Sequential decision-making
Premature Epistemic Action
When useful information changes future dynamics or feasible actions, the correct policy may be to acquire it later rather than now.
Independent researcher · technical builder
Saurav Sharma
Independent Researcher & Technical BuilderDecision systems for worlds where the rules are not fully visible.
I study reliable action under hidden mechanisms, consequential information acquisition, binding constraints, and rare failures—especially in economic and financial environments.
My work combines controlled experiments, exact planning where possible, adversarial baselines, frozen evaluations, and explicit negative results. A broader private program, Autonomous Economic Intelligence (AEI), connects these lines.
- AffiliationIndependent Researcher
- EducationBS student, IIT Madras
- BackgroundFinancial markets & trading
- BasedIndia
Fig. 0 — Prediction vs. identificationSchematic
01 — Research program
One recurring question, attacked from several directions.
The work asks when a capable system can still make the wrong decision because it has not identified the governing mechanism, has treated information as passive, or has ignored the constraint or tail regime that actually determines feasibility.
Financial AI · market microstructure
When Information Acquisition Moves the Market
Passive value-of-information can select the wrong inquiry policy when asking for information changes downstream execution values.
Financial AI · private credit
When Model Error Moves the Boundary
Model error becomes decision-critical when it moves a financial feasibility boundary reached by the optimized policy.
02 — Private research program
Autonomous Economic Intelligence
A broader architecture program for decisions under hidden rules, rare catastrophes, consequential evidence acquisition, and authoritative constraints. REC, microstructure, and private credit are standalone research probes of different pieces of the broader problem.
- evidence
- governing structure
- support
- stress
- planning
- verified execution
The public research boundary, summarized as a broad conceptual pipeline.
Public
What is shared
- Problem framing
- High-level research thesis
- Selected internally labeled evidence
- Evidence limitations
Private · not disclosed
What remains private
- Exact identification machinery
- Acquisition / probing policies
- Architecture-selection details
- Successor design
- Reconstruction-grade source and full evaluation suite
03 — Selected evidence
Results are presented with their failure boundaries.
Numbers below are mechanism-test results from synthetic or stylized environments, not claims of broad real-world prevalence or general intelligence.
Frozen diagnostic · REC
15/15
REC exact action match on the frozen diagnostic
Fresh 5,000-world IID holdout · RFQ
83.2%
inquiry-breadth policy mismatch on the fresh 5,000-world holdout; stylized exact-model experiments, not an estimate of real-market prevalence
Boundary-moving stress · credit
14/20
private-credit nominal-plan failures under the boundary-moving stress
Retained negative result
27/27
generic exact planner success retained as a negative result
Internally reproducible is not independently reproduced. Synthetic mechanism tests are not prevalence estimates.
Methods
How I work.
M.1
Mechanism isolation
Small controlled environments are used to separate a proposed failure mode from generic model weakness.
M.2
Prospective testing
When feasible, worlds, seed rules, criteria, and analysis are frozen before confirmatory evaluation.
M.3
Adversarial self-review
Oracle contamination, generic-planner attacks, non-identifiability, failed controls, and null results are retained.
04 — Background
Background.
Research affiliation: Independent Researcher.
Research affiliation
Independent Researcher
The research on this site is conducted independently.
Education
IIT Madras
BS student, Indian Institute of Technology Madras.
Background
Financial markets
Experience studying and trading financial markets shaped my interest in execution, liquidity, credit constraints, tail risk, and decision-making under uncertainty.
05 — Writing
Research notes.
Notes on the reasoning behind the research program.
Research agenda · Research note · 2026
Why Prediction Is Not Enough for Decisions Under Hidden Rules
A compact account of the recurring distinction between prediction, mechanism identification, consequential information acquisition, and reliable action.
06 — Current priority
Current priority
I am focused on externally specified evaluation, independent reproduction, and harder tests of decision systems operating under hidden mechanisms, consequential information acquisition, rare failures, and binding constraints.
Research discussion · Reproduction · Collaboration