Evidence-to-Abstraction Traceability Matrix¶
Artifact — instantiates Abstraction–Substrate Traceability Guardrail
Links each abstraction field, score, class, model output, or summary claim to source evidence, transformation steps, assumptions, and validation records.
An Evidence-to-Abstraction Traceability Matrix is a completeness table: one row per abstraction element, one column per warrant question, and a filled cell wherever an element can answer "what am I made of?" Its defining idea is coverage — the artifact's job is to prove that every field, score, or class an abstraction publishes has a documented source, a documented transformation, and a documented validation, so that a gap in the table is itself the finding. It is not a narrative and not a picture; it is a grid whose empty cells are alarms. Where a good abstraction can be consulted at a glance, this matrix records what would take a day to reconstruct: for each visible number, which substrate records fed it, what was done to them, and whether anyone ever checked that the result holds. It makes traceability auditable rather than merely asserted.
Example¶
A national body publishes hospital "quality star ratings" — one to five stars per hospital, driving patient choice and reimbursement. Behind each hospital's rating sit dozens of measures: readmission rates, infection counts, survey scores. An Evidence-to-Abstraction Traceability Matrix lays these out as rows (each measure that rolls into the star), and columns for source records (which admissions, which claims files, which survey responses), transformation (risk adjustment, exclusions, winsorizing, weighting into the composite), and validation (was this measure's reliability tested; when).
The exercise is unforgiving in the useful way. Most rows fill in. But the "patient experience" measure has an empty validation cell — no one can point to a test that its survey response rate is high enough to be stable at small hospitals. And the composite-weighting row's source cell reads "policy decision, 2019," not evidence. The matrix has not judged the rating; it has shown exactly where the rating is unlinked to substrate — two cells that now decide whether the star can be relied on for small hospitals, or must carry a caveat.
How it works¶
The artifact is built row-first and coverage-driven:
- Enumerate the reification candidates. List every element the abstraction actually exposes to a decision-maker — each field, sub-score, class, or headline claim gets a row. Naming them all is half the work; unnamed outputs are the ones that go unlinked.
- Bind each row to its substrate. Fill the source column with the specific records, samples, or observations the element derives from — not "our data" but the identifiable set.
- Record the transformation and its losses. Capture the path from those records to the element: the aggregation, filtering, exclusions, and weighting, and what each step drops.
- Read the empty cells as findings. An unlinked source, an undocumented transform, or a missing validation is the artifact's actual output — the places the abstraction floats free of what it claims to represent.
The distinguishing discipline is completeness: the matrix's value is the guarantee that nothing the abstraction shows is un-accounted-for, and its failures are gaps, not wrong links.
Tuning parameters¶
- Row granularity — one row per headline output versus one per intermediate field. Finer rows expose more unlinked internals but multiply maintenance.
- Column set — how many warrant questions each row must answer (source, transform, assumption, validation, owner). More columns are more thorough but raise the bar for "complete."
- Fill standard — whether a cell may cite a document, or must resolve to identifiable records. Strict standards make the matrix trustworthy but expensive to populate.
- Gap policy — whether an empty cell merely flags or actively blocks reliance on that element. Blocking forces resolution but can freeze a mostly-sound abstraction over one hole.
When it helps, and when it misleads¶
Its strength is turning "is this traceable?" from an assertion into a checkable grid: you can see whether every output links to evidence, and the holes announce themselves. It is the direct descendant of the requirements traceability matrix, which proves every requirement maps to a test and flags any that don't.[n1] Applied to abstractions, it is the artifact that makes substrate-traceability real instead of claimed.
It misleads through trace theater — a matrix every cell of which is filled, yet whose links don't answer any question a real reliance would ask. A cell that reads "source: warehouse table X" is technically full and practically useless if no one can tell what X actually contains. A dense, green-celled matrix can radiate a confidence the underlying links don't earn; completeness of form is not soundness of link. The classic misuse is populating cells to pass an audit rather than to answer a decision-maker's question. The guarding discipline is to test cells against real reliance questions ("if I doubted this number, does this cell let me find and check its source?") rather than grading the matrix on how few blanks it has.
How it implements the components¶
Evidence-to-Abstraction Traceability Matrix fills the coverage-and-provenance components — the guardrail's proof that the abstraction is fully accounted-for:
reification_candidate_abstraction— its rows enumerate every specific field, score, or class that has acquired authority, so none goes untracked.substrate_reference_set— its source column names, per row, the identifiable records and observations the element was abstracted from.transformation_and_loss_trace— its transform column records the path from substrate to element and what each step drops.
It does not propagate confidence or loss through a navigable graph (abstraction_confidence_label) — that is Source-to-Score Lineage Graph, which lets you traverse one output back to its roots where this artifact instead certifies that all outputs are covered; nor does it prompt a reviewer at decision time (reliance_decision_gate, validity_scope_boundary) — that is Map–Territory Review Checklist.
Related¶
- Instantiates: Abstraction–Substrate Traceability Guardrail — the matrix is the artifact that makes substrate-traceability auditable.
- Sibling mechanisms: Source-to-Score Lineage Graph · Model Card or Datasheet Linkage · Map–Territory Review Checklist · Counterexample Case Review · Category Language Audit
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: Evidence-to-Abstraction Traceability Matrix operates as a non-executable information artifact that externalizes static or prospective structure because it links each abstraction field, score, class, model output, or summary claim to source evidence, transformation steps, assumptions, and validation records.
Independent corroboration: The frozen evidence defines Evidence-to-Abstraction Traceability Matrix as 'Links each abstraction field, score, class, model output, or summary claim to source evidence, transformation steps, assumptions, and validation records', so its operative form is Representation, Specification & Plan.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Engineering & Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Bidirectional traceability matrices originated in systems and requirements engineering to link derived artifacts back to sources and transformations.
Related originating lineages:
- Computer Science & Software Engineering — Software requirements and model lifecycle practice materially standardized machine-checkable trace links through transformations.
- Data Science & Analytics — Modern model and data lineage materially extend traceability to scores, transformations, and validation records. Model cards, data lineage, and analytical provenance materially shaped traceability for scores and model outputs.
Review resolution: Both reviewers agree that engineering_design is primary. I retain data_science, computer_science only as formative origin lineages; cross_disciplinary_synthesis is appropriate because the final form materially combines the agreed primary with the retained formative lineages. Reach is multi_domain because the structure transfers across several fields but is not a near-universal human pattern, an applicability judgment kept separate from provenance. Encyclopedia synthesis is true because the exact generalized artifact is an encyclopedia-authored combination or refinement. No unresolved historical ambiguity remains after reconciling the secondary fields.
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] A requirements traceability matrix is a standard systems-engineering artifact: a grid mapping each requirement to the design elements and tests that satisfy it, so any requirement with no downstream link is immediately visible. Its whole value is that unmapped rows are the finding — the same logic this matrix applies to abstraction outputs and their substrate. ↩