Source Attribution Confidence Rubric¶

Confidence-grading rubric — instantiates Use-Time Source Attribution Calibration

A graded scale that scores how sure you are of an item's source — separately from whether the content is true — and trips a corroboration gate when the grade is low and the stakes are high.

Knowing which class an item belongs to is not the same as knowing how firmly you know it. Source Attribution Confidence Rubric takes a source label — however produced — and grades the confidence in that attribution on a defined scale, then crosses that grade with the stakes of the decision to decide whether the item may be used as-is or must be corroborated first. Its defining move is holding source confidence and content confidence apart: it scores how sure you are of where the item came from, not whether the item is true, because a claim can be correctly sourced yet weakly supported, or perfectly true yet badly misattributed. The rubric turns a private, sliding sense of "pretty sure" into a stated grade that means the same thing across people and items — and into an explicit trigger for when to go get more evidence.

Example¶

A genealogist is compiling a family history whose "facts" arrived by very different routes. She scores each attribution against the rubric. A birth date read directly off a scanned parish register grades A — documented primary: the source is the record itself, confidence in the attribution is high. A date auto-suggested by a genealogy site's hint engine and never checked against the underlying record grades D — machine-suggested, unverified. A great-grandfather's emigration "on a specific ship" that rests entirely on a cousin's retelling grades C — single oral source.

Then she crosses each grade with stakes. Most facts are low-stakes leaves and can carry their grade as-is. But the emigration claim is load-bearing — a whole branch of the tree hangs off which ship, which port, which year. Grade C on a high-stakes node trips the corroboration gate: before that branch is published as fact, go find a passenger manifest or naturalization record. The rubric has not decided the claim is false; it has decided this particular low-confidence attribution is carrying too much weight to pass unchecked.

How it works¶

Grade the attribution, not the content. Score confidence that the source label is correct — is the origin documented, inferred, single-sourced, reconstructed, machine-suggested? — on a defined scale with anchored descriptors, so a grade is reproducible rather than a mood.
Attach the grade as an annotation. The score rides with the item downstream, alongside its source class, so later users inherit the uncertainty instead of seeing a bare label that reads as settled.
Read the stakes tier. Establish what the item is about to be used for — brainstorming, an internal note, a published claim, a safety or legal decision — since the same weak attribution is tolerable in one and disqualifying in another.
Escalate on the low-grade / high-stakes cell. Where a low confidence grade meets a high stakes tier, the gate routes the item to corroboration before use; elsewhere it passes with its grade attached.

What distinguishes it is that it operates on the confidence axis — it presumes a source label already exists and asks how much weight that label can bear, scaling the answer to what is riding on it.

Tuning parameters¶

Scale granularity — a coarse high/medium/low versus a lettered or numeric ladder with anchored descriptors. Finer grades discriminate better but invite false precision and inter-rater drift on a soft judgment.
Anchor definitions — how concretely each grade is pinned to observable source conditions ("documented primary," "single uncorroborated report"). Tight anchors keep grades comparable across people; loose ones collapse back into gut feel.
Stakes tiering — how many stakes levels there are and where the thresholds sit. More tiers target corroboration effort precisely but complicate the call; fewer are simpler but blunt.
Escalation threshold — which grade-by-stakes cells trip the gate. A conservative setting corroborates more (safer, slower); an aggressive one lets more low-confidence items through on stakes grounds.
Annotation persistence — whether the grade is mandatory downstream or droppable for brevity. Sticky grades preserve calibrated uncertainty; droppable ones keep prose clean but let confidence silently normalize to certainty.

When it helps, and when it misleads¶

Its strength is that it makes a sliding, private sense of certainty into a stated, comparable grade — and, crucially, it keeps source confidence from being read as content truth, the confusion that lets a well-attributed guess and a poorly-attributed fact both pass as "solid." Tied to stakes, it aims corroboration effort where a weak attribution is actually carrying weight, rather than everywhere or nowhere.

Its failure modes are the ones every scoring instrument invites. A tidy grade can be mistaken for the accuracy itself — a bold "A" does not make the underlying claim true, and a rubric whose grades are never checked against outcomes is uncalibrated decoration.^[1] Its classic misuse is running backward: grading a source high to justify using material already chosen, rather than grading first and letting the grade gate the use. And a rubric can lend false precision to what is really a coarse judgment. The discipline that keeps it honest is to hold source confidence and content confidence on separate axes, to periodically check whether "high confidence" attributions actually turn out correct, and to treat the grade as an input to the corroboration decision, not a substitute for corroboration.

How it implements the components¶

Source Attribution Confidence Rubric fills the confidence-and-routing side of the archetype — the components that decide how much a source label can be trusted and what to do when it can't:

source_confidence_annotation — its core output: a graded, anchored confidence score attached to each item's source attribution and carried downstream.
corroboration_escalation_gate — the rule that, on a low-grade / high-stakes item, routes it to external grounding before it can be used as evidence.
decision_stakes_tier — the stakes classification the rubric crosses against the grade, so strictness scales with what the item is about to affect.

It does NOT produce the raw label from the source_cue_feature_set and use_time_attribution_rule it grades (the Reality Monitoring Checklist and Memory Source Probe do), resolve a provenance_backlink_or_trace (Provenance Lookup Before Publication does), or compile the source_attribution_error_profile across many attributions (the Source Confusion Matrix Review does).

Instantiates: Use-Time Source Attribution Calibration — this rubric is the confidence layer that decides how much weight a source label can bear and when to escalate.
Consumes: the raw source label produced by the Reality Monitoring Checklist or Memory Source Probe — the rubric grades that label rather than generating it.
Sibling mechanisms: Reality Monitoring Checklist · Memory Source Probe · Observation Recheck or Replication · Borrowed Idea Attribution Scan · Chain-of-Custody or Lineage Check · Generated Content Disclosure Gate · Hallucination Intrusion Triage · Provenance Lookup Before Publication · Source Attribution Training Set · Source Confusion Matrix Review · Source-Label Preserving Summary Template

References¶

[1] Calibration — the degree to which stated confidence matches actual accuracy, so that attributions graded "high confidence" turn out correct about as often as the grade implies. A confidence rubric only earns trust once its grades have been checked against outcomes; a well-formatted grade is not itself evidence that it is calibrated. ↩