Scientific Dependency Graphs: Who Actually Enabled a Discovery?

Getting your Trinity Audio player ready...

A scientific discovery can depend on a theorem, a dataset, an instrument, a software library, and years of maintenance by people absent from the final paper’s author list. Scientific dependency graphs map these enabling relationships. They can reveal overlooked contributions, but they cannot, by themselves, determine who caused a discovery or how its rewards should be divided.

That distinction matters for scientific recognition and for funding systems such as AI Internet-Meritocracy (AIIM). Identifying what researchers used is an evidence problem. Deciding how much credit each contributor deserves also requires judgments about alternatives, importance, and fairness.

What is a scientific dependency graph?

A scientific dependency graph is a network whose nodes represent research outputs, resources, or activities, and whose directed connections describe how one supported another.

Nodes can include papers, individual results, datasets, software versions, experimental protocols, instruments, and validation activities. People and organizations can be linked to the contributions they made.

For example, a graph might record that an analysis used a particular dataset and software release, and that its interpretation relied on an earlier theorem.

The W3C PROV framework supplies a foundation for representing provenance: information about the entities, activities, and agents involved in producing something. A scientific credit system can build on that foundation, but provenance records do not automatically establish importance or entitlement to payment.

Citation graphs and dependency graphs answer different questions

A citation records a reference. The reference might acknowledge a method, provide background, identify a competing explanation, or criticize an earlier result.

A dependency claim should be more specific:

Relationship	What it establishes	What it does not establish
Paper B cites Paper A	B explicitly references A	B needed A to reach its result
Analysis B uses Dataset A	A supplied data used in B	No other dataset could have worked
Software B imports Library A	A is a declared software dependency	Every component of A affected the result
Proof B invokes Lemma A	The recorded proof relies on A	No alternative proof exists
Study C independently replicates B	C provides additional evidence about B	C enabled B’s original discovery

A useful dependency graph records the relationship and its evidence, rather than treating every connection as an equal unit of credit.

How a discovery can depend on several kinds of work

Consider a hypothetical experimental discovery:

flowchart TD
    A["Instrument calibration"] --> B["Measurements"]
    B --> D["Scientific analysis"]
    C["Software release"] --> D
    E["Earlier theorem"] --> D
    D --> F["Reported discovery"]
    F --> G["Independent replication"]
    H["Independent dataset"] --> G

Arrows point from an input to the activity or result that uses it. This is an illustrative workflow, not a complete causal model.

The calibration specialist, software maintainer, and theorem author contribute through different paths. Independent replication adds support afterward; it should not be confused with producing the original result.

The graph makes those distinctions visible. A single author list or citation count cannot express them all.

A real example: the infrastructure behind AlphaFold

The 2021 AlphaFold paper documents a protein structure prediction system whose development drew on existing sequence and structure resources, including the Protein Data Bank. Its methods describe training data and computational components, while its acknowledgments recognize the researchers who contributed experimental structures. See Jumper and colleagues’ original AlphaFold paper.

Viewed through a dependency graph, this achievement connects model development with experimental data production, database curation, and computational tools.

That observation does not establish a percentage of the achievement attributable to each group. It establishes why evaluating only the final model or paper leaves out part of the documented enabling work.

A graph can reveal the contributors whom an evaluation should consider before deciding how to value their roles.

Why dependence does not settle causation

Three questions need separate answers:

1. Actual use: Did this research use the contribution?
2. Counterfactual importance: Without it, would the discovery have been delayed, more expensive, less reliable, or impossible?
3. Reward policy: How should that contribution affect recognition or funding?

A researcher may use one software package when an adequate substitute exists. Its contribution remains real, even if it was not indispensable.

Conversely, a small correction might prevent an entire experiment from producing a false conclusion. Its importance could far exceed its publication count.

Joint necessity does not produce additive percentages

Suppose a hypothetical discovery requires both a dataset and an algorithm. Removing either prevents the result.

Each can then be described as necessary under that scenario. But assigning each 100% of a fixed reward pool is impossible. Their contributions interact; they are not independent pieces of a pie.

A graph exposes the joint dependence. An allocation rule must decide how to divide the reward.

Historical dependence differs from logical dependence

A mathematician may discover a proof after reading another author’s work, yet later present a proof that does not invoke it. The intellectual influence is real but may be absent from the formal dependency structure.

The reverse can also occur: a proof invokes a convenient lemma even though a short alternative argument was available.

The route by which a discovery happened, the resources needed to reproduce it, and the alternatives that could have produced it are different objects of evaluation.

What evidence should each connection carry?

A practical dependency record should identify:

• The exact contribution: a result, dataset version, software release, or documented activity.
• The relationship: used, derived from, implemented with, corrected by, or independently validated by.
• Supporting evidence: a methods passage, execution record, proof reference, repository history, or contributor statement.
• Provenance and uncertainty: who asserted the connection, who reviewed it, and whether it remains disputed.

An installed software package is weaker evidence of use than a record showing that its relevant function executed. An author’s account of intellectual influence can be valuable without being independently verifiable.

Contributor descriptions can use NISO’s CRediT taxonomy, which defines 14 roles, including conceptualization, data curation, software, and validation. CRediT helps describe who did what; it does not specify how much money each role deserves.

Crucially, missing documentation should mean “unknown,” not automatically “no contribution.” Otherwise, a system risks favoring contributors with better records over those whose work was poorly documented.

How dependency graphs could improve AIIM

AI Internet-Meritocracy is Science DAO’s experimental approach to funding researchers and open-source developers through AI-assisted assessment of documented contributions.

For AIIM, dependency graphs could provide an inspectable evidence layer: reviewers could examine which upstream contributions an assessment recognizes and why. The following are design proposals, not claims that a complete, validated dependency-based allocation system is already operating.

Separate evidence from valuation. First establish a supported connection. Then assess its significance. Finally apply an explicit funding policy.

Recognize upstream work. Evaluation should consider reusable methods, maintained software, and curated data alongside the downstream outputs they support.

Control repeated counting. A contribution may appear through several paths. A funding model should not multiply its reward merely because the same dependence has several representations.

Make disputes correctable. Contributors should be able to challenge missing links, mistaken attribution, and exaggerated claims. AI-generated connections need source evidence and review.

Test resistance to manipulation. Adding irrelevant citations, splitting one contribution into many records, or creating circular endorsements should not reliably increase rewards.

A credible evaluation would ask whether the graph helps independent reviewers identify relevant contributors—and whether allocations remain reasonably stable when uncertain connections change.

Who actually enabled a discovery?

The answer is usually a network of people whose contributions played different roles.

Scientific dependency graphs can make that network inspectable. They can show which evidence was used, which tools performed the work, and which earlier results supported the reasoning. Their strongest contribution is to make credit decisions better informed and easier to challenge.

Funding still requires a purpose. Rewarding past achievement, sustaining essential infrastructure, and supporting future discoveries are related but distinct goals. A transparent system should explain which goal governs its payments.

For AIIM and other research funding models, the opportunity is to recognize enabling work that conventional records can obscure—while keeping every claimed dependency, valuation, and allocation open to scrutiny.