Saurav SharmaIndependent Researcher
R.01 · Sequential decision-making

Premature Epistemic Action

Information can be valuable and still be harmful to acquire now when acquisition changes the future decision problem.

Evidence / ArtifactLink pendingCode / ReproduceLink pending
Status
Research manuscript · 2026
Area
Sequential decision-making
Evidence
  • frozen diagnostic
  • ablations
  • offline artifact
  • negative result
Artifact
Paper-scoped offline artifact
Fig. 1 — Whether versus when to acquire informationSchematic
Schematic: probing early versus probing at the boundaryTwo lanes of decision states. In lane A information is acquired early and the consequence changes all later states. In lane B acquisition is delayed until the probe boundary.PROBE BOUNDARYA · PROBE EARLYB · PROBE AT THE BOUNDARYacquire nowconsequence propagates: dynamics, constraints, costs, feasible actions changeacquire at boundarySTATE 1STATE 7 · TIME →
SchematicAcquiring information can itself change future task dynamics, constraints, costs, or feasible actions. The decision is therefore not only whether to acquire information, but when. Illustrative, not experimental data.
01 · Question

Whether, and when.

Most information-acquisition models ask whether information is worth its informational value and cost. This project studies a different case: acquiring information can itself change future task dynamics, constraints, costs, or feasible actions. The decision is therefore not only whether to acquire information, but when.

02 · Selected results

Frozen diagnostic and ablations.

Frozen diagnostic · full REC
15/15

full REC exact action match on the frozen diagnostic

Ablation · prospective timing removed
0/6

pre-boundary accuracy after removing prospective timing

Ablation · consequence modeling removed
0/6

pre-boundary accuracy after removing consequence modeling

Generic exact depth-3 planning
27/27

small worlds solved by generic exact depth-3 planning

03 · Experimental design

Core diagnostic.

The frozen diagnostic contains 15 decision states spanning pre-boundary states, true probe-boundary states, and high-consequence no-probe controls. The tested frontier model remains correct on the actual probe boundary and high-consequence controls while repeatedly probing too early before the boundary.

04 · Strongest attack · negative result

Why the negative result matters.

Exact depth-3 expectimax solves all 27 tested small worlds. I therefore do not claim REC as a new generic planning primitive.

Retained explicitly in the reproduction package. No generic-planner novelty claim.
05 · What the evidence covers

A narrower contribution.

The paper's narrower contribution is the explicit immediate-versus-delayed epistemic-action interface, propagation of endogenous consequences, and selective verification.

Evidence: frozen 15-state diagnostic; full REC 15/15; principal timing/consequence ablations collapse pre-boundary performance.

06 · What it does not establish

What remains open.

Boundary

Independent external reproduction, cross-model frontier replication, real transaction-level validation, and broad non-economic generalization are not yet established.

07 · Evidence & artifact status

Reproducibility.

The paper-scoped offline artifact includes frozen worlds and queries, evaluator gold, prompt constructors and hashes, raw frontier responses, scorers, the REC reference implementation, principal ablations, support checks, deterministic holdout reproduction, the generic-planning negative result, SHA-256 manifests, and a top-level verification script.

python verify_all.py
Internally reproducible

Offline paper-scoped artifact with raw outputs, hashes, evaluator, reference implementation, and verification scripts.

Independently reproduced

Not yet established.

Status

Research manuscript · 2026.

Evidence / ArtifactLink pendingCode / ReproduceLink pending