trusting in-memory state.

Verify invariants

Compare graph, indexes and publication state against the defined valid outcomes.

Why AI changed the engineering economics

Two years earlier, the rational choice for this project likely would have been PostgreSQL beneath a graph/reasoning layer. In 2026, AI-assisted engineering makes deeper vertical integration economically realistic—not because a model can confer reliability, but because it can materially increase the rate of review, adversarial hypothesis generation, test construction, comparative analysis and regression expansion available to a small engineering effort.

More adversarial hypotheses

What if cardinality is zero? What if statistics are stale? What if the wrong graph is ambient? More failure hypotheses can be generated and checked before field deployment.

Faster conversion into evidence

A discovered defect can become a deterministic reproducer, qualification case and permanent regression gate rather than remaining tribal knowledge.

Broader comparative scrutiny

Design choices can be repeatedly challenged against established storage-engine techniques, prior failure patterns and competing implementations.

AI does not substitute for field exposure. It changes how quickly engineered experience can be accumulated, preserved and replayed.

Evidentiary maturity

AletheiaGraph does not claim that qualification can predict every open-world interaction. Production diversity remains a distinct and valuable source of evidence. But calling a system simply “mature” or “unproven” collapses several different questions into one chronological label.

Specification maturity

Are the properties the engine must preserve explicit enough to test and falsify?

Qualification maturity

Have those properties survived meaningful recovery, concurrency, corruption, scale and semantic-boundary challenges?

Exposure maturity

Has deployment diversity revealed interactions and failure classes that the qualification program did not anticipate?

Failure-memory maturity

Does each discovered failure become durable, replayable engineering knowledge?

Operational diversity

How many workload, hardware, timing, lifecycle and integration combinations have exercised the system?

Adaptation velocity

How quickly can a new failure class be converted into an invariant, reproducer and permanent gate?

For a machine system, maturity can be cumulative rather than merely chronological.

AletheiaGraph's maturity statement is therefore intentionally evidence-based: what behavior has been characterized, how was it challenged, what remains outside the evidence boundary, and can the accumulated lessons be replayed? That is what this page will continue to publish as v1 approaches freeze.