PaperlayerView exact revision on arXiv

Source-linked AI summary

Exact Record Omission in Delta Attention: A Transport Criterion, Its Cost, and a Replay Certificate

Vishwajith Ramesh

arXiv:2609.06872v3cs.LGcs.CLcs.CR

TL;DR

A deleted statement may stop appearing in answers while still shaping recurrent memory and later processing. This paper derives an exact receipt-transport criterion and evaluates replay; masking suppressed exact answers but candidate scoring still distinguished the edited state, while replay matched audited references.

  • Problem

    Answer suppression alone does not establish that a deleted record has stopped influencing recurrent memory or later processing.

    Source: §I · pp. 1–2

  • Method

    The paper derives a criterion for when transporting a saved boundary-state difference reaches the never-stored state and evaluates checkpoint replay as an auditable alternative.

    Source: §I · p. 2, §IV · p. 3

  • Results

    Masking suppressed all 40 exact target answers, yet candidate scoring distinguished masked from never-stored states with AUCROC 0.740; replay matched audited state arrays and logits across 80 contexts.

    Source: §I · p. 2

  • Takeaways & Limitations

    The criterion, behavioral probes, and state comparisons provide an audit of what a memory edit achieves, while measured replay and storage costs inform checkpoint-interval choices.

    Source: §VIII · p. 12

    The statistical conclusions cover fixed synthetic contexts, known candidate values, and probability-score access, not arbitrary prompting, relearning, or attacks limited to generated answers.

    Source: §VII · p. 12

Abstract

from arXiv · show

An assistant can stop repeating a deleted statement while its recurrent memory still carries that statement's influence. We examine this distinction by saving the state difference immediately after a record, carrying this saved difference, or receipt, through later updates, and comparing the corrected state with the state built from the same conversation with the record omitted. Unrolling the recurrence gives an exact criterion: transport reaches this never-stored state if and only if the additional differences created by later updates cancel after transport. We evaluate eleven conditions using 40 prospectively selected synthetic Kimi Linear contexts, with 40 separate calibration contexts and paired uncertainty estimates computed across complete records. Attention masking removed all 40 exact greedy target answers, yet a fixed three-query candidate-scoring attack with known candidate values and log-probability access attained AUC 0.740 (95% interval 0.695--0.809); the measured false-positive rate was 7.5%. Prompt-only forgetting still returned 33 of the 40 targets. Checkpoint replay matched the complete declared active state and audit logits in all 80 contexts, while preserving all 120 retained answers in the evaluation cohort. An independent Qwen cohort and matched original-bf16/8-bit Kimi controls confirmed native transport mismatch beyond the numerical error measured in matched controls. Replay also supported record replacement and successive deletions with a new record added between them. These results provide a practical audit of both the requested memory change and the assistant's remaining behavior, with replay work determined by the surviving suffix.

I. INTRODUCTION

A deleted statement can stop appearing in an assistant’s replies while still shaping its recurrent memory and later processing. The paper compares editing strategies with the state the same model would have built had the statement never been processed.

I. INTRODUCTION

  • Because a record can change both the initial recurrent state and how later messages are processed, subtracting its original footprint can leave downstream effects.The comparison is against the state built from the same conversation with the target omitted.Source: §I · pp. 1–2
  • The paper compares instruction-only forgetting, attention masking, receipt transport, and checkpoint replay as responses to a deletion request.Replay resumes from a checkpoint before the record, skips it, and processes the surviving discussion again.Source: §I · p. 2
  • An extraction adversary asks for the deleted value, a candidate-scoring adversary tests a suspected value using probabilities, and a privileged auditor compares against an omitted-record run.These access levels support different kinds of evidence about deletion.Source: §I · p. 2
  • The evaluation uses the released 48B Kimi Linear hybrid on synthetic administrative records, measuring both target disclosure and whether edited memory matches the never-stored reference.Each context contains one target assertion to remove and other information to retain.Source: §I · p. 2
  • AUCROC 0.740 distinguished masked states from never-stored references despite suppression of all 40 exact target answers; prompt forgetting elicited 33 of 40 targets.Replay matched the audited internal-memory arrays and next-token score vectors across 80 contexts, preserving all 120 retained answers.Source: §I · p. 2

II. TERMS AND SCOPE

  • The declared active state includes the arrays and caches needed to resume computation, and replay recomputes surviving-token representations from an unaffected prefix checkpoint.The paper calls the reference state built without the target the never-stored state.Source: §I · p. 1, §II · p. 2

III. RELATED WORK

The paper situates inference-time deletion in work on training-data unlearning, parameter and external-memory edits, and recurrent-state models. It derives an exact test for receipt transport and explains how different write rules affect what must be transported.

III. RELATED WORK

  • Unlike training-data unlearning or individually addressable memory edits, this work studies fixed model weights and records folded into a shared recurrent state.It uses recomputation and checkpoint replay to audit record omission in that setting.Source: §III · p. 3

IV. THE TRANSPORT CRITERION

  • Receipt transport reaches the never-stored state exactly if the accumulated differences created by later updates cancel after transport.The criterion follows by unrolling the recurrence; those additional differences are called forcing.Source: §IV · pp. 3–5
  • The criterion diagnoses the specified transported receipt, not every possible deletion method, and checking it requires computing the reference run.Large forcing vectors alone do not prove failure because they can cancel after transport.Source: §IV · p. 4
  • Additive writes preserve a static receipt, per-channel decay uses a diagonal ledger, and delta updates generally require an ordered, non-diagonal suffix product.The product can be applied sequentially without forming a dense matrix.Source: §IV · pp. 4–5
  • A diagonal ledger handles decay but not cross-channel mixing; full matrix transport handles the mixing, while the forcing condition still determines whether correction is exact.The two-channel example illustrates why the delta update can create a component that a diagonal ledger cannot represent.Source: §IV · p. 5

V. MODELS, CORPORA, AND AUDIT DESIGN

The audit combines Kimi Linear mechanism experiments with a prospective cohort, additional model-family checks, and explicit state comparisons. Its design separates answer suppression, score-based detection, and state agreement while measuring replay and transport under controlled conditions.

V. MODELS, CORPORA, AND AUDIT DESIGN

  • Kimi Linear’s recurrent KDA layers and attention caches were audited separately, with full receipt transport modifying only recurrent matrices.The main cohort used released 8-bit weights; matched original-bf16 and 8-bit controls were also run.Source: §V · p. 5
  • Mechanism experiments used synthetic patient records, TOFU facts, and MIMIC-IV-Ext-CDS notes, retaining record-context pairs only when the model met prespecified recall thresholds.They compared additive writes, decay, and full delta updates, using both frozen-input and native never-stored references.Source: §V · p. 5
  • The suffix sweep tested cuts from 0 to 4,096 tokens and measured target log-probability lift, first-token rank, KL divergence, and generated answers.Configurations shared corpus-specific suffix pools and were not independent records.Source: §V · p. 5
  • The prospective cohort used 40 evaluation and 40 disjoint calibration contexts across six synthetic domains to assess 11 conditions, including five edits and replay.The design also compared instruction-to-forget with a neutral instruction of equal token count; all evaluation contexts met the recall criterion.Source: §V · pp. 5–6
  • The candidate-scoring adversary averaged target log probabilities from three fixed prompts, with its detection threshold selected on calibration contexts and paired intervals resampled by complete context.Direct target and retained-record answers used greedy decoding and an exact-match rule.Source: §V · p. 6
  • Replay audits compared active recurrent, convolution, and attention-cache arrays, cache offsets, and complete audit logits; additional tests covered Qwen, precision controls, replacements, and sequential deletions.The prospective comparison required exact array and offset matches, with maximum absolute difference zero after conversion to float32.Source: §V · p. 6

A. Later context changes a record’s recurrent effect

Later suffixes can change a record’s recurrent effect, so transporting its receipt need not match the never-stored state. Replay matched that state in the evaluated configurations, while partial replay left substantial residuals.

A. Later context changes a record’s recurrent effect

  • Decay introduced up to 10.9% movement, reduced to numerical noise by a diagonal ledger; delta-rule movement was 12–49%, remaining 8–49% after that ledger.Additive writes were suffix-independent to numerical noise.Source: A. Later context changes a record’s… · p. 6

B. Transport corrects frozen inputs but leaves a native residual

  • Transport matched frozen-input effects to 6.1 × 10−6 maximum relative error, but native residuals remained 0.7–32.9% (median 12.6%) across 36 evaluations.Later keys, values, gates, and step sizes differed between trajectories, attributing the residual to changed later updates rather than transport error.Source: B. Transport corrects frozen inputs… · pp. 6–7

C. State residuals across suffix lengths

  • Replay matched the never-stored state exactly (maximum state difference 0) across all tested synthetic and TOFU configurations.At every suffix cut, the measured state difference agreed with transported receipt plus accumulated forcing within 9.0 × 10−4 relative error.Source: C. State residuals across suffix lengths · p. 7
  • At 4,096 suffix tokens, the imprint remained 4.5%; transported receipt subtraction left 4.4%, diagonal ledger subtraction 6.0%, and static receipt subtraction 27.5%.The static receipt moved the state further from the reference than leaving the imprint in place.Source: C. State residuals across suffix lengths · p. 7
  • The forcing norm ratio rose from a median of 0.67 after 16 tokens to 0.98 at the longest suffix.Individual layers ranged from 0.000 to 2.712; ratios above one reflect cancellation.Source: C. State residuals across suffix lengths · p. 7

D. Partial recomputation and deferred-transport costs

  • Recomputing the first eighth, quarter, and half of the suffix left residual norms of 0.77, 0.69, and 0.56 times the original state difference.At the evaluated cut points, a short initial replay did not remove most of the residual.Source: D. Partial recomputation and… · p. 7
  • The transition log costs 481 KiB per token, while a complete recurrent and convolution-state checkpoint costs 41.4 MiB regardless of context length.In this implementation, stacked transport took 9.4 ms per token versus 1.32 ms per token for prefill; the comparison excludes the hybrid’s other memory.Source: D. Partial recomputation and… · p. 7

E. Masking reduces recovery but leaves a distribution difference

Masking suppresses exact target answers but leaves detectable signal, while transport reduces that signal without establishing state equality. Replay matches the never-stored state in qualifying audits, with work determined by the surviving suffix.

E. Masking reduces recovery but leaves a distribution difference

  • All three receipt corrections left median synthetic target lift near 2.2 nats per token and greedy recovery in every configuration; masking sharply reduced recovery.Masking nevertheless left nonzero first-token KL on all 216 synthetic and 144 TOFU configuration–cut pairs; synthetic median KL was 0.052 nats.Source: E. Masking reduces recovery but leaves… · p. 7, Figure 2 · p. 9

F. Combining masking and transport reduces the measured target signal

  • Masking removed all 40 exact answers, yet candidate scoring detected 12/40 contexts (30%; 95% interval 5–50%) with a measured false-positive rate of 7.5%.The paired detection advantage over the null was 22.5 percentage points ([2.5, 37.5]).Source: F. Combining masking and transport… · pp. 7–8
  • 0.0178 nats per token was the mean lift with masking and full receipt transport; AUC fell from 0.740 to 0.519.Both masking alone and masking with transport produced zero exact target answers, but the residual AUC separation was not zero.Source: F. Combining masking and transport… · p. 8
  • All 120 retained answers were exact in every condition; retained-score and continuation-NLL intervals for masking with transport included zero.Continuation checks used a small synthetic passage bank and targets without deleted information.Source: F. Combining masking and transport… · p. 9

G. Replay exactly matches the audited state and preserves retained answers

  • 0.00 s of suffix processing followed deletion of the newest record, versus 6.70 s after the oldest, close to the 6.77-s full rebuild.Saving every sixteenth checkpoint reduced storage from 5.22 GiB to 0.36 GiB and raised estimated mean replay time to 3.89 s.Source: Figure 3 · p. 10, G. Replay exactly matches the audited… · p. 10
  • 21 of 22 primary corpus-audit configurations met the recall criterion; replay had zero residual for those, while the oldest MIMIC-Note target fell below threshold at 0.00451 nats per token.A separate fixed-position comparison found zero replay residual for all 12 targets.Source: G. Replay exactly matches the audited… · p. 10

H. Predicted transport behavior holds in the tested recurrent models

Across the tested recurrent families, receipt transport behaved as predicted under frozen inputs, but native trajectories retained differences that replay removed. Replay also matched corrected references through record amendments and successive deletions, while the discussion identifies evaluation and deployment boundaries.

H. Predicted transport behavior holds in the tested recurrent models

  • In Mamba-2 and Falcon-H1, diagonal-ledger and full-matrix receipts produced the same frozen-input correction and met tolerance; RWKV-7 required the full-matrix receipt.On native trajectories, every family showed nonzero forcing, while replay reached the never-stored state exactly.Source: H. Predicted transport behavior holds… · p. 10

I. Transport mismatch persists at both bf16 and 8-bit precision

  • In matched Kimi controls, aggregate residuals were 0.645–0.817 of the native record difference in bf16 and 0.665–0.780 in derived 8-bit, each over 20,000 times its matched-control numerical error.Three of 160 layer–context residuals in each variant were at or below their own floor, so the aggregate finding does not apply to every layer.Source: §I · p. 10

J. Replay supports record replacement and successive deletions

  • Replay matched the corrected-from-start reference in all 12 amendment contexts, retrieving all 12 new values and none of the old values; retained answers were 35/36 exact and all contained the correct value.All amended states and audit logit vectors matched their corrected-from-start references.Source: J. Replay supports record replacement… · pp. 10–11
  • All 36 transitions across six deletion-pair paths matched fresh surviving-record references, with a new record added between deletions.Both deletion orders produced the same record keys and probe scores in all six pairs.Source: J. Replay supports record replacement… · p. 11

VII. DISCUSSION

  • The tested scoring attack assumes known candidate values and probability-score access; it does not establish recovery of unknown codes through ordinary chat.The authors also identify larger, more varied conversations, adaptive or generated-answer-only attacks, and later copies as areas for further evaluation.Source: §VII · pp. 11–12
  • The criterion diagnoses when transported receipts reach the never-stored state, distinguishing incorrect transport from later changes that a correctly transported receipt does not remove.The matrix interventions leave convolution state and hybrid attention caches unchanged, so their component-specific contributions remain unresolved.Source: §VII · pp. 11–12

VIII. CONCLUSION

The paper presents exact record omission as a reference for auditing memory edits, using a transport criterion and replay comparisons alongside behavioral probes. It concludes that replay matched tested references and supported correction and sequential maintenance, while noting boundaries on data access and deployment.

VIII. CONCLUSION

  • The criterion, behavioral probes, and state comparisons provide an audit of what a memory edit achieves; replay matched the declared reference in Kimi and Qwen audits and supported record correction and sequential maintenance.Measured storage and replay costs give operators a basis for choosing a checkpoint interval to meet deletion requirements.Source: §VIII · p. 12

ETHICAL CONSIDERATIONS

  • Clinical experiments use credentialed PhysioNet access with local processing and aggregate-only reporting; deployment would additionally require authenticated authorization and tamper-evident deletion logs.Model weights remain fixed, and the audit compares active state and behavior under the paper's stated access assumptions.Source: ETHICAL CONSIDERATIONS · p. 12

OPEN SCIENCE

  • The reproduction package includes code, source-free results, protocols, and generated numerical macros, but credentialed MIMIC source data and record-bearing outputs cannot be redistributed.Rerunning those experiments requires independent authorized access.Source: OPEN SCIENCE · pp. 12–13

LLM USAGE CONSIDERATIONS

  • Language models assisted with literature search, code development, figure drafting, and conference-format preparation; the authors state they are responsible for the manuscript and scientific conclusions.The authors note that AI assistance does not independently verify generated code, references, claims, or experimental records.Source: LLM USAGE CONSIDERATIONS · p. 13

A. Preselected Kimi cohort and paired measurements

The experiments use prospectively fixed synthetic contexts, paired measurements, and separate calibration contexts to compare answer suppression, score-based detection, and state restoration. Replay matched the declared active state in the Kimi cohort, while attack and Qwen analyses tested remaining score separation and transport behavior.

A. Preselected Kimi cohort and paired measurements

  • The prospective Kimi cohort used 40 evaluation and 40 separate calibration contexts, with frozen inputs, probes, decoding, and analysis rules; contexts were independent synthetic constructions, not independent people or training runs.Records spanned six administrative settings and included retained records, a target record, and a frozen neutral suffix.Source: A. Preselected Kimi cohort and paired… · p. 14
  • Across the main cohort, all 120 retained answers were exact under every condition; target answers numbered 33 for forgetting and zero for masking, combined masking/transport, and the three replay/reference conditions.The measurement used three fixed target prompts and a 24-token greedy-generation cap.Source: A. Preselected Kimi cohort and paired… · p. 15
  • The study reports continuation NLLs of 9.518 for replay and never stored, 9.495 for masking, and 9.565 for masking with transport; the six-passage bank is not a broad language benchmark.NLL measures prediction of unseen passages, with lower values better.Source: A. Preselected Kimi cohort and paired… · p. 15

B. Three-call scoring attack and threshold calibration

  • The three-call attacker scored a supplied candidate value for a known record identifier using fixed prompts and log probabilities; it had no state access or never-stored evaluation references.The attack score was the mean of three per-token scores, with larger scores flagging a record as present.Source: B. Three-call scoring attack and… · p. 15
  • At the calibrated threshold, three of 40 evaluation never-stored scores exceeded it, a measured FPR of 7.5% with Wilson 95% interval 2.6–19.9%.The balanced synthetic design does not establish a deployment FPR guarantee.Source: B. Three-call scoring attack and… · p. 15
  • Masking with transport had AUC 0.5188, while the exploratory source-context-excluded AUC was 0.5115 [0.5033, 0.5238] for masking with transport and 0.7333 [0.6793, 0.8007] for masking alone.The exploratory calculation does not replace the primary AUC or change the attack.Source: B. Three-call scoring attack and… · p. 15

C. Complete active-state comparison

  • All 80 Kimi contexts had zero differences between replay and independent never-stored rebuilds across declared state arrays, cache offsets, and audit logits; native transport residual averaged 0.7011 [0.6861, 0.7146] of the original record effect in evaluation contexts.Unused attention-buffer capacity was excluded from the comparison.Source: C. Complete active-state comparison · p. 16

D. Independent Qwen confirmation

  • For Qwen3.5-4B-Base, all 64 replay and 64 repeat comparisons had exact numerical equality on 65 arrays or outputs and 82 bookkeeping flags, covering active layers, final logits, and logical cache position.The analysis resampled 32 records, and the suffix texts differed, so the paired contrast does not isolate a causal effect of length.Source: D. Independent Qwen confirmation · p. 16

E. Kimi precision comparison using the same source weights

Matched-source Kimi controls test whether native transport mismatch persists beyond measured arithmetic discrepancies; replay-based maintenance also installs corrections and supports successive deletions with an intervening new record.

E. Kimi precision comparison using the same source weights

  • Across eight preselected contexts per precision, replay matched all 94 active arrays, seven offsets, and audit logits; the comparison used each precision’s own never-stored reference.The matched control derived an 8-bit model from the original bf16 weights; its residuals support native mismatch beyond measured arithmetic discrepancies.Source: E. Kimi precision comparison using the… · pp. 16–17
  • 21,928–84,695× (8-bit) and 20,970–80,373× (bf16) were the global transport residuals relative to each precision’s measured arithmetic-error floor.Three of 160 layer–context residuals per precision were at or below their own floor, so the aggregate result does not establish above-floor mismatch in every layer.Source: E. Kimi precision comparison using the… · p. 17, Table VIII · p. 18

F. Amendment and sequential maintenance

  • All 12 amendment contexts returned the old answer before correction and the new answer afterward, with score changes relative to the corrected-from-start reference equal to zero.Retained answers matched exactly in 35 of 36 queries, and all 36 contained their targets.Source: F. Amendment and sequential maintenance · pp. 17–18
  • Across six target pairs, both deletion orders left both deleted answers absent and produced identical surviving-record keys and saved probe scores.Every retained and newly admitted answer was exact in both orders for five pairs; the independent sequential unit was the pair, not the path or transition.Source: F. Amendment and sequential maintenance · p. 18
  • All 48 intervention-to-reference comparisons and 48 independent reference repeats had zero measured state and logit differences.Runs overlapped other experiments and a weight download, so their times were not used as controlled latency measurements.Source: F. Amendment and sequential maintenance · p. 18

APPENDIX B MECHANISM PROTOCOLS AND SUPPORTING RESULTS

The appendix specifies cohort qualification and audit metrics, describes a three-record Qwen protocol, and estimates the storage–replay tradeoff from checkpoint spacing.

A. Replay records and state checks

B. Corpus, suffix and cross-family audit tables

  • The mechanism tables distinguish attempted from qualified cases, report target recovery and sampled-answer hits, and compare masking with replay and receipt checks across cohorts.The table descriptions specify metrics and cohort structure, but the supplied passages do not provide their cell values.Source: B. Corpus, suffix and cross-family… · pp. 18–19, Table X · p. 19

C. Three-record Qwen protocol

  • The Qwen confirmation pinned Qwen3.5-4B-Base and Transformers 5.12.1, testing three target records with two equal-length suffixes per record and comparing replay outputs across 65 specified arrays or outputs and 82 flags.The supplied passages describe the protocol but do not report the comparison’s outcomes.Source: C. Three-record Qwen protocol · p. 19

D. Sparse checkpoints

  • Every 16th checkpoint used 0.36 GiB with 3.89 s estimated mean replay, while every 32nd used 0.20 GiB and 4.31 s.The latency estimates came from a linear fit; keeping only the oldest checkpoint used 0.08 GiB but raised estimated replay to 6.89 s.Source: D. Sparse checkpoints · p. 19

E. Attention masking compared with replay

The comparisons distinguish answer suppression from matching the never-stored state: replay matches that reference, while receipt correction depends on the write rule and accumulated forcing.

E. Attention masking compared with replay

  • The ward-code example remained recoverable after recurrent matrices were zeroed, while replay made the model answer that the code was not provided and matched the reference’s final logits and 80 KDA arrays.The example illustrates why matching a greedy answer alone does not establish matching state.Source: E. Attention masking compared with… · pp. 19–20

F. Amendment

  • Record replacement restores the last unaffected checkpoint, ingests the correction, and replays the surviving suffix; the reported example matched the corrected-from-start reference on final logits and all 80 KDA arrays.The illustrated replay covered 55 tokens in 0.16 s.Source: F. Amendment · p. 21

G. Full write-rule hierarchy

  • Additive writes preserve a static receipt, per-channel decay requires a diagonal ledger, and KDA’s delta rule requires full-matrix correction.Whether transport removes the native difference is separate: later updates can create additional differences.Source: G. Full write-rule hierarchy · p. 21

H. Proof detail for Corollary 1

  • With zero accumulated forcing, full-matrix transport reaches the never-stored state; KDA’s diagonal-minus-rank-one factors generally yield non-diagonal products dependent on suffix keys.In 36 evaluations, the residual after full-matrix correction ranged from 0.7% to 32.9% of the record-present suffix-A state norm, with a median of 12.6%.Source: H. Proof detail for Corollary 1 · p. 21

I. Interpreting the forcing norm ratio

  • The forcing norm ratio compares accumulated forcing with the uncorrected state difference; it is not the fraction of that difference explained by forcing.The suffix sweep reached 2.712 at 4,096 tokens, where receipt and forcing could each exceed the remaining difference because they partly cancel.Source: §I · p. 21

J. Choosing how often to save checkpoints

Checkpoint cadence trades storage for replay work; the same replay procedure also installs corrections, while suffix-sweep audits check replayed state and cache contents.

J. Choosing how often to save checkpoints

  • Every 16th checkpoint used 0.36 GiB and an estimated 3.89 s replay, versus 5.22 GiB and 3.49 s when saving every boundary.Keeping only the oldest checkpoint used 0.08 GiB but raised estimated replay to 6.89 s; sparse-checkpoint latencies are estimates and depend on hardware and workload.Source: J. Choosing how often to save… · p. 21

K. Using replay to install a correction

  • Replay installs a correction by restoring the pre-correction checkpoint, ingesting the replacement, and replaying the surviving suffix; the amendment check reached the corrected-from-start reference exactly.The same procedure checks omission and amendment using the same set of state arrays.Source: K. Using replay to install a correction · p. 21

L. Suffix-length sweep protocol

  • At 4,096 suffix tokens, replay reproduced the never-stored state with zero maximum absolute difference; in eight configurations, logical MLA cache contents also matched exactly.The cache comparison covered keys and values up to the current offset in all seven layers; unused block-allocated buffer capacity could differ.Source: L. Suffix-length sweep protocol · p. 23

M. Cross-family protocol and pre-registered predictions

All pre-registered frozen-input predictions held across Mamba-2, Falcon-H1, and RWKV-7, while native forcing was nonzero. Fixed-trace replay reuses specified later messages and certifies only the declared active state after replay.

M. Cross-family protocol and pre-registered predictions

  • Adapters advanced suffix tokens one at a time and captured recurrent states before and after each token; RWKV-7 formed its recurrence from captured inputs, unlike Mamba-family residual capture.The behavioral probes changed recurrent matrices while retaining other record-present state components, so they did not isolate every persistent component.Source: M. Cross-family protocol and… · p. 23
  • The frozen-input counterfactual holds later inputs to the record-present run fixed, whereas the native reference follows the never-stored run with its own later inputs.Transport checks excluded layers whose frozen-input difference was below 10−3 of the never-stored state norm.Source: M. Cross-family protocol and… · p. 23
  • RWKV-7 recurrence error was 0; the Qwen3.5-4B check measured 1.5 × 10−6 recurrence error, 2.4 × 10−6 frozen-input transport error, and 0.018–0.938 native forcing norm ratio.Replay had zero residual across arrays, outputs, and flags on all six paths.Source: M. Cross-family protocol and… · p. 23
  • All pre-registered predictions held: static receipts are inexact for Mamba-2, Falcon-H1, and RWKV-7; exact corrections vary by family, and native forcing is nonzero.The frozen-input predictions used a maximum error tolerance of 1 × 10−4 of the state norm.Source: M. Cross-family protocol and… · pp. 23–24

N. Fixed-trace replay and causal regeneration

  • Replay reuses assistant turns, summaries, retrieval writes, and tool actions, including content influenced by the deleted record; causal regeneration was not evaluated.The certificate covers only the specified active state after authorized replay, not earlier outputs, external copies, or later influenced content.Source: N. Fixed-trace replay and causal… · p. 24

Related papers

Loading 2609.06872v3…