Diop Daily #108 — September 2026

The Experiment Must Leave a Trace

Research becomes faster when a machine can write code, run an analysis, inspect a result, and propose the next test. The speed is easy to see. The institutional consequence is harder: a laboratory can accumulate more experiments than its people can remember. If each branch disappears into a terminal, a notebook, or a model conversation, the system gains velocity while losing the ability to know why it believes anything.

An OpenAI News RSS item dated September 6 describes coding agents reshaping AI research and points to early data on agent usage, experiment velocity, task complexity, and research acceleration. The description comes from the company’s internal view, so it does not establish a universal productivity measurement; it remains useful because it names the transition directly: agents are moving closer to the machinery of inquiry. Once they write and execute more of the work, the scarce layer is the record that allows a research institution to distinguish a promising result from a fast repetition of an old mistake.

A laboratory does not become more intelligent when it runs more experiments. It becomes more intelligent when each experiment changes what the laboratory can remember and test next.

Speed without memory is only acceleration

Scientific work has always depended on records. A result without method cannot be reproduced. A method without conditions cannot be interpreted. A failed trial that is not recorded will return as a new hypothesis under a different name. The research notebook is therefore more than a place to store observations. It is a device for controlling the relationship between evidence and belief.

Agentic systems put pressure on that relationship because they can produce many plausible branches in a short period. An agent may alter a data-cleaning step, change a model route, run a new analysis, inspect the output, and generate another modification before a human has read the first diff. The central problem concerns both the correctness of the final result and whether anyone can reconstruct the path that made it appear credible.

A useful experiment ledger treats every run as a bounded claim. It records what the researcher expected, what changed, what was held constant, what the system observed, and what decision followed. This structure gives negative results a place beside successful ones. It also prevents a polished final artifact from hiding the branches that were abandoned along the way.

An experiment is a chain, not a result

The final chart or paragraph is only one state in a longer chain. A research-grade record should connect the question to the conditions under which an answer was produced. At minimum, each experiment needs:

  • Hypothesis: the claim being tested, including the expected direction and the reason it deserves a test.
  • Inputs: the dataset, documents, code, prompts, instruments, permissions, and prior findings available at the start.
  • Intervention: the exact change introduced, whether it was a new model route, parameter, transformation, tool, or human instruction.
  • Controls: the baselines, comparison groups, fixed conditions, and evaluation measures that make the result interpretable.
  • Observations: outputs, errors, anomalies, resource use, and uncertainty, including results that do not support the hypothesis.
  • Decision: the next state of the work: continue, revise, suspend, publish, reject, or repeat under a new condition.

These fields make the difference between an experiment and an execution trace. A trace says what the machine did. A ledger says what the action was supposed to establish and what the institution learned from the outcome. The distinction is essential when a research system uses several agents, because each handoff can otherwise detach an action from the question that justified it.

Google’s public description of ADK Go 2.0 supplies a relevant runtime signal. Its graph-based workflow engine includes human-in-the-loop orchestration, dynamic routing, retries, and built-in resilience. Those features can help a laboratory coordinate search, code execution, review, and publication, while scientific memory requires a ledger across the graph and preserve the relationship between a node’s action, the evidence it consumed, and the decision that came after it.

Negative results are part of the institution

Organizations often preserve success more carefully than failure. Published findings receive a title, a citation, and a location. Rejected hypotheses remain in private notes, deleted branches, or the memory of one researcher. This creates a costly asymmetry. The institution remembers what it chose to show and forgets what it already learned not to pursue.

Agentic research will intensify this problem unless the system gives failed branches a durable identity. A null result can reveal that a method is insensitive to a local condition. A failed data transformation can expose a hidden schema assumption. A discarded model route can show that a language, instrument, or population was poorly represented. Such findings are constraints on the next experiment, not editorial waste.

The ledger should therefore preserve rejected branches without confusing them with validated conclusions. A researcher must be able to ask which hypotheses were tested, under which conditions, with what evidence, and why the branch was closed. The answer should include uncertainty and ownership. A future agent can then avoid repeating the test, or reopen it when the data, method, or question has materially changed.

NIST’s AI Risk Management Framework gives this discipline a general vocabulary through Govern, Map, Measure, and Manage. Applied to experiment ledgers, Govern assigns ownership of the question and the record. Map identifies the people, data, systems, and obligations touched by the run. Measure records reproducibility, error, cost, and correction. Manage determines what happens when the result conflicts with prior evidence, the method changes, or a permission expires. The framework is a useful test rather than a laboratory notebook specification: it asks whether the research system can account for its own decisions.

Agents change the rate of inquiry

When a human team runs a small number of experiments, informal memory can appear sufficient. The team meets, compares notes, and carries a shared sense of what has already been tried. That arrangement weakens as agents increase the number of branches. More execution creates more state. More state creates more opportunities for duplicated work, contaminated comparisons, and forgotten assumptions.

The correct response is to make the narration part of the system instead of slowing every experiment until a person has manually narrated it. Agents should be required to declare the hypothesis they are testing, point to the baseline, attach the changed inputs, and record the reason for the next action. Human reviewers can then spend their attention on the meaning of the result rather than reconstructing basic chronology.

This also changes the meaning of research throughput. Counting runs is a weak measure. A better measure asks how many experiments become usable knowledge: how many can be reproduced, how many close a question, how many prevent duplicated work, how quickly a contradiction is surfaced, and how much of the record a new researcher can inherit without interviewing the original operator.

African laboratories must own the record of inquiry

Scientific sovereignty includes control over the memory of investigation. A laboratory that can access a foreign model but cannot export its hypotheses, failed branches, local terminology, evaluation sets, and reasons for rejecting a result has borrowed capability without securing knowledge.

This matters in African research environments where records may cross languages, institutions, funding cycles, and uneven technical infrastructure. A field study may combine community testimony, mobile data, paper forms, satellite imagery, and laboratory measurements. A model may process several of those sources while understanding only one language well. The experiment ledger must retain the original terms, the translation choices, the permissions, and the limits of the inference.

Local ownership means keeping the institutional question and the evidence chain under local authority while using external instruments where they help. An African research center should be able to move a ledger between model providers, audit it with its own reviewers, reproduce a result on available infrastructure, and decide which conclusions enter the public record. The machine can accelerate the inquiry while the institution retains the power to define what counts as knowledge.

Cheikh Anta Diop’s scientific method has a direct contemporary lesson here. Historical and scientific independence require more than access to instruments produced elsewhere. They require the capacity to compare evidence, expose assumptions, preserve the record, and build conclusions that can be tested by others. An experiment ledger is a small institution for that work. It gives a laboratory a memory that is structured enough to resist both erasure and false certainty.

Where the investable surface is widening

If coding agents increase the rate of inquiry, capital should examine the infrastructure that turns experimental activity into inheritable research capacity:

  • Experiment registries: systems that assign identity, version, ownership, status, and searchable relationships to hypotheses and runs.
  • Reproducibility runners: environments that snapshot code, data, model routes, dependencies, and permissions, then replay a result under declared conditions.
  • Branch and null-result management: tools that preserve rejected hypotheses and failed methods as reusable constraints without presenting them as validated findings.
  • Agent research control planes: orchestration layers that require hypothesis declarations, baseline references, approval gates, resource limits, and evidence-bearing handoffs.
  • Local scientific memory: multilingual, exportable systems that retain terminology, community permissions, evaluation sets, and institutional ownership across providers.

The underwriting question is concrete: does the product increase the amount of research that a new person, agent, or institution can reproduce and extend? Useful measures include reconstruction time, duplicate-experiment rate, percentage of runs with a declared baseline, share of negative results that remain discoverable, time to identify a contradictory finding, and the fraction of records that can be exported and replayed outside the original provider.

This is a different commercial object from a generic research assistant, a finished-artifact registry, or a workflow graph. Those products help a system answer, preserve an output, or route work. The experiment ledger preserves the institution’s changing map of what it tried, what failed, what remains uncertain, and why the next question is worth asking.

Build the ledger before adding another agent

A laboratory can test this architecture with one bounded research question. Before increasing the number of agents, make the inquiry legible:

  1. Write the hypothesis, expected observation, decision threshold, and owner before the run begins.
  2. Snapshot the inputs, code, model route, tools, permissions, and baseline used by the experiment.
  3. Require the system to record interventions, outputs, errors, resource use, and uncertainty as the run proceeds.
  4. Store successful, inconclusive, and rejected branches with distinct states and reasons.
  5. Reproduce one result on a second environment or provider, then record the differences rather than hiding them.

These practices make speed accountable without making inquiry timid. They let agents explore more branches while keeping a human institution able to inspect the question, challenge the method, and inherit the conclusion. They also create a better boundary for capital: the buyer can evaluate whether the system produces durable research capacity instead of merely a larger stream of machine activity.

The experiment must leave a trace because a laboratory is built from remembered questions as much as from published answers. When every branch has a place, failure becomes evidence, and the next generation can begin from the real frontier rather than from the fragments that survived.

Sources