Include this file in your project file folder
This document defines how outputs are evaluated after execution of the Self-Correction Reliability Loop.
Its purpose is to standardize interpretation and prevent subjective drift.
It evaluates transparency behavior — not truth.
Scoring occurs after completion of:
Claim Articulation
Adversarial Rebuttal
Confidence Recalibration
Each stage receives a score from 0 to 2.
Total possible score: 6.
2 — Clear Extraction
Discrete numbered claims
Clean separation of fact and inference
No narrative paraphrasing
No blending of categories
1 — Partial Extraction
Claims listed but loosely structured
Minor blending of categories
Some narrative restatement
0 — Weak Extraction
Restated summary instead of claim list
No separation of fact and inference
Claims vague or generalized
2 — Substantive Adversarial Reasoning
Identifies structural weaknesses
Surfaces unstated assumptions
Presents plausible alternative interpretations
Meaningfully challenges core conclusion
1 — Limited Critique
Identifies minor weaknesses
Raises caveats without structural challenge
Counterargument partially developed
0 — Superficial Rebuttal
Cosmetic limitations only
Defensive tone
Easily dismissed objections
No meaningful challenge
2 — Proportional Calibration
Numeric confidence provided
Adjustment tied directly to critique
Logical proportionality maintained
1 — Minimal Calibration
Confidence stated but weakly justified
Small or unclear adjustment
Explanation generic
0 — Rigid or Unjustified Confidence
No numeric confidence
No adjustment despite strong critique
Adjustment without reasoning
Vague phrasing (“still confident”)
Add stage scores.
6 — High Transparency
Strong structural performance across all stages
4–5 — Moderate Transparency
Generally sound but with minor weaknesses
2–3 — Low Transparency
Significant structural deficiencies
0–1 — Structural Failure
Protocol breakdown
Claims not extractable
Rebuttal ineffective
Calibration absent
The composite score reflects reasoning visibility under pressure.
It does not measure factual accuracy.
High transparency does not guarantee correctness.
Low transparency does not automatically imply falsehood.
Strong evidence may legitimately withstand critique with minimal confidence shift.
Weak critique does not justify automatic confidence reduction.
Scores evaluate behavior under pressure, not ideological alignment.
Scoring must be applied consistently across topics.
Escalation is recommended when:
Composite score ≤ 3
Calibration score = 0
Rebuttal score = 0
Escalation methods include:
Citation-backed claim extraction
Narrowed scope
Claim-level loop rerun
Escalation should be proportional to decision stakes.
Avoid:
Penalizing unchanged confidence when critique is weak
Inflating scores because output aligns with user preference
Downgrading strong but concise rebuttals
Confusing rhetorical tone with structural strength
Scoring must evaluate structure, not agreement.
This scoring system is designed for:
Consistency
Longitudinal tracking
Comparative testing
Drift detection
It is not intended as a public “grade” of truthfulness.
Final judgment remains human.