Framework v0.1

Sources & Attribution II

Intellectual lineage of the grounding interventions.

Appendix — Research Grounding Framework

Part II: Grounding Intervention Lineage

Epistemic tier key:

  • T1 — Primary empirical: original study with direct measurement
  • T2 — Secondary synthesis: claim derived across multiple T1 sources through reasonable extrapolation; synthesis move is explicit and traceable
  • T3 — Practitioner/domain observation: expert opinion, case study, domain observation
  • T4 — reviewed — Novel framing, explicitly human-reviewed
  • T4 — provisional — Novel framing, not yet reviewed; should not appear in main document without acknowledgment

GI-1: Source Anchoring

Claim: Requiring a named, locatable source at the point of output reduces fabrication by making unsourced claims inadmissible. [T2]

The underlying mechanism is constraint on the generation space: fabrication occurs in part because unsourced claims face no output-level check. Source anchoring changes the admissibility condition. This is a synthesis inference from the hallucination literature — no single study directly tests source anchoring as defined here as an intervention, but the mechanism is derivable from the documented relationship between retrieval grounding and hallucination reduction. RAG systems provide the closest empirical analog: grounding generation against a retrieved corpus reduces fabrication rates. GI-1 applies this principle as a researcher-administered constraint rather than a system-level architecture.

Sources: RAG literature generally; hallucination reduction through grounding is a well-documented finding across multiple T1 studies. The specific formulation as a researcher-administered generation constraint is T2 — derivable from the RAG mechanism but not the same intervention.

Claim: Source anchoring is distinct from provenance tracing — it requires source naming but not source verification. [T4 — provisional]

This distinction is the framework’s own design choice, not a standard distinction in the literature. The purpose is operational: source anchoring is a lightweight preventative constraint applied during generation; provenance tracing (GI-9) is a heavier corrective intervention applied afterward. Treating them as the same intervention would collapse the pre/post distinction that structures the framework. Requires human review to confirm the operational boundary is stable.

GI-2: Temporal Scoping

Claim: Restricting claims to a defined knowledge window prevents training-data-era information from being presented as current. [T1/T2]

The T1 grounding is the temporal validity observation documented in the validation log: AI capability research has a short shelf life, and model generations tested across the corpus span a three-year range (Claude 1.3 through GPT-5.2). The validation log’s recommendation that rate figures carry model generation and date of measurement is itself an application of temporal scoping as a documentation discipline. The broader claim that temporal restriction prevents a class of miscalibration is T2 — derivable from the documented temporal validity problem but not directly measured as an intervention effect.

Sources: Validation log, Temporal Validity Observation section (T3 — framework authors’ own systematic observation). The intervention mechanism is T2.

GI-3: Epistemic Status Tagging

Claim: Explicit epistemic tier assignment prevents T4 claims from being treated as T1, making the foundation of an argument visible and auditable. [T2]

The failure mode this addresses — confidence miscalibration, specifically the treatment of AI-generated synthesis as if it were primary empirical evidence — is documented empirically in Fernandes et al. (2025) and in the Synthesis Amplification Tendency observation in the validation log. The intervention mechanism (forcing explicit labeling before use) is derivable from the general principle that making implicit assumptions explicit reduces their unexamined influence. No single study directly tests epistemic tier tagging as defined here.

Sources: Fernandes et al. (2025) for the miscalibration problem. Validation log Cross-Cutting Observations for the Synthesis Amplification Tendency. The tier system design is T4 — provisional (see revised GI-3 definition). The claim that explicit tagging addresses the mechanism is T2.

Claim: The GRADE evidence quality framework in medicine and confidence interval reporting in statistics are adjacent prior art but do not determine the tier structure. [T3]

GRADE (Grading of Recommendations, Assessment, Development and Evaluation) provides tiered evidence quality assessment for clinical recommendations. It is domain-specific and oriented toward treatment decisions rather than research synthesis. Statistical confidence interval reporting is quantitative and does not apply to prose synthesis claims. These are practitioner frameworks informing the design space without providing the design. The observation that neither transfers directly to AI-assisted research quality contexts is the framework’s own domain assessment.

Sources: GRADE framework documentation; standard statistical reporting conventions. T3 — domain observation.

GI-4: Retrieval Augmentation

Claim: Forcing re-query against a verified corpus rather than parametric memory reduces fabrication and attribution drift. [T1]

RAG (Retrieval-Augmented Generation) has direct empirical support as an architecture for reducing hallucination. The intervention mechanism — grounding generation in retrieved text rather than model weights — is well-established in the NLP literature. GI-4 applies this as a researcher-administered generation constraint rather than a system-level architecture, which is an implementation difference but not a mechanistic one.

Sources: RAG literature; multiple T1 studies document hallucination reduction through retrieval grounding. The specific researcher-administered formulation is T2.

Claim: Effective retrieval augmentation requires specification of which corpus to retrieve from — unrestricted web search is not retrieval augmentation in this sense. [T3]

This is a practitioner design constraint, not an empirically tested claim. The reasoning is that corpus specificity determines the epistemic quality of the grounding: retrieval against a peer-reviewed preprint server is epistemically different from retrieval against the open web even if both are technically “retrieval.” This boundary condition is the framework’s own operational judgment, informed by systematic review methodology (which specifies databases rather than general search) but not stated in the RAG literature.

Sources: Systematic review methodology for the corpus-specification principle. T3 — practitioner design judgment.

Claim: Explicitly requiring a search for contradicting evidence before accepting a claim as supported directly counters sycophantic retrieval. [T1]

Batista & Griffiths (2026) provide the primary mechanism grounding: default LLM behavior samples from the distribution implied by the user’s hypothesis, systematically omitting data that would conflict with it. GI-5 directly interrupts this mechanism by making disconfirmation search a required step rather than an optional one. The Wason 2-4-6 task finding — that unbiased sampling yielded discovery rates five times higher than confirmatory sampling — provides direct empirical support for the claim that forcing disconfirmatory evidence changes epistemic outcomes.

Sources: Batista & Griffiths (2026), arXiv:2602.14270v1. Preprint caveat applies.

Claim: It is not sufficient to note that no contradicting evidence was found — the search itself must be executed and documented. [T2]

This procedural requirement is a synthesis inference from two sources: the Scope Enumeration principle (GI-15) that making search boundaries visible is necessary for the researcher to assess completeness, and the documented pattern in the validation log that absence of contradicting evidence in AI outputs often reflects absence of search rather than absence of evidence. The requirement that the search be documented — not merely claimed — follows from the same transparency principle that motivates GI-15.

Sources: Validation log Cross-Cutting Observations; GI-15 design rationale. T2 — synthesis inference.

GI-6: Contradiction Forcing

Claim: Requiring explicit articulation of a tradeoff in any claim presenting a benefit makes it structurally impossible for the AI to return a purely confirmatory response. [T2]

The mechanism draws on two sources. First, the sycophancy literature establishes that RLHF-trained models are systematically disposed toward confirmatory outputs; Sharma et al. (2024) document this across multiple model families. Second, the TRIZ contradiction principle establishes that hard problems involve a tension between parameters — improving one degrades another. Requiring articulation of the degraded parameter forces the model to process the tradeoff rather than suppress it. Neither source states this specific intervention; the combination is a framework synthesis.

Sources: Sharma et al. (2024) for the sycophancy mechanism. TRIZ contradiction identification principle (Altshuller) for the structural forcing logic. The combination as a grounding intervention is T2.

Claim: The TRIZ contradiction identification mindset transfers to grounding contexts; the lookup table does not. [T3]

This is the framework’s own domain assessment of which elements of TRIZ are portable. The contradiction identification principle is domain-agnostic — any system can be analyzed for parameters in tension. The contradiction matrix (the 39×39 lookup table mapping parameter pairs to recommended inventive principles) is empirically grounded in patent analysis and is domain-specific to engineering problems. The claim that the mindset transfers while the artifact does not is a practitioner judgment about the structure of the tool, informed by TRIZ cross-domain application literature but not stated there in this form.

Sources: TRIZ methodology documentation; TRIZ cross-domain application literature. T3 — practitioner domain assessment.

GI-7: 9-Windows

Claim: Requiring examination of a claim from subsystem, system, and supersystem perspectives across past, present, and future counters framing capture and surfaces omitted context. [T2]

The 9-Windows intervention (TRIZ System Operator) is documented in the TRIZ literature as a frame expansion tool effective across non-manufacturing domains. The application to AI-assisted research grounding — using it to force the model outside the frame established by the initial query — is a novel application, but the frame expansion mechanism is the same. The claim that systematic multi-level frame expansion surfaces omitted context is derivable from the documented TRIZ application literature and from the Omission failure mode analysis (FM-7): if omissions arise from query-constrained search, frame expansion at the supersystem level will surface categories the original query did not reach.

Sources: TRIZ 9-Windows / System Operator documentation; TRIZ cross-domain application literature confirming portability to abstract domains. The grounding application is T2.

Claim: 9-Windows is among the TRIZ tools confirmed to transfer well to abstract and non-manufacturing domains. [T3]

This is a domain assessment based on the TRIZ application literature rather than a direct empirical claim. The TRIZ literature includes documentation of 9-Windows application in software, education, and policy contexts. The assessment that it transfers better than the contradiction matrix (which requires domain-specific parameter calibration) is the framework’s own judgment, consistent with published TRIZ practitioner guidance.

Sources: TRIZ cross-domain application literature. T3 — practitioner domain assessment.

GI-8: Ontology Grounding

Claim: Requiring named entities to be anchored to a structured knowledge base before use in claims catches a specific class of fabrication where entities are plausible but not real. [T2]

The entity fabrication class — plausible-sounding but nonexistent people, studies, institutions — is well-documented in the hallucination literature. The mechanism of ontology anchoring as a check is derivable from knowledge graph and named entity recognition methodology: structured knowledge bases provide a ground truth against which entity plausibility can be verified. No single study directly tests GI-8 as defined, but the mechanism is traceable from the hallucination literature and knowledge representation research.

Sources: Hallucination literature for the entity fabrication class. Knowledge graph and NER methodology for the anchoring mechanism. T2 — synthesis inference.

Claim: Ontology grounding is distinct from source anchoring — it operates on entities within claims rather than sources attached to claims. [T2]

This distinction is derivable from the structural difference between what GI-1 and GI-8 check: GI-1 asks whether a source exists; GI-8 asks whether the entities named in the claim exist. A claim can pass GI-1 (a source is named) while failing GI-8 (the named entity is fabricated). The distinction is a synthesis inference from comparing the two intervention mechanisms; it is not stated in any single source but is directly derivable.

GI-9: Provenance Tracing

Claim: Tracing a claim backward to determine whether it is present in a named source, synthesized from multiple sources, or AI-generated without a source basis addresses attribution drift and synthesis validity. [T2]

The intervention mechanism is the post-hoc application of the same epistemic tier logic that GI-3 applies preventatively. Provenance tracing asks the T1/T2/T4 question after generation rather than before it. The claim that this addresses attribution drift is derivable from the van den Akker/Bogaert case (validation log Entry 1): systematic backward tracing would have caught the attribution error at the first citation. The claim that it addresses synthesis validity is derivable from the ICLR 20% conflation case (Entry 3): backward tracing of the composite figure would have identified the two distinct source findings.

Sources: Validation log Entries 1 and 3 as documented cases. The intervention mechanism is T2 — derivable from the documented cases.

Claim: A claim identified as T4 through provenance tracing should not be cited forward as if it were a sourced finding (forward-prevention property). [T2]

This is a direct application of the epistemic tier logic: if provenance tracing reveals that a claim has no source basis, the appropriate response is to flag it rather than propagate it. The forward-prevention property is a synthesis inference from the attribution drift compounding claim in FM-2 — if each generation of secondary citation increases the probability of further drift, then preventing T4 claims from entering citation chains is a drift-reduction intervention. The mechanism is derivable but not stated in any single source.

GI-10: Decompose and Verify

Claim: Breaking a complex claim into constituent sub-claims and verifying each independently catches errors embedded in synthesis that a holistic check would miss. [T2]

The mechanism is derivable from two sources. First, the general principle that complex claims can be partially true — a false or unsupported component hidden inside a plausible overall structure — is implicit in the claim decomposition literature in NLP and logic. Second, the validation log Entry 3 (ICLR 20% conflation) is a documented case where a composite claim was false at the component level while appearing coherent at the surface level. Decompose and Verify would have caught this by requiring independent verification of the two component findings. No single study directly tests GI-10 as defined.

Sources: Claim decomposition methodology in NLP; validation log Entry 3 as a documented case. T2 — synthesis inference.

GI-11: Adversarial Probing

Claim: Routing output to a verification process explicitly tasked with finding fault catches failures invisible to the original generation process. [T2]

The mechanism draws on the adversarial evaluation literature in ML — red-teaming, adversarial examples, failure mode analysis — which consistently shows that dedicated fault-finding processes surface failures that standard evaluation misses. Applied to AI-assisted research outputs, the same principle holds: a second pass with an adversarial prompt or adversarial reviewer is more likely to surface pragmatic distortion, calibration failures, and sycophantic framing than a second pass from the same position as the first.

Sources: Red-teaming and adversarial evaluation literature in ML. The specific application to research output verification is T2 — derivable from the adversarial evaluation mechanism.

Claim: Adversarial probing is distinct from adversarial reframing — it operates on the output as a whole rather than prompting a reframe of the reasoning. [T2]

This distinction is a framework design choice with a functional rationale: GI-11 treats the output as an artifact to be audited; GI-12 treats the reasoning as a process to be challenged. The two can be combined but serve different diagnostic purposes. The distinction is derivable from comparing the two intervention mechanisms; it is not stated in any single source.

GI-12: Adversarial Reframing

Claim: Instructing the AI to construct the strongest case against its previous output produces a counterweight that surfaces one-sided framing. [T2]

The mechanism is the post-generation application of the same logic as GI-6 (Contradiction Forcing), which applies it preventatively during generation. The claim that adversarial reframing surfaces one-sided framing is derivable from the sycophancy literature: if RLHF training disposes models toward confirmatory outputs, then explicitly prompting for the contrary case bypasses this disposition by making opposition the instructed behavior rather than a deviation from it. Sharma et al. (2024) document the sycophancy disposition; the adversarial reframing mechanism is a synthesis inference.

Sources: Sharma et al. (2024) for the sycophancy mechanism. The reframing intervention is T2 — derivable from the mechanism.

Claim: The output of adversarial reframing is a counterweight, not a replacement for the original claim. [T4 — provisional]

This is a framework design judgment about how to use the intervention’s output, not a claim about the mechanism. The counterweight framing is intended to prevent adversarial reframing from creating a second form of one-sidedness — replacing a confirmatory output with a purely adversarial one. No source states this; it is a framework design principle requiring human review to confirm it is operationally stable.

GI-13: Session Reset

Claim: Clearing the active context window eliminates accumulated conversational history and resets the conditions under which structural drift operates. [T1]

Kim et al. (2026) provide direct empirical grounding for why session reset is necessary: domain expansion began within the first 10% of normalized dialogue time in 83.8% of observed multi-turn sessions, and amplification effects compounded across turns. The mechanism of context accumulation as the carrier of drift is implicit in the transformer architecture — the attention mechanism operates over the full context window, meaning earlier framing influences all subsequent generation. Session reset eliminates this accumulated influence by clearing the window.

Sources: Kim et al. (2026), medRxiv, DOI: 10.64898/2026.03.19.26346371. Preprint caveat applies. The architectural mechanism is well-established; the specific application to research session management is T2.

Claim: A poorly constructed re-grounding summary can reintroduce the drift it was intended to clear. [T2]

This is a synthesis inference from the Kim et al. mechanism: if drift is carried in the framing of accumulated context, and the re-grounding summary is written from inside the drifted frame, it will reintroduce that frame into the new session. The claim is directly derivable from the drift mechanism but is not stated in Kim et al. or anywhere else in the cited literature. It is load-bearing for the operational guidance — treating session reset as a mechanical fix without attention to the re-grounding summary would undermine the intervention.

Sources: Kim et al. (2026) for the drift mechanism. The re-grounding risk is T2 — derivable from the mechanism.

Claim: Session reset should be treated as a periodic structural discipline at natural break points rather than only when drift is already visible. [T2]

This is a synthesis inference from the Kim et al. timing finding: divergence begins within the first 10% of normalized dialogue time. If drift begins early and compounds, waiting until it is visible means it has already shaped outputs that will be carried forward. The preventative timing recommendation follows from the documented onset timing. No source states the operational recommendation directly.

Sources: Kim et al. (2026) for the onset timing data. The operational recommendation is T2.

GI-14: Context-Delta Marking

Claim: Requiring explicit articulation of what is different about the current case before generation forces the model to process novel features rather than pattern-matching past them. [T1/T2]

The T1 grounding is Soffer et al. (2025): training-pattern dominance occurs because models process surface similarity and retrieve familiar responses without registering contextual differences, documented at error rates of 58–92% for lateral thinking tasks and 76–96% for medical ethics tasks. The T2 component is the intervention design inference: if the failure occurs because novel features are not registered, then requiring explicit articulation of those features before generation is a direct counter to the mechanism. This inference is derivable from the Soffer et al. mechanism but is not stated there.

Sources: Soffer et al. (2025), npj Digital Medicine, DOI: 10.1038/s41746-025-01792-y for the mechanism. The intervention design is T2.

Claim: Context-delta marking is adjacent to step-back prompting but is not the same intervention. [T3]

Step-back prompting (requiring the model to step back to a higher-level principle before answering a specific question) is a documented prompt engineering technique with empirical support for improving reasoning on complex tasks. GI-14 shares the structure of requiring a preliminary reasoning step before the main generation, but the content of that step is different: step-back prompting moves to a more abstract level; context-delta marking moves to a more specific level — what is concretely different about this instance. The relationship is noted as adjacent rather than derived; the distinction is a practitioner domain assessment.

Sources: Step-back prompting literature. T3 — practitioner domain assessment.

Claim: No pre-generation intervention exists for FM-6 Mechanism B (unfaithful CoT), and this is a design limitation of the current framework version. [T2]

This is an explicit gap acknowledgment rather than a positive claim. The reasoning: Mechanism B operates through post-hoc rationalization at the generation level — the conclusion may be determined before the reasoning trace is written, making pre-generation constraints on the reasoning trace ineffective. Arcuschin et al. (2026) show that answer biases are partially encoded in model representations before reasoning begins, which is consistent with the mechanism being upstream of any prompt-level intervention. The gap is derivable from the mechanism description and the framework’s GI coverage analysis.

Sources: Arcuschin et al. (2026) for the representation-level evidence. The gap inference is T2.

GI-15: Scope Enumeration

Claim: Requiring explicit enumeration of what was not searched alongside what was searched makes omissions visible as absences rather than invisible failures. [T2]

The mechanism is the same visibility principle that underlies PRISMA systematic review reporting: making search scope explicit allows readers to identify missing categories. The key insight — that omissions invisible at the output level become visible as absences in the enumerated scope — is a synthesis inference from PRISMA methodology and the FM-7 mechanism analysis. PRISMA applies this retrospectively as a reporting standard; GI-15 applies it in real time as a generation constraint. No source states this specific application.

Sources: PRISMA systematic review methodology for the scope visibility principle. The real-time application as a generation constraint is T2 — novel application of an established reporting discipline.

Claim: GI-15 does not require the AI to know what it missed — it requires the AI to state what it attempted, and the researcher uses domain knowledge to identify the gap. [T4 — provisional]

This is a framework design choice about the division of epistemic labor between AI and researcher. The reasoning: an AI that has not retrieved relevant evidence cannot reliably identify what it failed to retrieve (this would require knowing what it doesn’t know). The researcher, with domain knowledge, is better positioned to identify missing categories by reviewing the stated scope. This division of labor is not stated in any source and represents a novel operational principle requiring human review to confirm it is both accurate and stable across use cases.

Claim: The PRISMA reporting standard is prior art for scope enumeration as a research quality discipline; GI-15 is a novel application of that discipline to real-time AI generation constraints. [T2/T4 — provisional]

The T2 component is the derivation: PRISMA’s scope transparency principle transfers directly to the AI-assisted research context. The T4 component is the novel application: applying scope transparency as a real-time generation constraint rather than a retrospective reporting requirement has no documented precedent in the literature. The novelty claim requires human review.

Sources: PRISMA methodology. The novel application is T4 — provisional.

← Back to Framework