Evidence Sufficiency Rubric¶

Evidence review — instantiates Regress Termination Rule

Grades whether a body of evidence is strong, diverse, and relevant enough that the chain of 'but what backs that?' can stop — calibrated to the evidentiary regress specifically.

Any factual claim can be met with "and what backs that?", and the backing can be questioned in turn, forever. An Evidence Sufficiency Rubric ends the evidentiary form of that regress by grading the support itself against explicit criteria — quality, quantity, independence, and relevance — and declaring the chain grounded once it rests on support that clears the bar. What makes it THIS mechanism is its narrow, sharp focus: it judges the evidence's own backing, not whether the overall decision is ready and not whether to keep collecting more. It applies only where the demand is genuinely for evidence — and refuses to be used to settle a value or authority question it cannot touch.

Example¶

A newsroom is about to publish that a public official steered contracts to a relative. Before it runs, an evidence sufficiency rubric is applied claim by claim. For the central allegation the rubric traces the backing chain — what supports this? two sources. What supports them? one is a signed contract (a primary document); the other is a person with direct knowledge — and asks whether the two lines are truly independent or secretly trace to the same origin. It grades each key claim against the standard: contested factual claims need at least two independent lines, documentary where obtainable, and a genuine attempt to find contrary evidence. Claims that clear the bar are treated as grounded; one detail that rests on a single anonymous source does not clear it, so it is either cut or published with an explicit caveat. The rubric doesn't decide whether to run the story — it decides which of its claims are evidentially solid enough to stop questioning.

How it works¶

Walk the backing chain per claim. For each load-bearing claim, follow "what supports this, and what supports that?" down to where it rests on primary evidence or independent corroboration.
Grade against explicit criteria. Score quality, quantity, independence (not just count), and relevance against the standard — echoing corroboration that traces to one origin is not two sources.
Ground or flag. A claim that clears the bar is treated as sufficiently supported and the evidentiary regress stops there; one that doesn't is escalated, cut, or carried with its gap named.
Stay in its lane. The rubric only adjudicates evidentiary demands; a value or procedural "why" is handed off, not answered with more evidence.

Tuning parameters¶

Independence requirement — how many genuinely separate lines a claim needs. Raising it kills laundered corroboration but can stall on hard-to-verify truths.
Quality-versus-quantity weighting — whether one strong primary source outweighs several weak ones. Favoring quality resists volume-gaming; favoring quantity guards against a single point of failure.
Relevance and proximity strictness — how directly the evidence must bear on the exact claim. Tighter proximity cuts inferential leaps but discards useful circumstantial support.
Contrary-evidence duty — whether an active search for disconfirming evidence is required before grounding. Requiring it fights confirmation bias at the cost of speed.

When it helps, and when it misleads¶

Its strength is a defensible floor for factual closure: it stops the "what backs that?" regress with a rule that catches the two commonest evidentiary sins — the single unverified source and the echo-chamber "corroboration" that isn't independent. It is the right tool precisely when the regress is about facts.

Its failure modes are proportional to its narrowness. Applied to the wrong regress — a value premise, an ought — it manufactures false closure, piling up evidence for a question evidence cannot settle. Within its lane, it is gamed by counting sources that clear the bar only on paper (two outlets both quoting the same press release), and it can be run backwards, with the independence bar quietly lowered for the story someone wants to publish. The discipline that keeps it honest is to test independence rather than count it, to publish the residual gap rather than bury it, and to refuse the rubric any question that is not actually evidentiary.^[1]

How it implements the components¶

sufficiency_threshold — it sets and calibrates the bar the evidence must clear on quality, quantity, independence, and relevance.
regress_chain — it makes the evidentiary support chain visible, walking each claim down to its backing to find where it genuinely grounds.
uncertainty_residue — it names the evidential gaps that remain after grading, so a claim can be published or acted on without pretending it is airtight.

It does not decide whether the overall decision may close (stopping_criterion → Decision Closure Criteria), type non-evidentiary demands or allocate who must prove them (link_type_distinction → Burden-of-Proof Rule), or decide when to stop gathering evidence in the first place (Research Stopping Rule).

Instantiates: Regress Termination Rule — it terminates the evidentiary regress with a graded sufficiency bar.
Sibling mechanisms: Decision Closure Criteria · Research Stopping Rule · Burden-of-Proof Rule · Axiom Set · First-Principles Statement · Five Whys with Stop Rule · Governance Authority Chain · Review Trigger Register · Timeboxed Inquiry Record · Assumption Log

Notes¶

A rubric grades a fixed body of evidence for adequacy; a Research Stopping Rule decides when to stop adding to that body. Confusing them lets a team keep collecting because the rubric isn't satisfied, when the honest answer is that the available evidence has been exhausted and the residual gap should be carried forward instead.

References¶

[1] The journalistic two-source rule — a contested claim needs at least two independent confirmations — is a familiar evidence sufficiency threshold; its well-known weakness, sources that merely echo a common origin, is why independence must be tested rather than counted. ↩