# TRACE v0.1: novelty and contribution matrix

Every entry is a proposed artifact or empirical question. None is an established methodological innovation or observed improvement. Reference IDs resolve in references.json and in the paper.

| Item | Closest precedent | Status | Necessary comparison or decision |
|---|---|---|---|
| Invariance and response to transformations | Against All Likelihood, ICP/IRM and behavioral testing (R01–R09) | Inherited motivation and safeguards | H1 diagnostics; stability alone never certifies truth |
| Explicit claim scope and attribution | Claim Cards (R43), existing SNO work | Inherited representation | Measure extraction/interpretation errors; do not claim schema novelty |
| Dependence-aware evidence accounting | Source-dependence literature (R10) | Inherited technique integrated | D–C with identical lineage facts; explicit weights, eligibility and warrant rules vary together |
| Abduction from contradictory findings | Behfar/Okhuysen (R47), PARADOX (R48), clinical evidence synthesis (R49) | Established premise; no priority claim | Inspect full PARADOX article beyond the released detection/graph code before definitive operational differentiation |
| Contexts, derived variables and conditional relationships | AutoDiscovery (R50) | Existing generator capabilities | Preserve these capabilities in comparators; test localization within the same native discovery method |
| Competing explanations and discriminating evidence | Contrastive Discovery (R41), experiment-design/rule-learning work (R20/R31/R40) | Inherited loop | Faithful contrastive method-level comparator and controlled scaffold intervention |
| Agentic falsification | POPPER (R42) | Inherited; guarantees not transferred | Later comparison on tasks satisfying implication and conditional-validity assumptions |
| Multi-agent arbitration | Debate, co-scientist and Perceive-Then-Adjudicate (R13/R14/R35/R46) | Optional implementation | Simpler single-model implementation unless benefit is demonstrated |
| Numerical rule compiler | Ordinary induction/ILP (R21/R38) | Implementation option | Classical learner and affordable exhaustive search; no artificial cap |
| Source-conflict localization changes candidate generation | Close overlap above; operational differentiation uncertain | Primary H2 research question | Static C–B with same facts, model and capacity; blinded common selector and summary-call control; report native-method localization transfer comparison separately |
| Explicit lineage improves discovery conditional on localization | R10 and source-grounded reasoning | Secondary D–C question | Same metadata; no information advantage; full factorial needed for an independent main-effect claim |
| Localization improves extraction only | Existing structured interpretation approaches | Possible narrower result | End-to-end versus common-input comparison; do not call it discovery improvement |
| Candidate prioritization saves search cost | Existing search/ranking | Secondary mechanism diagnostic | Frozen pool, changed order only; metered full budget curve |
| Shared records enable cross-case analysis | Existing provenance and experiment logging | Proposed reproducibility artifact | Inspect source fidelity and data availability; schema validity alone is insufficient |
| A contradiction/discovery benchmark | ConfRAG, DiscoveryBench and related benchmarks (R44/R25) | Contingent artifact, not yet produced | Demonstrate new coverage, reliable labels, identifiable targets and independent cases |

## Outcomes that change the proposal

- C improves distinct measurable candidates and sealed predictions over B at acceptable cost: investigate the localized-conflict mechanism on fresh cases; no general superiority claim follows from a small pilot.
- Static C–B improves but native-method augmentation does not: retain the scaffold-specific finding; do not claim a transferable component. Report uncertainty before treating a null result as evidence of absence.
- C changes candidates but not predictions: hypothesis diversity alone does not support H2.
- Only D improves: investigate dependence handling; do not attribute the result to conflict localization.
- Only extraction/reliability improves: narrow the research claim and compare against structured verification methods.
- Ordinary methods match or exceed TRACE: simplify or retire the proposed method; nonsignificance in a small pilot does not establish equivalence.
- A benchmark reveals failures: assess its independent contribution; negative results do not automatically establish a useful benchmark.
- Prior work covers the operational distinction: attribute it and withdraw the novelty claim, even if the experiment remains a useful replication.
