ScienceThree Axes

Scientific Metrics

ECAM-X, AEI-Delta, and TCA as three independent falsifiability axes — not one super-score.

No super-score

AIJIM must not collapse ECAM-X, AEI-Delta, and TCA into one quality number. A single score would destroy scientific information: attribution validity, explanation stability, and human reliance are different claims with different falsification protocols.

Interpretation Matrix

The three axes answer different scientific questions. The most important column is what each axis does not measure: it prevents scope creep, green-washing, and accidental authority transfer between metrics.

Axis	Scientific Question	Measures	Does Not Measure	Authority	Falsifiable Via	Reader Label	Audit Depth	Fail-Closed Form	Paper Section
ECAM-X	Does this evidence support this claim?	NLI label, similarity, evidence-link coverage, negative controls, judge configuration, and attribution confidence.	Causality, source bias, editorial relevance, explanation stability, or human reliance quality.	Publication-required for claim/evidence support metadata.	Canonical JSON, model snapshot/hash, evidence content hash, judge config, and reproduced NLI link output.	Evidence grounding	Run IDs, evidence hashes, thresholds, NLI labels, similarity scores, and judgeConfig.	claim_missing_evidence_links or explicit evidence grounding unavailable.	Attribution Validity
AEI-Delta	Does the explanation remain stable under perturbation?	Seeds, perturbations, delta distributions, confidence intervals, sanity checks, and stability drift.	Truth, evidence quality, causality, bias, or editor trust. A stable explanation can still be wrong.	Publication-strengthening disclosure. Absence is honest not-measured, not failure and not green.	Seeds, perturbation recipe, model snapshot, stability artifact, and deterministic delta recomputation.	Stability evaluated / not evaluated	Perturbation set, delta histograms, CI bounds, failed sanity checks, model snapshot.	stability not evaluated; never synthesize a neutral stability score.	Explanation Stability
TCA	How much editorial reliance is justified and documented?	HITL decision coverage, override rate, review decisions, calibration curves, and reliance patterns.	Per-user surveillance, writing quality, factual truth, or source quality.	Publication-strengthening disclosure when HITL history exists. Absence is honest not-measured.	Timestamped HITL decision log, actor identity, revision identity, and deterministic calibration reconstruction.	Editorial trust calibration	Decision history, override patterns, review gates, calibration summaries, and actor-bound audit events.	trust calibration not available; never present missing HITL history as low risk.	Human Reliance Calibration

Three Kinds of Falsifiability

Bundle integrity

Core hashes, signatures, and canonical JSON bind the bundle. ECAM-X supplies reproducible claim/evidence attribution metadata inside that signed integrity envelope.

Experimental

AEI-Delta exposes seeds and perturbations so stability can be re-run like a reproducibility capsule.

Audit-log

TCA reconstructs reliance from timestamped HITL decisions, actor identity, and revision-bound choices.

Two-Tier Surface Discipline

Reader UI	Quiet labels: evidence grounding, stability evaluated/not evaluated, trust calibration. No raw run IDs.
Editor UI	Actionable state: which axis is missing, blocked, measured-positive, or measured-negative.
Audit UI	Technical depth: run IDs, hashes, thresholds, model snapshots, verifier outputs, and RFC conformance warnings.
Publication Bundle	Immutable disclosure: first-class publicationId plus axis artifacts that actually exist.
Verifier CLI	Strict publication-integrity PASS/FAIL first; EU AI Act readiness as PASS/PARTIAL/FAIL; optional quality report second. No quality score may weaken cryptographic verification.
Protocol Paper	Triple-Falsifiability Publication Protocol: cryptographic bundle integrity, experimental stability evidence, and audit-log human reliance calibration.

Honest absence

Every axis has three UI states: measured-positive, measured-negative, and not-measured. Not-measured is not green and not red; it is a truthful absence category. This is mandatory for AEI-Delta and TCA because they strengthen publication quality without becoming hidden requirements for every publication. A measured-positive axis can still carry an audit-only with-warning sub-state when RFC-003/RFC-004 reason codes disclose non-blocking gaps; Reader UI keeps the top-level state quiet, Audit UI exposes the reason codes.

EU AI Act readiness

The verifier reports EU AI Act readiness as PASS, PARTIAL, or FAIL from three artifacts:model_card.v1,risk_register.v1, andevaluation.v1. PARTIAL is an honest-absence state, not a score and not a cryptographic publication failure.

Verifier Stance

aijim-verify remains strict about publication integrity: publication identity, body hash, bundle hash, Merkle root, schema version, and mandatory ECAM-X support metadata must verify or the publication does not pass. AEI-Delta and TCA belong in a separate scientific quality report unless a publication explicitly claims those axes. Optional quality disclosures may enrich verification output, but they must never turn a cryptographic failure into a pass.

Protocol Paper Framing

The strongest scientific framing is a Triple-Falsifiability Publication Protocol: AIJIM publishes AI-assisted work with signed bundle integrity, evidence-bound attribution validity, experimental explanation stability, and audit-log human reliance calibration. Unlike source-display systems or quality dashboards, AIJIM keeps these axes orthogonal at reader, editor, audit, and verifier surfaces.