My first paper draft. I ask for an discussion and evaluation

Learning on an Experience Clock: A General Method for Repeatable and Paired Evaluation of Stochastic Learners, Demonstrated on a Biologically Inspired Text Learner

„…Draft of 17 November 2026. Preprint draft, not peer reviewed.

Abstract

Problem. Stochastic learners are hard to evaluate. Random replay, sampling and state-dependent thresholds make two runs of the same code end differently. Decay and recency measured in wall-clock seconds make learning depend on the speed of the machine. A code change then cannot be separated from chance.

Method. We propose a general method in three parts:

  1. An experience clock. Learning time advances by one tick per learning cycle and is stored with the learner’s state. This follows evidence that forgetting depends mainly on intervening experience and that the brain builds time from events.
  2. Deterministic noise. Every random stream is seeded from run seed, cycle, component and purpose, in the spirit of counter-based generators. Iteration and ties are totally ordered, and each cycle is atomic. A run then depends only on its start state, its code and its run seed.
  3. Seeds as replicate individuals. Variants are compared on common seeds by paired differences against a non-inferiority margin fixed in advance.

We state the requirements as a checklist with a test for each, so the method can be applied to other learners.

Demonstration. We apply the method to a hard case: a biologically inspired text learner with sleep replay and consolidation.

  • Before the change, two runs from identical copies differed in 101 of 156 database tables. Afterwards, all 157 tables agreed, including every timestamp, across hash salts, start times and a restart.
  • Two 1,000-cycle runs that took 1.5 h and 5.8 h agreed in all 24,000 logged values.
  • The spread between seed replicates was at most 1.6 %.

Statistical check. With five seeds, the percentile bootstrap used for the paired interval covers only 84 % instead of 95 %. It passes a variant that is exactly at the margin in 7.9 % instead of 2.5 % of cases. The paired t interval holds the nominal rate and is recommended.

Scope. The evidence shows repeatability on one machine and one software stack. It does not show portability.

Keywords: reproducibility, determinism, experience time, counter-based random numbers, common random numbers, non-inferiority, seed variance, stochastic learners, biologically inspired learning


1 Introduction

1.1 The problem

Many learning systems are stochastic by design. Biologically inspired ones borrow their noise from the brain:

  • replay picks memories at random;
  • consolidation samples among candidates;
  • thresholds react to the system’s own state.

This variability is part of the model, but it makes evaluation hard. If two runs of the same code end differently, a single comparison between an old and a new variant mixes the effect of the change with the effect of the run.

Deep learning shows the same problem. Identical training runs differ because of seeds, tooling and hardware (Pham et al. 2020; Summers & Dinneen 2021; Zhuang et al. 2022), and reported results often rest on too few runs (Henderson et al. 2018; Bouthillier et al. 2021).

The usual remedies fit poorly:

  • Fixing one global seed hides the variance instead of measuring it, and it breaks as soon as any component draws one random number more than before.
  • Record and replay (LeBlanc & Mellor-Crummey 1987) repeats one run but does not make two code variants comparable.
  • Many independent runs are expensive when a run takes hours, and they still confound the change with the run.

Learners that model memory have a further, less obvious source of irreproducibility: time. A system that decays, orders or windows its memories by wall-clock seconds learns differently on a fast machine and on a slow one.

1.2 The method

We treat the timing and the noise of a learner as part of its stored state (Section 3):

  • An experience clock replaces wall-clock time inside learning. It is grounded in research on forgetting and on the neural representation of time (Section 2.1).
  • Counter-based random streams, total ordering and atomic cycles make a run a function of its start state, code and run seed.
  • Run seeds become replicate individuals. Variants are compared on common seeds against a margin fixed in advance.

Section 3.7 turns these requirements into a checklist with a test for each.

1.3 What becomes possible

Once identical code gives identical results, four kinds of measurement become possible:

  • a refactoring can be tested for exact equality with the old code;
  • the variance between replicates can be measured exactly, and a comparison margin can be derived from it;
  • a property that holds in every replicate can be told apart from an accident of a single run;
  • an experiment can branch from one exact state into a perturbation arm, a sham arm and a control.

1.4 Demonstration and contributions

We demonstrate the method on BrainStem, a neuro-symbolic text learner. BrainStem starts without word lists, grammar or pretrained knowledge. It reads raw German Wikipedia text and runs a chain of 36 components per learning cycle, which include sleep replay, consolidation and neuromodulator-like control variables. This makes it a hard test case: noise enters in many places, and many quantities depend on time. The system is at an early developmental stage and does not understand language. It is the test bed of this paper, not its claim.

The contributions are:

  1. the method and its checklist (Section 3);
  2. evidence in the demonstration for repeatability, restart equivalence, independence from execution speed and genuine variation between replicates (Section 5);
  3. a paired benchmark protocol with a five-seed baseline (Section 5.5);
  4. an analysis of the protocol’s operating characteristics, including a weakness in its interval estimate and a remedy (Section 6).

All numbers come from the project’s reports and raw data, which are named next to each result. The raw data ship with the paper in data/. A script that needs only the Python standard library rebuilds Figures 2 to 7 and the simulation from them (figures/make_figures.py)…”

Complete document: