Module Glossary¶
Page Maps¶
graph LR
family["Reproducible Research"]
program["Deep Dive DVC"]
section["Reproducibility Failures in Real Teams"]
page["Module Glossary"]
capstone["Capstone evidence"]
family --> program --> section --> page
page -.applies in.-> capstone
flowchart LR
result["trusted result"] --> influence["influential state"]
influence --> identity["recorded identity"]
identity --> reconstruction["clean reconstruction"]
reconstruction --> recovery["team recovery"]
recovery --> claim["bounded reproducibility claim"]
Use this glossary when an early workflow claim sounds plausible but its evidence boundary is unclear. The distinctions determine what a rerun, digest, Git revision, DVC record, or recovery test can actually support.
Claim and evidence terms¶
Claim¶
A statement someone asks another person to trust.
Example: “Another maintainer can reconstruct this risk summary.”
The claim should name result, boundary, and expected evidence.
Bounded claim¶
A claim limited to the conditions its evidence tests.
“Two runs in this prepared workspace produced equal JSON” is bounded. “The whole workflow is reproducible everywhere” is not.
Evidence¶
An observed fact relevant to a claim: file identity, command receipt, parameter value, runtime record, clean rebuild, or recovery result.
Evidence does not carry its conclusion automatically. Equal outputs are evidence of local repeatability, not scientific validity.
Receipt¶
A durable record connecting an action to identities and result. A command receipt may include invocation, status, runtime, inputs, and output digest.
A README command is an intended route. It becomes a receipt only when an execution is recorded.
Verified¶
Direct evidence supports the bounded claim.
Contradicted¶
Direct evidence shows the claim is false. A current input digest different from the recorded producing digest contradicts an identity match.
Unverified¶
Required evidence has not been collected. The same-workspace case has no clean rebuild receipt, so reconstruction is unverified.
Unavailable¶
Required state is identified but cannot be obtained through the promised boundary. This differs from an unknown identity.
Out of scope¶
The route does not promise to answer the question. Scientific threshold validity is out of scope for the workflow evidence audit.
Proof limit¶
The boundary beyond which evidence does not support inference.
Example: a Python version record does not prove every dependency or hardware influence.
Closure evidence¶
The observation that changes a finding's result. For a missing clean rebuild, closure requires the clean execution and matching verified output, not a prose promise to run it later.
Repeatability and reconstruction terms¶
Local repeatability¶
The same operator can execute a workflow again in substantially the same workspace and obtain the same bounded result.
It may inherit private data, environment, caches, and memory.
Team reproducibility¶
Another authorized maintainer can obtain recorded state, reconstruct the result outside the author's private workspace, and explain the match.
This is the operational definition used in this course; other disciplines may use the term differently.
Determinism¶
The property that equal effective inputs and execution assumptions produce equal output. Determinism helps repeatability but does not supply missing inputs or recovery.
Clean reconstruction¶
Execution from a boundary that excludes unreviewed private workspace state and uses recorded inputs, controls, implementation, and runtime evidence.
“Delete the output and rerun here” is not automatically clean.
Independent execution¶
A reconstruction performed by another actor or shared system at the promised team boundary. It tests transfer beyond the original operator.
Mechanical reproducibility¶
The ability to identify and reconstruct workflow state and result. It makes semantic and scientific review possible against stable evidence but does not approve those concerns.
Scientific validity¶
Whether the method, data, assumptions, and interpretation support the research claim. DVC evidence can stabilize the objects under review; it does not decide validity.
State and influence terms¶
Influential state¶
Any file, value, condition, decision, or prior state with the power to change the trusted result.
Hidden state¶
Influential state absent from the recorded contract. It may be stable and deterministic while still hidden.
Undeclared input¶
An influential file or value not represented in the workflow's input or control boundary.
Examples include a prepared file, environment threshold, optional lookup, or database snapshot.
Control input¶
A value that chooses workflow behavior: threshold, seed, feature list, profile, or parameter. Controls should have effective values and sources.
Effective value¶
The value computation actually used after defaults and overrides. The value in
params.json is not effective if an environment variable replaces it.
Manual transformation¶
A human-performed operation that changes state between source and result. Manual work can be valid; reproducibility requires its behavior, authority, and lineage to be represented.
Runtime influence¶
An execution-context property that can change the result: language, package, tool, locale, hardware, concurrency, or randomness.
External state¶
Influential data or behavior owned outside the repository, such as a database snapshot, API response, object version, or hosted model.
Identity and recovery terms¶
Path¶
A location or role name such as data/current.csv. A path can remain stable
while content changes.
Content identity¶
Evidence distinguishing exact bytes, commonly a cryptographic digest or content-addressed object identity.
Role¶
How the workflow uses state: raw input, parameter, derived output, metric, or promoted release.
Provenance¶
The recorded connection from a result to the implementation, inputs, controls, runtime evidence, and decisions that produced it.
Reconstruction¶
Executing recorded computation to produce a result.
Recovery¶
Obtaining required historical state after local absence or loss.
Reconstruction can fail with available data and an unknown runtime. Recovery can fail with a perfect workflow and unavailable input bytes.
Recoverability¶
The supported ability for an authorized maintainer to retrieve and verify protected state through the promised boundary.
Social memory¶
Workflow knowledge available mainly through people's recollection: the real command, correct extract, manual filter, or location of historical data.
An interview can discover social memory. A contract and receipt must replace it as the durable interface.
Git, scripts, and DVC terms¶
Source revision¶
A Git-identified state of tracked source, configuration, documentation, and DVC metadata.
It does not automatically identify external artifact bytes unless the recorded metadata connects them.
Script behavior¶
The computation implemented by source code. A script can expose logic without declaring all input roles, controls, output contracts, or execution evidence.
Intended route¶
A documented command sequence describing how a workflow should run. It is not evidence that a historical result used that route.
Artifact boundary¶
The identity, storage, recovery, and retention boundary for data and derived files that do not belong in ordinary source history.
DVC metadata¶
Lightweight repository state connecting paths, pipeline declarations, or recorded execution to artifact identities. Metadata without reachable objects supports diagnosis but not materialization.
DVC remote¶
A configured content storage boundary used for transfer and recovery. Its existence does not by itself define access, retention, or recovery objectives.
Pipeline declaration¶
A reviewed statement of commands, dependencies, parameters, and outputs. Declarations can be incomplete; controlled invalidation tests challenge their truthfulness.
Recorded execution state¶
Observed input and output identities associated with a pipeline execution, such as DVC lock state. It records what DVC observed under the declaration, not every possible hidden influence.
Adoption contract¶
The narrow claim a team chooses DVC to support first, with evidence, acceptance route, adversarial test, owners, and proof limits.
Workflow inventory terms¶
Trusted result¶
An output used in a decision, review, publication, or downstream contract. Trust role determines which evidence deserves protection.
Evidence inventory¶
A present-state record of source inputs, controls, transformations, runtime, outputs, reconstruction, recovery, owners, and unknowns.
Evidence gap¶
A required identity, declaration, receipt, recovery route, or owner missing from the current claim.
Repair boundary¶
The smallest complete change that restores one affected contract.
Example: declare and test a threshold control rather than redesigning the whole pipeline.
Responsibility map¶
A record assigning one authority to each fact: Git source history, DVC artifact identity, environment runtime, storage retention, domain validity, and governance approval.
Common distinctions¶
| Do not collapse | Distinction |
|---|---|
| local rerun and clean reconstruction | one inherits workspace; one starts from recorded boundary |
| repeatability and team reproducibility | one proves stability here; one proves transfer |
| determinism and reproducibility | stable behavior does not supply missing state |
| path and content identity | location can stay while bytes change |
| identity and recovery | knowing which bytes does not make them obtainable |
| script and workflow contract | behavior is not the full declared interface |
| README command and execution receipt | intention differs from observation |
| Git revision and result provenance | source state needs an explicit result connection |
| DVC metadata and remote object | reference differs from available content |
| clean rebuild and scientific validity | reconstruction does not approve method |
| unverified and contradicted | missing observation differs from observed mismatch |
| mechanical and semantic trust | stable state does not define correct meaning |
When a sentence overreaches, rewrite it with both sides of the relevant distinction.