PaperlayerView exact revision on arXiv

Source-linked AI summary

Epistemic Typing as a PostgreSQL Table Access Method: Adversarial Conflict Resolution Under Confidence Forgery and Sybil Coordination

Venkata Maguluri, Swapna Kommuri, Kritica Sinha, Ashish Sood

arXiv:2609.36795v1cs.DBcs.CR

TL;DR

Conventional stores treat measured, inferred, and derived writes alike, leaving confidence ranking vulnerable to high-confidence hallucinations and coordinated Sybils. KNDB adds engine-assigned epistemic kinds and write-time conflict resolution in PostgreSQL’s table access method. It beats confidence-only arbitration on two datasets and is competitive below, then dominant at or above, the truth-discovery density-saturation cell.

  • Problem

    Conventional stores treat observed, inferred, and derived writes identically, while high-confidence hallucinations and coordinated Sybils can challenge confidence-based arbitration.

    Source: §1 · p. 1

  • Method

    KNDB assigns epistemic kinds in the engine and resolves per-slot conflicts at write time using a PostgreSQL table access method.

    Source: §9 · p. 12

  • Results

    KNDB beats a confidence-only arbitrator by 63 and 92.7 points across two workloads, and is competitive below but dominant at or above the truth-discovery density-saturation cell.

    Source: §9 · p. 12

  • Takeaways & Limitations

    The density-saturation cell is predictable from per-slot honest support, with the threshold prediction validated within ±20% on Book-Author and ±30% on Zheng.

    Source: §1 · p. 2, §9 · p. 12

    KNDB scores 0.000 at c = 1 on three temporal knowledge-editing benchmarks because their last-writer-wins target conflicts with its first-committer-wins tiebreak.

    Source: §6.14 · p. 11

Abstract

from arXiv · show

We describe KNDB, a PostgreSQL 18 table access method (TAM) that types every row with an engine-assigned epistemic kind (MEASURED, INFERRED, or DERIVED) and resolves per-slot conflicts inside every write-time heapam callback. Rows land as ordinary heap tuples; seven of the 44 TAM callbacks are overridden (tuple_insert, multi_insert, tuple_update, tuple_delete, tuple_insert_speculative, tuple_complete_speculative, relation_toast_am), the other 37 delegate to heap; we provide a completeness argument over the interface as a paper artefact. This paper reports the engineering behind that decision and the adversarial evaluation that motivated it. On a confidence-forgery workload where an attacker asserts INFERRED writes with confidence in [0.95,1.0] against honest MEASURED writes with confidence in [0.5,0.9], KNDB beats a confidence-only baseline by 63 percentage points on the Book-Author fusion dataset and 92.7 points on the Zheng crowdsourcing dataset. Both wins are proven load-bearing on the kind axis by a source-rebuild disable-and-test in which the lattice is neutralised and the win vanishes. Against four truth-discovery baselines (TruthFinder, CRH, CATD, ACCU) reimplemented from the original equations and validated to within 0.3 percentage points of the published numbers, KNDB is competitive below a per-dataset density-saturation cell and dominant at or above it. We formalise the cell as k* ~ rho_alg * h_top, where h_top is per-slot top honest surface-form support, and validate the prediction within +/-20% on Book-Author and +/-30% on Zheng. Because the kind axis is assigned by the engine from independent metadata and cannot be forged at write time, KNDB's k* is unbounded. The paper is honest about where KNDB loses: CRH and ACCU outperform KNDB below saturation on Zheng, and KNDB scores zero on three temporal knowledge-editing benchmarks whose ground truth is last-writer-wins.

1 INTRODUCTION

KNDB makes epistemic kind part of PostgreSQL’s write-time enforcement, addressing conflicts among observed, inferred, and derived claims. The paper frames its contribution around adversarial robustness, a predictable Sybil-collapse threshold, and explicit scope limitations.

1 INTRODUCTION

  • The motivating risk is that confidence-ranking arbitration can prefer high-confidence hallucinations over directly observed claims, while coordinated Sybils can outvote honest support.Conventional relational stores treat these producer claims alike, often allowing the latest write or a user-space trigger to determine the current value.Source: §1 · p. 1
  • KNDB assigns each row an engine-enforced epistemic kind and resolves slot conflicts inside the table access method rather than relying on triggers.Its precedence order uses kind, specificity, confidence, and xmin; the paper describes seven overridden callbacks and an audit of all 44 TAM callbacks.Source: §1 · pp. 1–2
  • KNDB matches confidence-only arbitration when kind and confidence coincide by construction, and scores zero on temporal benchmarks whose ground truth is last-writer-wins.Below the density-saturation cell on Zheng, CRH and ACCU outperform KNDB.Source: §1 · p. 1
  • The paper enumerates nine user-space write paths blocked by TAM callbacks and three DDL-privileged threats it leaves out of scope.The batch COPY, UPDATE/DELETE, and speculative-insertion paths required dedicated overrides.Source: §1 · pp. 1–2
  • The paper predicts Sybil-collapse from per-slot honest support, validating the threshold within ±20% on Book-Author and ±30% on Zheng.The stated predictions are [5] versus an observed cliff N=5...10 for Book-Author, and [14,20] versus observed cell N=20 for Zheng.Source: §1 · p. 2
  • The paper also reports a 15,000-row-per-transaction ceiling at PostgreSQL’s default lock setting and identifies SET ACCESS METHOD heap as outside the enforcement envelope.Its custom WAL resource manager is described as an annotation channel, not a durability mechanism.Source: §1 · p. 2

2 MOTIVATION AND THREAT MODEL

The threat model centers on adversarial writes that exploit confidence ranking, coordinated identities, or write paths that bypass expected checks. KNDB assigns kind from independent metadata and places enforcement in the TAM callback layer, while excluding table-owner schema changes.

2.1 Confidence forgery

  • An attacker’s high-confidence INFERRED claims can outrank honest MEASURED claims under confidence-only arbitration, but KNDB’s kind order rejects them.The evaluated attack uses confidence uniform in [0.95, 1.0] against honest confidence uniform in [0.5, 0.9].Source: §2.1 · p. 2

2.2 Sybil coordination

  • Sybil agents may coordinate falsified claims and outvote honest sources once their count exceeds the top honest surface-form support.The model bars adversaries from forging existing honest identities or obtaining a MEASURED assignment; kind uses independent source or worker metadata.Source: §2.2 · p. 2

2.3 Write-path bypass

  • The bypass analysis treats TAM callbacks as reachable from user-space write paths and distinguishes nine blocked paths from three DDL-privileged paths out of scope.The figure’s routing claim is tied to PostgreSQL 18 REL_18_STABLE callback sources.Source: §2.3 · p. 2, §2 · p. 2
  • F20 rejects UPDATEs that change the epistemic prefix and refuses DELETE outright, preventing prefix forgery and incumbent removal.Content-only updates remain permitted; the stated threat boundary excludes the table owner.Source: §2.3 · pp. 2–3
  • The TAM callback fires regardless of trigger settings or replication role, unlike trigger-based enforcement.The speculative-insertion overrides are included defensively even though the described SQL path currently lacks a usable arbiter index on epistemic tables.Source: §2.3 · pp. 2–3

3 DESIGN

KNDB stores ordinary heap tuples but adds write-time epistemic ranking and validation. Its design combines a precedence order, candidate-row rules, a per-slot lock, and eviction auditing.

3.1 Precedence lattice

  • The precedence lattice ranks MEASURED above DERIVED above INFERRED, with kind as the primary key and specificity and confidence as later ranking levels.A true tie is resolved using xmin.Source: §3.1 · p. 3, §3.4 · p. 3

3.2 Write-time rules R1–R5

  • R1–R5 validate candidate rows independently of the precedence contest, including source requirements, registered source identifiers, confidence bounds, and optional slot-kind constraints.An INFERRED candidate with confidence 1.0 is refused; a candidate that passes these rules still competes with any live incumbent.Source: §3.2 · p. 3

3.3 Per-slot advisory lock (F6)

  • A per-slot transaction-scope advisory lock serializes writers before the live-incumbent scan, so the scan can observe a peer’s committed row.Without the lock, two concurrent same-slot inserts both committed in 50 of 50 trials; with it, session1 won all 50 trials.Source: §3.3 · p. 3

3.4 First-committer-wins tiebreak (F8)

  • The xmin decision reads the incumbent’s raw xmin and compares it with the current transaction ID after the advisory lock ensures the incumbent has committed.The implementation uses TransactionIdPrecedes to handle xid wraparound.Source: §3.4 · p. 3
  • The xmin tiebreak gives the incumbent the win on a true rank tie, replacing a content-hash rule vulnerable to attacker-controlled grinding.The hash-grinding script recorded 30 of 30 attacker wins under the earlier rule and 0 of 20,000 under xmin; survivor identity is not stable across dump and restore.Source: §3.4 · pp. 3–4

3.5 Reason codes and audit

  • Eviction records the loser and reason, closes the incumbent’s sys_time, and inserts the winner within one PostgreSQL transaction.The paper reports 1–2 torn-state violations per 25 crash trials when an audit was staged through dblink, versus 0 of 25 for the in-transaction path.Source: §3.5 · p. 4

4 IMPLEMENTATION

The implementation overrides selected PostgreSQL 18 table-access callbacks while delegating ordinary tuple storage to heap. The section details the write paths requiring special handling and the operating and threat boundaries of that design.

4.1 Handler and delegation

  • KNDB targets PostgreSQL 18 without a PostgreSQL patch and overrides seven TableAmRoutine callbacks while inheriting the other 37.The inherited callbacks preserve heap behavior for reads, index builds, VACUUM, and analyse; the paper describes an exhaustive callback audit.Source: §4 · p. 4, §4.1 · p. 4

4.2 The tuple_insert wrapper

  • The insert wrapper checks R1–R5, locks the slot, scans for a live incumbent, applies precedence, and either rejects the candidate or inserts and audits the winner.It also emits a custom WAL annotation record, which the paper says is not load-bearing for durability.Source: §4.2 · p. 4

4.3 The multi_insert wrapper (COPY FROM)

  • The multi_insert override applies the single-row enforcement path to every COPY row, trading bulk throughput for correctness and compatibility with restore.The fan-out uses one heap insert and WAL record per row, running at roughly INSERT ... SELECT throughput.Source: §4.3 · p. 4

4.4 The tuple_update and tuple_delete wrappers (F20)

  • The UPDATE wrapper rejects changes to ep_kind, ep_specificity, or ep_confidence but allows content-only changes; the DELETE wrapper refuses deletion outright.The authors leave a more permissive audited-delete policy to a future eviction API.Source: §4.4 · pp. 4–5

4.5 The speculative-insertion wrappers (F21)

  • Speculative insertion runs the normal checks before delegating the speculative tuple write to heap, then defers eviction bookkeeping until the speculative tuple succeeds.On failure, pending eviction is discarded so no audit or incumbent closure is left without a winning row.Source: §4.5 · p. 5

4.6 WAL: annotation, not durability

  • The custom WAL resource manager emits an annotation record, while heap’s WAL record carries the row needed for recovery.Recovery succeeded byte-for-byte with the custom emitter disabled in the reported 110-row test.Source: §4.6 · p. 5

4.7 Bulk-load ceiling

  • At the default max_locks_per_transaction setting, the per-row advisory lock imposes a roughly 15,000-distinct-slot transaction ceiling that also applies to COPY.The reported ceiling scales linearly with the setting; raising the GUC is an operator-controlled alternative to removing the lock.Source: §4.7 · p. 6

4.8 Threat surface, restated; 4.9 Exhaustive TAM audit

  • The table owner can move a relation to plain heap outside the enforcement envelope, while the callback audit classifies all 44 interface entries and discloses DDL bypasses.The audit identifies TRUNCATE and CLUSTER/VACUUM FULL alongside SET ACCESS METHOD heap as DDL-privileged threats.Source: §4.8 · p. 6, §4.9 · p. 6

5 FORMALISM

The formalism predicts when agreement-based truth-discovery systems succumb to Sybil coalitions from per-slot honest support. KNDB’s engine-assigned kind rank instead gives it an unbounded collapse threshold when kind is based on independent metadata.

5 FORMALISM

  • The PG 18 audit covers 44 callbacks, including seven overridden callbacks and three DDL-privileged threats disclosed as out of scope.The remaining 37 callbacks delegate to heap.Source: Table 1 · p. 6

5.1 Model and threat

  • The model defines h_top as the maximum honest support for any gold-matching value at an item, so string-valued tasks can have h_top ≤ h.For Book-Author, distinct surface forms of the same gold answer divide support.Source: §5.1 · p. 6

5.2 Per-algorithm monotonicity

  • The evaluated algorithms’ coalition weight grows with Sybil count, through agreement-based trust, a log-ratio update, a confidence bound, or Bayesian per-source accuracy.The passage describes the corresponding mechanisms for TruthFinder, CRH, CATD, and ACCU,,, and.Source: §5.2 · p. 7

5.3 Threshold theorem

  • For TruthFinder, CRH, CATD, and ACCU, the Sybil-collapse threshold is Θ(h_top), with per-algorithm slope close to one under the stated hyperparameters.The formalism gives k* ~ ρ_alg · h_top; the constants are bounded above by 1 + o(1).Source: §5.3 · p. 7

5.4 KNDB invariance

  • With kind assigned from independent metadata, KNDB’s k* is unbounded because one Tier-A MEASURED write outranks any coalition of INFERRED writes.The metadata examples are source history on Book-Author and a disjoint qualification test on Zheng.Source: §5.4 · p. 7
  • When the kind-primary ordering is disabled, KNDB collapses to pg_conf’s numeric behavior on both evaluation datasets.The source-rebuild disable-and-test patches out the kind branch so ranking falls through to specificity and confidence.Source: §5.4 · p. 7

5.5 Empirical validation

  • The predicted Book-Author collapse bracket [3, 5] falls within ±20% of the observed cliff at N=5...10.TruthFinder, CATD, and ACCU show a sharp cliff there; CRH lags one step.Source: §5.5 · p. 7
  • On Zheng, the predicted collapse bracket [14, 20] is within ±30% of the observed N=20 collapse for CRH, CATD, and ACCU.The observed cell coincides with the density-saturation cell.Source: §5.5 · p. 7

6 EVALUATION

In confidence-forgery tests, KNDB outperforms confidence-only ranking on Book-Author and Zheng. Source-rebuild tests show that neutralising kind rank removes those gains, while the evaluation also documents its setup and workload conditions.

6.1 Experimental setup

  • The measurements used PostgreSQL 18.4 on one Apple M5 Pro, and the source-rebuild test restored the source tree byte-identical after patching kind precedence.Integrity remained intact under the patched build because advisory locking and the xmin tiebreak still operated.Source: §6.1 · p. 7, §6.4 · p. 8

6.2 Confidence-forgery workload (Book-Author)

  • The Book-Author workload injects INFERRED claims with confidence [0.95, 1.0] against Tier-A MEASURED claims with confidence [0.5, 0.9].The dataset contains 895 bookstores, 1,265 books, and gold for 100 ISBNs; adversarial counts were swept over N = {1, 3, 5, 10}.Source: §6.2 · p. 7
  • 63 points on Book-Author and 92.7 points on Zheng: KNDB’s precision advantage over pg_conf disappears when kind rank is disabled.Book-Author precision is 0.630 with kind on and 0.000 with kind off; Zheng precision is 0.927 and 0.000, respectively.Source: §6.2 · p. 7, Table 2 · p. 8, §6.3 · p. 8

6.3 Confidence-forgery workload (Zheng)

  • On Zheng, adversarial INFERRED writes use confidence [0.95, 1.0] and flip the gold binary label; the result reports a 92.7-point delta over pg_conf.Tier A is based on worker accuracy on a disjoint qualification test.Source: §6.3 · p. 8

6.5 UPDATE and DELETE forgery attacks (F20)

The F20 UPDATE and DELETE callbacks block two reproduced forgery paths, and disabling their wiring makes both attacks succeed. The section also tests speculative insertion and compares KNDB with truth-discovery baselines under Sybil coordination.

6.5 UPDATE and DELETE forgery attacks (F20)

  • F20 rejects UPDATEs that alter a row’s epistemic prefix and refuses DELETE outright, while allowing content-only UPDATEs to delegate to heapam.The UPDATE callback compares the new kind, specificity, and confidence with the incumbent.Source: §6.5 · p. 8
  • Disabling the F20 callback wiring makes both UPDATE and DELETE forgery scripts succeed; restoring it makes both reject the attacks.The honest build passes 10/10 end-to-end checks, including the two new forgery checks.Source: §6.5 · p. 8

6.6 Speculative-insertion attack (F21)

  • F21’s C-level probe confirms that speculative-insertion overrides reject R3- and R4-violating rows when wired, but let them land when disabled.The honest build passes 8/8 install checks, including the speculative-insertion test.Source: §6.6 · p. 9

6.7 Truth-discovery baselines

  • At Book-Author N=10, KNDB scores 0.630 precision versus CRH’s 0.230, while all four truth-discovery baselines collapse under Sybil coordination.Under the independent-value attack, KNDB’s 0.630 leads the best TD baseline, CRH at 0.580, by 5 points.Source: §6.7 · p. 9, Table 4 · p. 9
  • Under independent-value attacks, all four TD baselines remain flat across N=1...10; pg_mv alone changes between the scrambled and Sybil constructions.KNDB, pg_conf, pg_lww, and pg_trigger decide from the epistemic prefix and are invariant to how adversarial values are distributed across colluding identities.Source: §6.7 · p. 9

6.8 Zheng Sybil sweep and density saturation

On the Zheng Sybil sweep, KNDB remains flat at density saturation while several truth-discovery methods collapse; below saturation, some methods outperform majority vote. The section also reports integrity-grid results and a systematic disagreement between the LLM and KNDB’s lattice ordering.

6.8 Zheng Sybil sweep and density saturation

  • KNDB stays at 0.927 on the Zheng sweep, while CRH, CATD, and ACCU cliff to 0.000 at N=10 or N=20.ACCU reaches 1.000 at N=5; TruthFinder degrades gradually from 0.905 to 0.482.Source: §6.8 · p. 9
  • At sub-saturation N=1...5, CRH, CATD, and ACCU beat majority vote by 4–26 percentage points, but CATD and ACCU fall 29 points below it at N=10.At N=20, all three collapse to the majority-vote floor.Source: §6.8 · p. 9

6.9 Integrity column

  • KNDB is the only system passing every Stage-2 integrity-grid cell; trigger-based baselines fail at 32 clients on hotter contention settings.Two concurrent writers can both observe no incumbent in an empty SELECT FOR UPDATE result; KNDB’s advisory lock closes this window.Source: §6.9 · p. 10, Table 6 · p. 10

6.10 Real LLM calibration and non-determinism

  • The mock LLM was calibrated to 92.5% correctness and 1120 ms mean latency from 200 real Claude Haiku 4.5 API calls.The earlier mock assumptions of 0.65 correctness and 300 ms latency were both updated.Source: §6.10 · p. 10
  • Across 500 real API calls, the LLM had zero flips across 50 conflicts but was unanimously wrong on two; both errors involved a higher-specificity candidate within the same kind.KNDB selects that candidate, whereas the LLM consistently favors the incumbent’s higher confidence; replay-and-vote does not help.Source: §6.10 · p. 10

6.11 Bypass table

The bypass evaluation covers trigger paths and write-time enforcement, while the surrounding tests expose throughput, batch-size, and workload boundaries. KNDB’s correctness advantages do not extend to last-writer-wins temporal benchmarks.

6.11 Bypass table

  • Table 7 summarizes bypass scenarios and outcomes; its marker bypassed* means ENABLE ALWAYS TRIGGER survives replica mode but not DISABLE TRIGGER ALL.The ON CONFLICT attack’s SQL surface is not currently reachable because epistemic tables cannot create the required arbiter.Source: Table 7 · p. 10, §6.11 · p. 10

6.12 Batch-size ceiling under COPY

  • At PostgreSQL’s default max_locks_per_transaction=64, COPY succeeds through 14,900 rows but first fails at 14,950; raising the setting to 1024 pushes the ceiling to about 215,000.Disabling the F18 multi_insert override makes the failing COPY succeed, linking the ceiling to per-row enforcement locks.Source: §6.12 · p. 10

6.13 Contention control (YCSB, F9 gate)

  • At 𝜃=0.0, KNDB reaches 2057 tps median versus pg_heap’s 7119 tps, a 71.1% throughput drop; pg_heap has no correctness guarantee.The passage attributes the drop to per-insert rule checks and overlap scans.Source: §6.13 · p. 10
  • pg_heap’s correct-goodput is 0 on every Stage-2 cell despite its higher raw throughput, with tens of thousands of duplicate live rows per 20-second window.Correct-goodput is defined as tps × (1 – abort rate) × correctness.Source: §6.13 · pp. 10–11

6.14 Temporal-shaped benchmarks (honest zero)

  • KNDB scores 0.000 on three temporal benchmarks because their last-writer-wins ground truth conflicts with its first-committer-wins tiebreak.Both writes rank equally, leaving the tiebreak to decide the survivor.Source: §6.14 · p. 11

7 RELATED WORK

The related work spans truth discovery, provenance, agent-memory warrant, and temporal databases. KNDB’s stated distinction is write-time survivor selection by an engine-assigned epistemic kind, rather than query-time provenance or temporal-only resolution.

Truth discovery

  • The four evaluated truth-discovery designs are TruthFinder, CRH, CATD, and ACCU; their methods use source trust, joint minimization, confidence bounds, and Bayesian accuracy, respectively.The Zheng et al. VLDB 2017 survey supplies the d_sentiment dataset and reproducibility baseline.Source: §7 · p. 11

Provenance in PostgreSQL

  • ProvSQL propagates provenance through query evaluation, whereas KNDB selects surviving rows at write time; the mechanisms are described as orthogonal and potentially composable.A PW25 demonstration adds update provenance through temporal databases with time-travel and undo.Source: §7 · p. 11

Agent-memory and epistemic warrant

  • Orogat and Mansour [5] question database foundations for long-term agent memory, while Romanchuk and Bondar [7] argue that tool boundaries alone do not confer epistemic warrant.Source: §7 · p. 11

TOKI

  • TOKI [10] resolves contradictions by projecting assertions onto valid-time and system-time axes, without an epistemic-kind axis; its temporal semantics and KNDB’s kind axis are described as compatible.TOKI layers a bitemporal operator algebra over an unmodified engine; KNDB implements write-time kind ranking as a table access method.Source: §7 · p. 11

Serialisable snapshot isolation and the TAM substrate

  • KNDB inherits PostgreSQL SSI predicate locking from heapam and stores ordinary heap tuples, allowing tools including pg_dump, pg_upgrade, logical replication, and pg_basebackup to work unchanged.The paper builds on the PostgreSQL 18 TAM interface.Source: §7 · p. 11

8 LIMITATIONS

KNDB’s advantages are bounded by workload, adversary, and PostgreSQL implementation choices. The paper reports losses below saturation and on last-writer-wins tasks, alongside operational and inherited-engine constraints.

Truth discovery wins sub-saturation

  • Below saturation on Zheng d_sentiment, CRH matches or beats KNDB and ACCU reaches 1.000 versus KNDB’s 0.927 at N=5; KNDB’s distinction is an unbounded k* at and above saturation.The paper does not claim KNDB beats every truth-discovery baseline on every cell.Source: §8 · p. 11

Zero on temporal knowledge-editing benchmarks

  • KNDB scores 0.000 on three temporal knowledge-editing benchmarks because their ground truth expects later writes to win, while KNDB’s tiebreak is first-committer-wins.The current implementation does not expose a last-writer-wins tiebreak configuration; pg_lww’s 1.000 at c=1 falls to 0.11–0.27 at c=8.Source: §8 · p. 11

Reputation farming defeats k* = ∞

  • The unbounded-k* argument assumes fresh, unfarmed adversarial identities; reputation farming into Tier A defeats the lattice, and long-lived farming requires external audit or stricter tier assignment beyond scope.The stated defense covers an unfarmed adversary and a writer without tier control.Source: §8 · pp. 11–12

R2 and R5 SPI cost per non-MEASURED insert

  • R2 and R5 each incur a full SPI cycle per non-MEASURED candidate before the F6 lock; a 5,000-row test measures R2 overhead at about 7 μs per row, roughly 35% above the MEASURED baseline.Source: §8 · p. 12

Batch-size ceiling

  • At the default lock setting, transactions face a batch ceiling of about 15,000 rows, including COPY; raising max_locks_per_transaction to 1024 raises it to about 215,000.Dropping the lock reopens the F6 integrity leak.Source: §8 · p. 12

Schema-change and WAL boundaries

  • A table owner can move a relation to plain heap with ALTER TABLE SET ACCESS METHOD, unbinding the TAM callback; the custom WAL resource manager is annotation-only, while heap WAL provides durability.Source: §8 · p. 12

Inherited heap semantics and atomicity

  • KNDB inherits heap behavior, including bugs, on 37 of 44 callbacks; eviction atomicity is PostgreSQL’s, not a novel guarantee of the TAM.Source: §8 · p. 12

Serialisable-level guarantees

  • Under SERIALIZABLE, inherited relation-granularity predicate locks can cause SSI aborts between writers on nonoverlapping slots; slot-level predicate locking is out of scope.Source: §8 · p. 12

9 CONCLUSION

KNDB implements engine-level epistemic typing in PostgreSQL and evaluates it against confidence forgery and truth-discovery baselines. Its reported advantage is strongest at and above density saturation, with explicit workload and implementation boundaries.

9 CONCLUSION

  • The PostgreSQL 18 TAM uses seven callback overrides, engine-assigned epistemic kinds, per-slot advisory locks, and a first-committer-wins tiebreak to enforce write-time conflict resolution.The remaining 37 callbacks delegate to heap.Source: §9 · p. 12
  • KNDB beats a confidence-only arbitrator by 63 and 92.7 percentage points on two workloads, with the kind-axis wins confirmed by source-rebuild disable-and-test.Against truth-discovery baselines, it is competitive below the density-saturation cell and dominant at or above it; the threshold is predicted by k* ≈ ρ_alg · h_top.Source: §9 · p. 12
  • KNDB does not win on last-writer-wins workloads, and the conclusion flags batch-size, schema-change, WAL, and reputation-farming boundaries.Source: §9 · p. 12
Loading 2609.36795v1…