PNS-World

1. Purpose

A deterministic event-stream benchmark for transcript-free persistent memory. A model consumes each event once, in order, and must answer later questions from bounded persistent state alone: no accumulated chat template, no transcript replay, no re-reading history. Because the evidence has already scrolled past, any cross-event dependence must survive inside whatever fixed-size state the model carries — which makes the benchmark a causal instrument rather than a retrieval one. You can swap one lifetime's state into another, reset it, zero it, or lesion one store, and measure what the behaviour loses.

A lifetime is one simulated agent-and-environment stream: typed events, interleaved questions, exact-solver gold labels, 256–4096 events long, deterministic from a seed. No human annotation, no model-generated labels.

2. Event and task families

element	what it forces
drifting enum attributes (location, phase, status, …)	semantic state that must be updated, not accumulated
exact values (codes, readings, budgets) with correction chains	exactness where approximation is unsafe
revert corrections naming an arbitrary earlier assertion	defeats "newest record wins" — the memory-hard core
relations (owner, project, vault)	two-hop binding
operations → commit → delayed asynchronous result	self-reference and environment handles
goals with priorities: set, re-rank, close	dynamically changing goals and subgoals
deadlines against an integrated clock (only per-event dt is visible)	elapsed-time transitions

Eleven question families — SEM_LATEST, SEM_2HOP, EXACT_DELAYED, EXACT_2HOP, SELF_REF, HANDLE_REF, TEMPORAL_ORDER, DEADLINE, GOAL_TOP, IMMEDIATE_CMP, OP_EMIT — about 26 answered questions per 256-event lifetime.

3. Splits

directory	split id	lifetimes	events	length	first seed	labels
data/e1_original/train	train	120,000	30,720,000	256	0	original
data/e1_original/val	val	2,000	512,000	256	1,000,000	original
data/e1_original/val_1k	val_1k	400	409,600	1024	1,100,000	original
data/e1_original/val_4k	val_4k	120	491,520	4096	1,200,000	original
data/e1_original/test	test	2,000	512,000	256	2,000,000	original
data/e1_original/test_1k	test_1k	400	409,600	1024	2,100,000	original
data/e1_original/test_4k	test_4k	120	491,520	4096	2,200,000	original
data/e1_corrected_eval/train_relabel	train_relabel	2,000	512,000	256	0	corrected SEM
data/e1_corrected_eval/val_relabel	val_relabel	2,000	512,000	256	1,000,000	corrected SEM
data/e3_corrected/train	e3_train	120,000	30,720,000	256	3,000,000	corrected
data/e3_corrected/dev	e3_dev	2,000	512,000	256	3,500,000	corrected
data/e3_corrected/dev_1k	e3_dev_1k	400	409,600	1024	3,600,000	corrected
data/e3_corrected/confirmation	e3_conf	2,000	512,000	256	3,900,000	corrected
data/e3_corrected/confirmation_1k	e3_conf_1k	400	409,600	1024	3,950,000	corrected

Seed namespaces are disjoint by construction (the two *_relabel splits deliberately reuse the E1 seeds; see §4) and cross-checked by content digest — 0 collisions. Exact ranges, per-shard SHA-256 and gold mode: manifests/splits.csv, manifests/seed_namespaces.json, manifests/corpus_manifest.json.

val_1k / val_4k and their test counterparts are 4× and 16× beyond any trained horizon, so they measure extrapolation, not interpolation. test* and confirmation* were each read exactly once in the study under a sentinel that refused a second run; both uses are spent and recorded, and the splits carry no further restriction here.

4. Original vs corrected labels

A defect was found in the benchmark's own gold labels after the first experiment was complete: Chain.current() returned the last appended valid assertion rather than the chronologically latest one, so for ~21% of SEM_LATEST / SEM_2HOP questions the gold was stale (audit/AUDIT.md §7). The original is not replaced. Three label modes are published separately:

mode	directory	what it is
v1 — original	data/e1_original/	the corpus exactly as generated and trained on, so Experiment 1's published results still describe the data they were measured on
sem — corrected evaluation labels	data/e1_corrected_eval/	same seeds, byte-identical event streams, corrected SEM_LATEST/SEM_2HOP golds only — this is what makes forensic re-evaluation of already-trained checkpoints possible
v2 — corrected	data/e3_corrected/	corrected everywhere. Exact and temporal control flow depends on the resolver, so these streams necessarily differ from v1; Experiment 2 was generated and trained here from scratch

Verified when train_relabel was generated: tokens, event types, dt and offsets byte-identical to train; SEM golds differ on 21.8% of questions; every non-SEM gold identical.

A comparison that mixes v1 and v2 corpora is invalid and none is made in the paper — the corrected-corpus control was retrained from scratch rather than re-evaluated, for exactly this reason.

5. Load and reproduce

import numpy as np
z = np.load("data/e3_corrected/confirmation/shard_00000.npz")
print(z["lt_off"].shape, z["etype"].shape, z["family"].shape)

Shards are flat arrays plus offset indexes; bounded record stores are per-event add/evict deltas against a per-lifetime record table, so every consumer materialises the exact stores the generation-time reducer certified. benchmark/pns/data/view.py is the reference reader.

sha256sum -c manifests/SHA256SUMS.txt          # verify what you downloaded

pip install -r requirements.txt
export PNS_DATA=$PWD

# regenerate the corpus, deterministically (~35 s on 64 cores).
# PNS_GOLD selects the resolver: v1 | sem | v2
PNS_GOLD=v2 python3 benchmark/generate.py --n-train 120000 --e3-train 120000 --workers 48

# re-certify shortcut resistance. Normalised advantage H = (acc-base)/(1-base),
# one-sided: H >= +0.05 is a leak. Certified output ships as
# benchmark/null_attack_results_{e1,e3}.json -- on the memory-hard core every
# deterministic attack sits BELOW its within-chain baseline (H -0.19 to -0.35).
python3 benchmark/attack_nulls.py --out out/NULLS_E3.json

preregistration/ (five preregistrations with dated, append-only amendments) and audit/AUDIT.md (eleven recorded deviations, four of which bear on a claim) are published because the methodological claims rest on them. The 20 invariant tests need the model code and live in the companion repo.

6. License

CC BY 4.0 (LICENSE). Fully synthetic: generated by a deterministic simulator, labelled by an exact solver. No human-authored text, no scraped content, no personal data.

7. Citation

CITATION.cff.

@article{pnsworld2026,
  title  = {Beyond Transcript Replay: Causal Persistent Memory Through Protected Semantic Bindings},
  year   = {2026},
  note   = {Dataset: https://huggingface.co/datasets/nur-dev/pns-world, tag paper-v1.0}
}

8. Code and models

nur-dev/pns-bind-25m — the implementation, the 24 evaluated checkpoints, the causal-intervention code, and one-command reproduction of every headline number.