Initializing PRISM000
› loading world model
VΔLK ResearchPRISM · WMA-V · v1.0.0
VΔLK/ PRISM
Read paperRead paper
[ PRISM · WMA-V ]
v1.0.0 · production release
arXiv preprint 2026

World modelspredict. PRISMverifies.

A verifiable world-model agent architecture where prediction errors drive constraint-checked state updates and trace-consolidated memory. The policy never writes state directly — a verified update cycle does.

PRISM_LIVEPREDICT s_{t+1}OBSERVE o_tERROR ε_tADMITPRESSUREGATE_τUPDATE ΔsCOMMIT τREPLAY ✓POLICY πCONSOLIDATEROLLOUT
SCROLL
PRISMWorld-Model Agent · VerifiablePrediction error drives update9 typed error channelsReplay-ExactHash-chained memoryPRISMWorld-Model Agent · VerifiablePrediction error drives update9 typed error channelsReplay-ExactHash-chained memoryPRISMWorld-Model Agent · VerifiablePrediction error drives update9 typed error channelsReplay-ExactHash-chained memoryPRISMWorld-Model Agent · VerifiablePrediction error drives update9 typed error channelsReplay-ExactHash-chained memory
[ 01 / architecture ]
the core problem

why current agents silently drift

Policyshould not writestate.

PRISM · update rule·§3 formal definition

// existing LLM agents — ReAct, Reflexion, Voyager:

st+1 πLLM(st, ot)// policy mutates state directly

// PRISM — the policy never writes state:

εt = d(ôt+1, ot+1) admit(ε) Δs commit

// prediction error is the only update signal

// every Δs passes admissibility before commit

// the chain is hash-verified before replay

“A model that predicts well
is not yet a model that commits well.”

— §1 motivation
No write authority
  • I.
    Policy network
    via token sampling
  • II.
    Structured parser
    via extraction output
  • III.
    Retriever / RAG
    via top-k context
  • IV.
    World-model latent
    via raw rollout
  • V.
    Tool callback
    via function return
  • VI.
    Operator
    via manual override

State writes only via the verified update cycle.

[ 02 / cycle ]

The update
cycle,
eight steps.

Every observation passes through the full inference cycle. Enforced as a durable workflow. Scroll to traverse one execution.

STEP010102030405060708
[ step 01 ]

The agent's current latent belief. A typed structure with entities, claims, and a posterior over the world.

formal
s_t = (entities, claims, posterior)
BELIEF_TYPED·enforced
CYCLE_PROGRESS
013%
[ 03 / properties ]

Six architectural
guarantees.

Each one is checked at compile time and verified at runtime. Remove any one and the architecture degrades to a vanilla retrieval-augmented loop.

[ property_01 ]
Replay-exact memory
commits since genesis
847,291
LIVElast hash: 7c3a...b9e1

State at any t is bit-exactly reconstructible from genesis.

Memory is a hash-chained replay log of typed transitions. Vector store is retrieval — not memory. Graph store is projection — not belief.

[ property_02 ]
Latent world model
z_tẑ_t+1

Prediction error becomes the update signal.

The latent predictor forecasts ẑ_t+1 before observation. Mismatch enters as one typed error channel — never as direct state write.

[ property_03 ]
Four-gate action commit
01
ADMISSIBILITY GATE
02
EVIDENCE GATE
03
POLICY GATE
04
SAFETY GATE
EXTERNAL_ACTION

Four gates. Then the world changes.

Every action emits a post-execution observation. The agent always closes the loop.

[ property_04 ]
Local-only inference
vLLM
SGLang
PyTorch
llama.cpp
sentence-T
CUDA

No external API in the update path.

Deterministic audit. Reproducible provenance. Sovereign data.

[ property_05 ]
Strict cycle
21
cycle steps
enforced as
durable workflow

No skips.

14_INVARIANTS
no_update_without_error/no_direct_state_write/no_policy_authority_over_state/no_parser_authority_over_state/no_retriever_authority_over_state/no_memory_storage_collapse/no_commit_without_trace/no_admission_after_error_violation/no_action_without_gate/no_action_without_transition/no_latent_to_state_without_error/no_operator_to_state_mutation/no_action_result_to_state_without_observation/predictor_decorrelation > 0.05/no_update_without_error/no_direct_state_write/no_policy_authority_over_state/no_parser_authority_over_state/no_retriever_authority_over_state/no_memory_storage_collapse/no_commit_without_trace/no_admission_after_error_violation/no_action_without_gate/no_action_without_transition/no_latent_to_state_without_error/no_operator_to_state_mutation/no_action_result_to_state_without_observation/predictor_decorrelation > 0.05/
[ 04 / live ]
Live inference console

Watch the model
commit, step by step.

Live simulation of PRISM's telemetry. Every transition is hash-chained, every metric streams in real time. Connect your own predictor — same surface.

prism.valkresearch.org / inference / live
RUNNING
Cycle pipeline
PREDICT
OBSERVE
ERROR
ADMIT
processing...
COMMIT
UPDATE TEMP τ
0.624
τ_t+1 = τ_t · d + u · g − r · k
Replay log · hash-chained
CHAIN_VERIFIED ✓
ε channels · current step‖Δ‖ = 0.348
struct
ent
rel
claim
temp
caus
evid
telos
lat
Metrics report
novelty0.71
relevance0.89
evidence0.62
contradiction0.18
risk0.09
pressure0.58
TELOS_ALIGNMENT
truth●●●●
provenance●●●●
safety●●●●
coherence●●●●

The console is a live simulation. In production it streams from PostgreSQL ledger via SSE.

[ 05 / theorem ]
§4 · formal guarantee
bit-exact reconstruction

Replay-Exact.

Let A be a PRISM-compliant agent. Under assumptions 1–3, for any time t:

  1. I.

    State sₜ is bit-exactly determined

    by the genesis state s₀ and the trace sequence (τ₀, τ₁, …, τ_{t−1})

  2. II.

    Every state update has a witness

    each transition Δs corresponds to exactly one logged τᵢ — no silent drift

  3. III.

    Tampering is detectable

    any modification to τᵢ breaks the chain h_τ = H(h_prev ∥ serialize(τ)) w.h.p.

[ corollary · necessity ]

“Any agent satisfying replay-exact verifiability must satisfy the PRISM separation. The architecture is implied, not chosen.

Assumptions
  • A1commit: S × Δs → S is deterministic and pure
  • A2Hash H is collision-resistant
  • A3Replay log {τᵢ} is append-only, tamper-evident
What the theorem does NOT claim

T1 guarantees auditability and reconstructibility — not improved task success, not predictor accuracy. PRISM bounds error commit, not error occurrence.

[ 06 / comparison ]

Five axes.
Nine architectures.
One row closes all five.

ReAct & descendants ignore the write boundary entirely. Active Inference and JEPA address prediction but not commit authority. Cognitive architectures satisfy parts. No prior LLM-agent design combines all five.

ReAct

Yao et al. 2023
  • Write boundary
    separated
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

Reflexion

Shinn et al. 2023
  • Write boundary
    separated
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

Voyager

Wang et al. 2023
  • Write boundary
    separated
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

Toolformer

Schick et al. 2023
  • Write boundary
    separated
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

Constitutional AI

Bai et al. 2022
  • Write boundary
    separated
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

Active Inference

Friston 2010
  • Write boundary
    separated
    N/A
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

JEPA / V-JEPA

LeCun 2022
  • Write boundary
    separated
    N/A
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

Soar / ACT-R

Laird 2019
  • Write boundary
    separated
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

PRISM

this work
  • Write boundary
    separated
  • Prediction
    required
  • Typed delta
    structurally
  • Gated action
    4× gates
  • Replay
    bit-exact

← drag horizontally · Table 1 · §2 of preprint →

[ 07 / benchmarks ]

An order of magnitude
less silent drift.

On suites measuring what PRISM was designed for — error-driven update fidelity, silent commit of unsupported claims, replay traceability — the architecture dominates baselines by 10–30×.

CDR↑ better
Contradiction Detection
0%
PRISM v1.0.0
ReAct47
RAG agent31
SCR↓ better
Silent Commit Rate
0%
PRISM v1.0.0
ReAct38
RAG agent52
TR↑ better
Traceability
0%
PRISM v1.0.0
ReAct0
RAG agent0
RVP↑ better
Replay-Validation Pass
0%
PRISM v1.0.0
ReAct0
RAG agent0
[ honest disclosure · §7.3 ]

Latency is the price.
We pay it openly.

PRISM pays a per-cycle cost. On the B1/B2 suites that cost is approximately 2.4× ReAct p95. For chat-only workloads this is prohibitive. For agents that must replay their own decisions, this is the price of verifiability.

0ms
PRISM p95
1800ms
ReAct p95
[ 08 / stack ]

Standard tools.
Verifiable wiring.

PRISM uses production ML components — V-JEPA, GLiNER-Relex, DeBERTa-MNLI, vLLM — as encoders and serving layers, never as state-writing actors. All inference runs locally.

Inference
  • Python 3.12runtime
  • PyTorchcore
  • vLLMLLM serving
  • SGLangstructured gen
  • llama.cppedge
Encoders
  • sentence-transformerstext emb
  • V-JEPAvisual latent
  • GLiNER-Relexentity / relation
  • DeBERTa-v3-MNLIclaim NLI
  • TimeMoEtemporal
State / Memory
  • PostgreSQLhash-chained log
  • Qdrantretrieval index
  • Neo4jgraph projection
  • RedisSTM cache
  • Apache WALdurability
Runtime / Obs
  • Temporaldurable workflow
  • FastAPIcontrol plane
  • Pydantic v2contracts
  • OpenTelemetrytracing
  • structlogaudit
$production gateruff check .mypy --strictpytest invariants/docker compose upreplay --verify-chainall green · sealed
[ 09 / paper ]

Read.Replay.Refute.

PRISM v1.0.0 reference implementation, the full B1/B2 benchmark suites, baseline configs, and replay logs from all benchmark runs — open-sourced for independent verification. PRISM is a discipline applied to existing models, not a new model. Plug in your own predictor. Audit the result.

Authors
[anon. for review]
Affiliation
VΔLK Research
Status
preprint
Year
2026
bibtex / cite
@article{prism2026wmav,
  title   = {PRISM: Predictive Recurrent Inference
             State Machine — A Verifiable World-Model
             Agent Architecture},
  author  = {[Anonymized for Review]},
  journal = {Preprint},
  year    = {2026},
  note    = {V{\Delta}LK Research,
             Institute for Complex Cognitive Systems}
}
When to adopt PRISM
  • long-horizon agents where memory must replay
  • multi-turn reasoning where drift is a known failure mode
  • regulated domains: finance, healthcare, law, audit
  • embodied / robotic systems with safety constraints
  • research benchmarks that score traceability
v1.0.0 sealed · reference release