Source-to-Score Lineage Graph¶
Software / tool — instantiates Abstraction–Substrate Traceability Guardrail
Visualizes lineage from substrate records through transformations to the final score, label, dashboard value, or decision artifact.
A Source-to-Score Lineage Graph is an interactive tool that renders the computation behind a single abstraction as a navigable directed graph — substrate records at the leaves, transformations as edges, the published score at the root — so a user can start from one output and walk backward to everything that made it. Its defining idea is traversal for one instance: unlike a static catalog of what exists, the graph answers "why is this score what it is?" by letting you expand the node, follow each contributing edge, see the transformation applied and the loss it introduced, and read the confidence that each hop propagates forward. It is built for the moment someone disputes or doubts a specific value and needs to trace its provenance and shakiness, edge by edge, down to the raw records — a live lineage explorer, not a coverage grid and not a written report.
Example¶
A consumer disputes their credit score: 612, dense enough to deny them a lease. A Source-to-Score Lineage Graph lets an analyst (and, in a redacted view, the consumer) start at the 612 node and traverse. One edge leads to a "payment history" sub-factor; expanding it reveals a transformation that treated a single 45-day-late medical bill as a major derogatory. Another edge shows a "credit age" factor computed from an account the consumer says isn't theirs — a leaf record that shouldn't be in the graph at all. Each node carries a confidence annotation: the disputed account's contribution is flagged low-confidence because it failed an identity match, and that low confidence propagates up toward the 612.
The tool renders no verdict — it does not decide whether the score is fair. What it does is make the path from raw records to the three-digit number walkable, so the one erroneous leaf and the one harsh transformation become visible and challengeable instead of buried in an opaque pipeline. The score stops being a bare fact and becomes a traceable computation.
How it works¶
The tool is graph-structured and traversal-first:
- Model the pipeline as a DAG. Represent every substrate input, intermediate feature, and transformation as nodes and edges, terminating in the published abstraction.
- Enable backward and forward walks. From any output, let a user expand contributing nodes down to raw records; from any input, trace forward to see what it affects (impact analysis when a source is found bad).
- Annotate edges with transformation and loss. Each edge names the operation applied and what it discarded, so the graph shows not just that a value flowed but what was lost at each hop.
- Propagate confidence along the graph. Attach a confidence to nodes and carry it forward through the edges, so a shaky input visibly weakens the outputs that depend on it.
The distinguishing discipline is navigability of a single instance: the tool's power is letting you traverse the provenance of one specific value, with loss and confidence attached to every step.
Tuning parameters¶
- Graph depth — how far down toward raw records the traversal reaches. Deeper graphs enable full provenance but grow large and slow and can expose sensitive substrate.
- Confidence model — how uncertainty is propagated along edges (simple flags versus formal error propagation). Richer models are more informative but harder to keep honest and easy to over-read.
- Redaction layer — how much of the substrate a given viewer may traverse to. Tight redaction protects privacy but can stop a trace short of the record in question.
- Loss annotation detail — whether edges note that loss occurred or quantify it. Quantified loss is powerful but expensive and can imply a precision the transforms don't have.
When it helps, and when it misleads¶
Its strength is that it makes an opaque score interrogable one instance at a time: when someone doubts a specific value, they can follow it to the exact record or transformation responsible, which is the difference between a challengeable decision and a black box. It is the applied form of data lineage — the recorded flow of data through its transformations from origin to output — turned into something a non-engineer can walk.[n1] Propagating confidence along the graph is what keeps a shaky input from wearing the same solid font as a trustworthy one at the output.
It misleads when the graph is complete but incomprehensible — thousands of nodes that technically show lineage while answering no human's actual question, a visual form of trace theater. Propagated confidence can also mislead: a tidy percentage at the root can imply the whole pipeline's uncertainty is understood when much of it was never modeled, false precision dressed as rigor. And a lineage tool tempts people to think tracing a computation validates it — but a faithfully-traced path from bad records to a bad score is still bad. The classic misuse is shipping the graph as proof of transparency while the transformations it renders remain unjustified. The guarding discipline is to design the traversal around real dispute questions, mark where confidence is unmodeled rather than implying zero, and remember the tool shows how the score was computed, never whether it should be believed.
How it implements the components¶
Source-to-Score Lineage Graph fills the traversal-and-uncertainty components — the guardrail rendered as a walkable computation:
transformation_and_loss_trace— its edges are the trace: each renders the operation from substrate toward output and the information that step discarded, walkable in either direction.abstraction_confidence_label— it attaches confidence to nodes and propagates it along the graph, so a value's shakiness is visible at the output it feeds.
It does not enumerate every abstraction field to certify coverage, nor catalog the full substrate reference set (reification_candidate_abstraction, substrate_reference_set) — that is Evidence-to-Abstraction Traceability Matrix, which proves all outputs are linked where this tool lets you traverse one; nor does it adjudicate disputes or retire the abstraction (revision_or_retirement_path) — that is Counterexample Case Review.
Related¶
- Instantiates: Abstraction–Substrate Traceability Guardrail — the graph makes a single output's provenance and uncertainty walkable to its roots.
- Consumes: Evidence-to-Abstraction Traceability Matrix supplies the certified source-and-transform links the graph renders as a navigable path.
- Sibling mechanisms: Evidence-to-Abstraction Traceability Matrix · Counterexample Case Review · Model Card or Datasheet Linkage · Point-of-Use Reification Warning · Proxy Drift Dashboard
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: Source To Score Lineage Graph operates by externalizes inputs, transformations, intermediate features, and outputs as a traversable lineage DAG. That concrete deployed or enacted form is Representation, Specification & Plan under the frozen taxonomy.
Nearest alternative: Interface, Display & Cue — Although Interface, Display & Cue can support this mechanism, the frozen evidence makes its operative form the act that externalizes inputs, transformations, intermediate features, and outputs as a traversable lineage DAG; the alternative is therefore secondary rather than defining.
Review outcome: Adjudicated after independent review; medium confidence.
Origin Attribution¶
Primary origin: Data Science & Analytics
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Tracing records through transformations into scores and decisions is analytical data lineage.
Related originating lineages:
- Accounting & Auditing — An audit trail supports reperformance and control review.
- Computer Science & Software Engineering — Computer science and software-engineering practice supplies a parallel or contributing lineage for the mechanism's defining operation: visualizes lineage from substrate records through transformations to the final score, label, dashboard value, or decision artifact.
- Library & Information Science — Provenance metadata preserves source and derivative relationships.
- Statistics & Experimental Design — Statistics, experimental design, and measurement theory supplies a parallel or contributing lineage for the mechanism's defining operation: visualizes lineage from substrate records through transformations to the final score, label, dashboard value, or decision artifact.
- Ethics of Technology & AI Governance — Decision accountability requires visibility into how inputs produced consequential outputs.
Review resolution: The blind reviewers agree that data_science is the primary origin and differ only on alternate origin disagreement, domain reach disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain cross_disciplinary_synthesis because the combined evidence shows material contributions from several lineages. The broader reach of multi_domain records portability separately from historical provenance; encyclopedia_synthesis=true preserves the affirmative synthesis judgment where either reviewer identified one.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Data lineage is the documented flow of data from its origins through each transformation to its final form, used to trace how a given output was produced and to assess the downstream impact of a bad input. This tool turns that record into an interactive graph so a specific score's provenance can be walked rather than reconstructed. ↩