acceptodds
Under review as a conference paper at ICLR 2027

Machine Causality Stratification: Derived Authority, Reversible Causal Ledgers, and Coverage-Gated Capability

Abstract

Memory systems for agents increasingly store causal claims, yet the causal lines surveyed in Section 2 report capability as a single level rather than a graded, evidence-derived one. That framing cannot answer what matters when a system’s own memory encodes causal claims: how deep is the claim, why should it be believed, can its authority be revoked, and can one high-level edge make the whole system look causally mature. We present CKB, a causal-memory substrate with a machine-governed causal-authority lifecycle: a graded CDL (Causal Depth Level, associative to interventional), a predicate-derived and reversible CAL (Causal Authority Level, a strict accumulation over evidence predicates), a per-edge capability coordinate CCC, and a coverage-gated envelope CCE that stops one high-depth edge from inflating the reported depth. Authority is captured in an append-only, reversible witness DAG whose transitions are receipted and replayable. We separate what kind of causal claim an edge expresses (CDL) from how much authority the system may assign to it (CAL). An implemented Rust substrate and a cross-language Python reference oracle execute the calculus at the substrate and conformance layer; we report structural certificates, exact indexed recall, a registered counterfactual edge audit, and downstream measurements including a null result for incremental explicit-structure presentation. Downstream, a preregistered paired test (N = 80, a LongMemEval-S public-development subset) finds that external memory helps substantially (+32.5 pp over no memory, p = 3.0e-8), but that presenting the same evidence as typed causal/temporal structure yields no detectable gain over a flat presentation (+1.25 pp, p = 1.0); the null persists under a matched-token placebo arm. On the full 500-question public development set, the frozen official gpt-4o-2024-08-06 evaluator scores a frozen candidate at 493/500; with no baseline and no recorded answer generator, we report this as an observability result, not a comparative or generalization claim (Section 8).

Primary area:
causal reasoning
Keywords:
Causal reasoning, Knowledge representation and reasoning, Neurosymbolic and hybrid AI systems, Memory and retrieval systems, Datasets and benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_bcBEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CKB, a causal-memory framework that separates the depth of causal claims from their evidence-derived authority. It introduces reversible authority updates and a coverage-based measure of system capability, alongside a prototype implementation and preliminary evaluations.

Critical strengths
  1. The paper addresses an interesting problem: how agents should maintain, justify, and revise causal claims in persistent memory as supporting evidence changes.
Critical weaknesses
  1. The paper is hard to follow and reads very AI, with excessive terminology, repetitive qualifications, and unclear details.

  2. The experiments are largely small-scale demonstrations and synthetic checks of the proposed rules. They provide limited evidence of practical usefulness.

Questions for authors

N/A

Reviewer_AGFHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes CKB, a causal-memory substrate for agents. Each stored edge gets a Causal Depth Level (CDL 0-6, from associative to interventional) describing what kind of causal claim it expresses, and a Causal Authority Level (CAL) derived by the system, not the claim's author, from a chain of evidence predicates P1-Pk with strict accumulation. The pair is the edge's capability coordinate (CCC), and a coverage-gated envelope (CCE) reports the system's verified depth as the highest depth supported by at least a fraction tau of edges at sufficient authority, so one deep edge cannot inflate reported capability. Authority transitions are appended to a reversible witness DAG. The authors implement this in Rust with a Python reference oracle and report passing adversarial invariant tests (24 + 17 tests, 16 conformance checks), exact indexed recall that is 26.3x faster than a full scan on a synthetic 2000-edge fixture, and a synthetic CCE demonstration. Downstream, a preregistered paired test (N = 80, LongMemEval-S dev subset) finds that external memory beats no memory (+32.5 pp) but typed causal/temporal presentation gives no detectable gain over a flat presentation (+1.25 pp, p = 1.0), and this null survives a matched-token placebo arm. A frozen gpt-4o judge scores one candidate 493/500 on the 500-question dev set.

Critical strengths
  1. The motivation / idea in this paper sounds and implemented well. Separating the memory components with kinds of causal claim is practical and the initial result from Section 7. Experiment is encouraging.
Critical weaknesses

1.. The writing of this paper needs cleaning up of AI report which consists of bullet points and enumerates interchangeably rather than mostly paragraphs. also many writing style is not standard academic writing, e.g. M0 - M3 on section title, scope note, and bullet contributions/limitation with component explanation only.

  1. The empirical result can be more comprehensive. Right now the result is small and non-independent. We have 493/500 performance with no baseline number reported, and section 7.3's result comes from author-assigned adjudications made under a rule written after seeing the judges' outputs.
Questions for authors

N/A. I will read others' comment and questions as well.

Reviewer_SACyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on causal memory and auditable representations for intelligent agents, proposing a framework that incorporates causal depth hierarchies, causal authority hierarchies, capability coordinates, and coverage constraint envelopes. Experiments conducted on LongMemEval-S indicate that while external memory is beneficial, explicit causal structures offer no significant incremental gains over flattened inputs. Overall, the paper attempts to establish a formal, auditable causal memory system; however, its core definitions, experimental setup, and the boundaries of its conclusions remain difficult to grasp with clarity and consistency in their current presentation.

Critical strengths
  1. Emphasizing the auditability and revocability of agent memory. This motivation is valuable, particularly in high-risk agent systems, where evidence traceability and the revocation of permissions are critical issues.
  2. Attempting to organize causal claims into a hierarchical structure, rather than relying on simple Boolean judgments. This approach helps advance more nuanced modeling of causal memory.
Critical weaknesses
  1. The clarity of the exposition is severely lacking, making core concepts difficult to grasp. There is no consistent or accessible framework for interpreting the numerous terms, abbreviations, and internal notations, and the text frequently shifts between different levels of abstraction. For a conference paper intended for a broad readership, the current writing style raises the barrier to understanding and increases the risk of misinterpretation.
  2. The core framework lacks a sufficiently clear motivation and formalization of the problem. It is difficult for readers to discern whether concepts such as "causal depth" and "causal authority" are objective metrics, model predictions, or merely human-defined labeling schemes. Furthermore, many of the rules appear to be ad-hoc engineering conventions rather than theoretical constructs supported by rigorous argumentation.
  3. The experimental evaluation is weak and insufficient to substantiate strong claims. The primary experiments are limited to the development set of a single dataset, lacking both an independent test set and comparisons across a broader range of tasks. Additionally, the presentation of certain results lacks transparency, making it difficult to assess the true contribution of the proposed method.
Questions for authors
  1. Could you rephrase the core distinction between CDL and CAL more concisely, and explicitly clarify whether each represents an objective definition, a subjective annotation, or a model output?
  2. Could you provide additional experimental results—using independent test sets or a broader range of datasets—to verify whether the current negative results are robust?
  3. Could you clearly outline the shared assumptions underlying all major experiments, and specify which conclusions are applicable only to evaluation environments like LongMemEval-S?
  4. Could you consolidate key definitions, notational conventions, and experimental protocols into a single, unified section to spare readers from having to flip back and forth through the main text?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.