entroly

PRISM-R: Reversible Neural Evidence Compression

Status: experimental research prototype. It is not enabled in Entroly’s default compression path and is not yet evidence for a general quality or production-cost claim.

Research question

Can a model-agnostic agent context layer use a local transformer to improve query-conditioned evidence selection while preserving readable source text, exact recovery, deterministic fallback, and auditable omission risk under future-query shift?

The current answer is promising but incomplete. A frozen SQuAD v2 pilot passes the repository’s evidence-retention gate at tighter budgets, but downstream LLM answers, long-agent trajectories, domain shift, and stronger learned-compressor baselines remain untested.

Scoped 2021-2026 prior-art synthesis

This is a focused map of directly relevant primary work, not a claim to have read every paper published in five years.

Direction	Representative primary work	Design constraint for Entroly
Learned latent prompts	Gisting, NeurIPS 2023	Latent compression can be efficient but is model-coupled and opaque.
Learned summary vectors	AutoCompressors, EMNLP 2023	Precomputed soft memories are useful, but exact textual provenance is not inherent.
Dynamic learned units	Nugget, ICML 2023	Variable-rate neural selection is established prior art.
Query-aware hard compression	LongLLMLingua	Query conditioning and position control matter; lexical evidence recovery remains essential.
Bidirectional token classification	LLMLingua-2	A learned compression objective is stronger than generic embedding similarity.
Selective augmentation	RECOMP, ICLR 2024	Abstaining from irrelevant augmentation is established; Entroly extends abstention to unsafe compression.
Memory-slot autoencoding	ICAE, ICLR 2024	Soft memory slots can compress aggressively but require a compatible decoder.
One-token RAG fusion	xRAG, NeurIPS 2024	Extreme latent compression is possible but is retriever/model specific.
Rate-distortion limits	Fundamental Limits of Prompt Compression	Query-aware variable-rate policies are theoretically necessary; token reduction alone is not the objective.
Certified retrieval risk	C-RAG, ICML 2024	Finite-sample risk claims require explicit assumptions and calibration data.
Evaluator-head salience	EHPC, NeurIPS 2025	Early transformer attention can identify useful tokens without a separate large compressor.
Ultra-long chunk compression	ParallelComp, ICML 2025	Chunk-level coherence and long-context memory limits must be evaluated, not inferred from short QA.
Semantic chunk integrity	ChunkKV, NeurIPS 2025	Token-level pruning can fragment meaning; source-addressable spans are safer units.
Real latency break-even	Prompt Compression in the Wild	Compression overhead can erase inference gains; production claims require hardware-specific end-to-end measurement.
Multiplex graph pruning	RAGP, 2026 preprint	Distributed evidence and bridge relations motivate graph-aware future evaluation.

Hybrid hard/soft compression, learned selection, submodular diversity, abstention, and provenance each have prior art. The plausible contribution is their formal and operational coupling under recoverable agent-memory shift, not any component in isolation.

Objective

For normalized semantic vectors e_i, define:

• u_i(q): current-query transformer relevance;
• a_i: exact-evidence value supplied by deterministic detectors or policy;
• h_i: future-query utility learned from agent access history;
• s_i: confidence that the semantic representation is reliable;
• c_i: token cost.

PRISM-R approximately maximizes

F_q(S) = sum_{i in S} [u_i(q) + gamma*a_i + rho*h_i]
         + lambda log det(I + sum_{i in S} w_i(q) e_i e_i^T)

w_i(q) = s_i * max(0, u_i(q) - tau)^p

subject to:

L_q subset S
sum_{i in S} c_i <= B

L_q contains mandatory evidence locks. Separating modular utility from the determinant weight avoids counting relevance twice. The floor tau prevents an irrelevant but orthogonal span from winning on novelty. With non-negative weights, the static core remains monotone submodular; the implementation compares gain-greedy and density-greedy schedules under the token knapsack.

Acceptance and recovery contract

1. Load only an explicitly supplied local transformer directory. Remote model identifiers are rejected.
2. Normalize every embedding and reject zero, malformed, or non-finite output.
3. Preserve mandatory exact evidence even when that exceeds the requested budget; expose the budget failure instead of deleting evidence.
4. When lexical and neural champions disagree, either retain both or require a threshold tied to a named calibration artifact. Otherwise abstain.
5. Persist every omitted character span atomically with source offsets and a content hash. A single stale offset or hash aborts the write.
6. Rehydrate exact source text by receipt/span ID and debit the recovered tokens from realized savings. Repeated retrieval IDs are idempotent.

Measured results

Frozen answer-passage retrieval

On 600 deterministic SQuAD v2 trials with 15 distractor paragraphs, calibrated on 300 trials and evaluated on a disjoint 300:

Selector	Held-out top-1 evidence recall	Notes
Deterministic BM25	99.0%	Exact lexical evidence remains the strongest primary scorer.
Local all-MiniLM-L6-v2 transformer	97.7%	Frozen revision c9745ed1d9f207416be6d2e6f8de32d1f16199bf; one transformer-only success, five lexical-only successes; no neural-primary claim.
Dual-channel disagreement guard	99.3%	Selects 1.02 of 16 passages on average, a 15.69x passage compression ratio.

The generic transformer failed the primary-scorer gate. PRISM-R therefore uses it as a residual semantic channel rather than replacing lexical evidence.

Artifacts:

• neural-evidence-frontier.md
• neural_evidence_frontier.json
• neural_evidence_frontier.md

Future-query shift and exact rehydration

Each of 200 trials contains two SQuAD questions whose answers occur in different sentences of the same paragraph. The context is compressed for q1; q2 is revealed only afterward. Results use the same frozen trials and local model at every ratio.

Nominal active ratio	Actual active ratio	Lexical q1	PRISM-R q1	Direct unseen q2	q2 after exact recovery	Active + recovery
50%	44.2%	92.0%	95.0%	34.0%	95.0%	72.7%
25%	24.1%	60.5%	87.0%	9.0%	90.5%	50.6%
12.5%	18.2%	24.0%	81.0%	3.5%	87.0%	43.5%

The 25% and 12.5% settings pass the pilot’s paired current-evidence and rehydration gates. The 50% current-query delta does not (p=0.109375) and is not presented as a win. Actual ratios differ from targets because sentence spans are indivisible and safety guards may retain both channels.

Artifacts:

• neural_query_shift_2x.json
• neural_query_shift.json
• neural_query_shift_8x.json

What remains before a paper or public breakthrough claim

• Replace the generic MiniLM scorer with a trained evidence-necessity model and compare against LLMLingua-2, RECOMP-style extraction, and other current extractive compressors.
• Evaluate NoLiMa-style lexical-gap retrieval, RULER multi-needle and tracing, multi-hop bridge evidence, LoCoMo-scale conversations, and repeated compaction trajectories.
• Measure downstream answers, unsupported claims, exact locked facts, calibration under domain/model/compression shift, recovery latency, and net tokens after repeated rehydration.
• Run multiple target LLMs and hardware classes. The committed pilots do not establish production latency or billing savings.
• Audit emerging work such as EXIT, IterCOMP, SARA, agent-memory action policies, and recent provenance graphs against exact papers and released code before making a novelty statement.

The defensible current statement is: PRISM-R is an implemented, locally tested research prototype for relevance-gated neural selection with deterministic evidence fallback and exact future-query rehydration. It has passed a bounded evidence-retention pilot, not the full breakthrough gate.

This site is open source. Improve this page.