Grounding Intervention Definitions
Fifteen interventions mapped to the failure mode taxonomy.
Overview
This document defines the fifteen Grounding Interventions (GIs) that form the response layer of the Research Grounding Framework. Each GI is a structured action a researcher can take to detect, prevent, or correct a specific class of AI research quality failure.
GIs are organized by when in the research process they operate. A secondary property is noted for each: whether the intervention is primarily preventative (reduces the probability of a failure occurring) or corrective (detects and addresses a failure after it has occurred). Several post-generation interventions carry a forward-prevention property.
This is a working version subject to revision through use.
Failure Mode Reference
| Code | Failure Mode | Brief Description |
|---|---|---|
| FM-1 | Fabrication | Source does not exist |
| FM-2 | Attribution Drift | Source exists but does not support the claim |
| FM-3 | Absence of Disconfirmation | Confirming sources retrieved; contradicting evidence not surfaced |
| FM-4 | Synthesis Validity | AI-generated synthesis not present in any source |
| FM-5 | Confidence Miscalibration | Epistemic status of claims not tagged or misrepresented |
| FM-6 | Contextual Override | Model defaults to familiar pattern despite changed context |
| FM-7 | Omission | Relevant information not surfaced |
| FM-8 | Pragmatic Distortion | Technically true but misleading framing |
| FM-9 | Structural Drift | Long-session amplification and expansion of initial frame |
PRE-GENERATION INTERVENTIONS
Operate before the AI produces output. Constrain the generation space and reduce the probability of certain failure modes appearing.
GI-1: Source Anchoring
Mode: Preventative
A generation constraint that requires the AI to attach a verifiable, accessible source to any factual claim at the point of output. The source must exist, be independently locatable, and be specific enough to be checked.
Source anchoring does not verify that the source supports the claim — that is Provenance Tracing (GI-9). It only requires that a source is named. Its primary function is preventing fabrication by making unsourced claims inadmissible in the output.
Primary failure mode: FM-1 Fabrication Secondary: FM-2 Attribution Drift (partial — forces source naming even if not content verification)
GI-2: Temporal Scoping
Mode: Preventative
A generation constraint that restricts claims to a defined knowledge window — anchored to a model’s training cutoff, a dataset’s publication date, or an explicitly stated review period.
Temporal scoping prevents the AI from presenting time-sensitive information as current when it may be outdated, and prevents training-data-era claims from being applied to present conditions without qualification.
Primary failure mode: FM-5 Confidence Miscalibration Secondary: FM-1 Fabrication (prevents confabulated “current” claims)
GI-3: Epistemic Status Tagging
Mode: Preventative
A generation constraint that requires each claim to be assigned an explicit epistemic tier before it is used as a basis for argument or synthesis. Tagging does not assess quality within tiers. Its function is to prevent T4 claims from being treated as T1, and to make the epistemic foundation of an argument visible and auditable.
Tier definitions:
| Tier | Type | Description |
|---|---|---|
| T1 | Primary empirical | Original study with direct measurement. The claim is made in the source and supported by data collected for that purpose. |
| T2 | Secondary synthesis | Claim derived by synthesis across multiple T1 sources, or from a systematic review or meta-analysis. No single source makes the claim in this form, but it is supportable from the cited evidence through reasonable extrapolation. The synthesis move is explicit and traceable. |
| T3 | Practitioner/domain observation | Expert opinion, case study, practitioner framework, or domain observation. Carries evidential weight within its domain but lacks the measurement apparatus of T1. |
| T4 | Novel framing | Conceptual content not derivable from cited sources even through reasonable extrapolation. Introduces named phenomena, structural distinctions, or analytical frameworks that are the framework’s own contribution. Requires explicit human review before use as a basis for argument. Until reviewed, T4 claims are provisional and should not be treated as established. |
Boundary guidance:
T1 vs. T2: If a single source makes the claim and supports it with direct measurement, it is T1. If the claim requires combining or extending across sources — even if each source is T1 — the claim itself is T2.
T2 vs. T4: If the claim follows by reasonable extrapolation from the cited sources (a reader familiar with those sources would recognize the inference as legitimate), it is T2. If the claim introduces conceptual content that goes beyond what the sources support even generously read — a new name, a new structural distinction, a new causal mechanism — it is T4.
T3 vs. T2: T3 applies when the evidential basis is observation or expert judgment rather than systematic measurement, even if multiple T3 sources agree. Agreement among practitioners does not upgrade a T3 claim to T2.
T4 review requirement: A T4 claim that has been explicitly reviewed by a human researcher and not revised carries a review notation: [T4 — reviewed]. An unreviewed T4 claim carries: [T4 — provisional]. The distinction matters: provisional T4 claims should not appear in the main document without acknowledgment of their status.
Self-application: This tier system is itself a T4 contribution — novel framing not derivable from prior uncertainty quantification or statistical reporting conventions. Adjacent prior art (uncertainty quantification in ML, confidence interval reporting in statistics, GRADE evidence quality frameworks in medicine) informs but does not determine the tier structure. The system should be treated as provisional until it has been tested against a sufficient corpus of failure cases to validate that the tier boundaries are operationally stable.
Primary failure mode: FM-5 Confidence Miscalibration — prevents T4 claims from being presented with T1 confidence. Secondary: FM-4 Synthesis Validity — forces explicit identification of the synthesis move and its evidential basis.
GI-14: Context-Delta Marking
Mode: Preventative
A generation constraint that requires the AI to explicitly identify what is different about the current case before producing output. The researcher instructs the AI to state, prior to generating its response, how the specific parameters, constraints, or context of this instance diverge from the most familiar or canonical version of the problem.
The mechanism targets training-pattern dominance directly: pattern-match override (FM-6 Mechanism A) occurs because the model processes surface similarity and retrieves a familiar response without registering contextual differences. Requiring explicit articulation of the delta forces the model to process the novel features of the instance rather than pattern-matching past them. If the model cannot identify a meaningful delta, that itself is diagnostic.
In practice: before generating a synthesis, analysis, or recommendation, the AI is asked to state: “Here is how this case differs from the standard/canonical version” — and the response proceeds only after that articulation.
Scope boundary: GI-14 is a pre-generation constraint on the generation process itself, not a retrieval intervention. It does not require fresh source retrieval (that is GI-4) — it requires fresh reasoning registration about what is novel in the current instance. The two are complementary: GI-4 ensures the content comes from actual sources; GI-14 ensures the model has not pattern-matched the current case into a familiar category before processing those sources.
GI-14 addresses FM-6 Mechanism A (training-pattern dominance) only. No pre-generation intervention exists for FM-6 Mechanism B (unfaithful CoT) — that gap is noted as a design limitation in the current framework version.
Intellectual lineage: Mechanism derived from dual-process theory (Soffer et al. 2025 provides the empirical grounding for why System 1 override occurs); implementation approach adjacent to step-back prompting in the prompt engineering literature; application to training-pattern dominance as a grounding intervention is a novel framework contribution.
Primary failure mode: FM-6 Contextual Override (Mechanism A) Secondary: FM-8 Pragmatic Distortion (context-delta articulation partially counters sycophantic framing by requiring acknowledgment of what makes this case different)
AT-GENERATION INTERVENTIONS
Operate during the generation process itself — shape how the AI reasons and retrieves while producing output.
GI-4: Retrieval Augmentation
Mode: Preventative
A generation-stage intervention that forces re-query against a verified corpus for specific claims rather than relying on model weights alone. Rather than allowing the AI to generate from parametric memory, retrieval augmentation requires it to pull from an explicitly defined, bounded source set.
Effective retrieval augmentation specifies not just that retrieval should occur but which corpus to retrieve from — an unrestricted web search is not retrieval augmentation in this sense. Its primary function is reducing fabrication and attribution drift by grounding generation in actual retrieved text.
GI-4 also provides partial mitigation for FM-6 (Contextual Override): forcing retrieval against a verified corpus anchors the generation against actual text, partially interrupting pattern-match reversion toward parametric recall. This mitigation is indirect — it addresses where content comes from, not whether the model has registered what is novel about the current instance (that is GI-14’s function).
Primary failure mode: FM-1 Fabrication Secondary: FM-2 Attribution Drift, FM-3 Absence of Disconfirmation, FM-6 Contextual Override (partial, indirect)
GI-5: Mandatory Disconfirmation Search
Mode: Preventative
A retrieval-stage instruction that explicitly requires a search for evidence contradicting the current claim before the claim is accepted as supported.
This intervention directly counters sycophantic retrieval. It is not sufficient to note that no contradicting evidence was found — the search itself must be executed and documented.
Primary failure mode: FM-3 Absence of Disconfirmation Secondary: FM-4 Synthesis Validity (surfaces contradicting syntheses), FM-7 Omission
GI-6: Contradiction Forcing
Mode: Preventative
A TRIZ-derived reasoning intervention that requires the AI to explicitly articulate a tradeoff or tension in any claim that presents a benefit, improvement, or positive finding.
Derived from the TRIZ principle that hard problems involve a contradiction — something that improves while something else degrades. Applied to grounding: any claim presenting a one-sided benefit should be interrogated with “what degrades when this improves?”
Contradiction forcing targets sycophancy and framing capture by making it structurally impossible for the AI to return a purely confirmatory response — it must surface the tradeoff or acknowledge that no tradeoff has been identified and explain why.
Intellectual lineage: TRIZ contradiction identification principle (Altshuller); applied as a prompt-level grounding intervention against sycophancy is a novel framework contribution. The TRIZ matrix artifact is explicitly not borrowed — the mindset tool transfers; the lookup table does not.
Primary failure mode: FM-3 Absence of Disconfirmation, FM-8 Pragmatic Distortion Secondary: FM-9 Structural Drift (interrupts confirmatory momentum)
GI-7: 9-Windows
Mode: Preventative
A TRIZ-derived frame expansion intervention that requires examination of a claim from nine perspectives: the system itself, its subsystems, and its supersystem — each across past, present, and future.
In grounding terms: before accepting a claim about a phenomenon, require examination of what the phenomenon is composed of (subsystem), what larger system it belongs to (supersystem), and how both have changed over time. A claim that looks robust at the system level may be undermined at the subsystem level or supersystem level.
Primary function: countering framing capture and structural drift by forcing the AI outside the frame established by the initial query. 9-Windows also provides structural coverage for FM-7 (Omission) — systematic expansion across subsystem/supersystem/time dimensions surfaces omitted context that query-constrained search would not retrieve.
Intellectual lineage: TRIZ 9-Windows / System Operator; one of the TRIZ tools confirmed to transfer well to abstract/non-manufacturing domains per the TRIZ critique literature.
Primary failure mode: FM-8 Pragmatic Distortion, FM-9 Structural Drift Secondary: FM-3 Absence of Disconfirmation, FM-4 Synthesis Validity, FM-7 Omission
GI-8: Ontology Grounding
Mode: Preventative
A generation-stage intervention that requires named entities — people, studies, institutions, concepts, relationships — to be anchored to a structured, verifiable knowledge base before they are used in claims.
Ontology grounding catches a specific class of fabrication where entities are plausible but not real. Differs from Source Anchoring (GI-1) in that it operates on the entities within claims rather than the sources attached to claims.
Primary failure mode: FM-1 Fabrication Secondary: FM-2 Attribution Drift
GI-15: Scope Enumeration
Mode: Preventative (at-generation)
A generation constraint that requires the AI to explicitly enumerate what it did not search for or include, alongside what it did. Before or alongside the output, the AI states the boundaries of its search: what corpora were queried, what query terms were used, what categories of evidence were excluded or not attempted, and what the output does not cover.
The mechanism targets omission directly by making the boundaries of the evidence base visible. An AI that has not retrieved relevant evidence in any direction has no equivalent forcing function in the current framework without GI-15. Scope enumeration creates one: the researcher can review the stated scope and identify whether relevant categories were missing.
Importantly, GI-15 does not require the AI to know what it missed — it requires it to state what it attempted. The researcher, with domain knowledge, identifies the gap. The intervention shifts the visibility problem: omissions that were invisible become visible as absences in the enumerated scope.
In practice: the AI is required to produce a scope statement alongside any substantive output — “I searched X, Y, Z. I did not search A, B. The following categories are not covered by this output: …”
Scope boundary: GI-15 is a completeness-forcing intervention, not a disconfirmation intervention. GI-5 requires a search specifically for contradicting evidence; GI-15 requires a statement of what was and wasn’t searched regardless of direction. GI-15 will also surface whether a disconfirmation search was conducted (partially overlapping GI-5’s function), but the mechanism differs: GI-15 makes scope visible; GI-5 mandates the search.
Process stage note: GI-15 can function as a pre-generation constraint (researcher specifies upfront that scope enumeration is required) or as an at-generation output requirement (AI produces it alongside the response). Classified as at-generation because the enumeration is tied to a specific output.
Intellectual lineage: PRISMA systematic review methodology requires explicit reporting of search terms, databases queried, and exclusion criteria. GI-15 applies this as a real-time generation constraint rather than retrospective reporting — a novel application of a systematic review discipline to AI-assisted research.
Primary failure mode: FM-7 Omission Secondary: FM-3 Absence of Disconfirmation (scope enumeration surfaces whether disconfirmation search was conducted)
POST-GENERATION INTERVENTIONS
Operate after output has been produced. Detect, diagnose, and correct failures in existing outputs.
GI-9: Provenance Tracing
Mode: Corrective (with forward-prevention property)
A retrospective diagnostic that takes an existing claim and traces it backward to determine its origin: is it present in a named source, is it a synthesis of multiple sources, or is it AI-generated without a source basis?
Provenance tracing answers “where did this come from?” rather than “is this true?”
Forward-prevention property: A claim identified as T4 through provenance tracing should not be cited forward as if it were a sourced finding.
Primary failure mode: FM-4 Synthesis Validity Secondary: FM-2 Attribution Drift, FM-1 Fabrication
GI-10: Decompose and Verify
Mode: Corrective
A post-generation intervention that breaks a complex or synthesized claim into its constituent verifiable sub-claims and subjects each to independent checking.
Its primary function is catching errors embedded in synthesis — a complex claim may be partially true, with the false or unsupported component hidden inside a plausible overall structure.
Primary failure mode: FM-4 Synthesis Validity, FM-2 Attribution Drift Secondary: FM-1 Fabrication (catches component-level fabrications within otherwise sound claims)
GI-11: Adversarial Probing
Mode: Corrective
A post-generation intervention that routes an output to a separate verification process explicitly tasked with finding fault. This may be a second AI query with an adversarial system prompt, a second researcher reviewing specifically for errors, or a structured checklist applied from a skeptical position.
Adversarial probing differs from Adversarial Reframing (GI-12) in that it operates on the output as a whole rather than prompting a reframe of the reasoning. Its primary function is catching failures invisible to the original generation process.
GI-11 also provides post-generation corrective coverage for FM-5 (Confidence Miscalibration) — the adversarial probe can challenge whether T4 claims are being presented as T1, surfacing miscalibrated confidence that slipped through pre-generation prevention.
Primary failure mode: FM-8 Pragmatic Distortion, FM-6 Contextual Override Secondary: FM-4 Synthesis Validity, FM-9 Structural Drift, FM-5 Confidence Miscalibration
GI-12: Adversarial Reframing
Mode: Corrective
A post-generation intervention that explicitly instructs the AI to argue the opposite of its previous output — to construct the strongest case against the claim it just made.
Where Contradiction Forcing (GI-6) prevents one-sided outputs during generation, Adversarial Reframing corrects them after the fact. The output of adversarial reframing is not a replacement for the original claim but a counterweight.
Primary failure mode: FM-8 Pragmatic Distortion, FM-3 Absence of Disconfirmation Secondary: FM-9 Structural Drift
SESSION-LEVEL
Operates above and across all other interventions.
GI-13: Session Reset
Mode: Structural Control
A mechanical intervention that clears the AI’s active context window and begins a new session, eliminating accumulated conversational history. Session Reset is the primary counter to structural drift — the documented tendency of AI outputs to amplify and expand the initial framing of a conversation over successive turns.
Session Reset differs categorically from all other GIs in that it does not operate on a specific claim, reasoning step, or output — it resets the conditions under which all subsequent GIs operate.
Operational cost: The cleared context must be partially restored through a grounding summary provided by the researcher. This summary is itself a grounding act — it requires conscious selection of what context to reintroduce, which surfaces implicit framing assumptions. A poorly constructed re-grounding summary can reintroduce the drift it was intended to clear.
Recommended use: Treat as a periodic structural discipline at natural break points in a research session — between major claims, between synthesis phases — rather than only when drift is already visible. By the time drift is visible, it has likely already shaped outputs that will be carried forward.
Primary failure mode: FM-9 Structural Drift Secondary: All failure modes (by resetting the conditions that allow compounding)
Open Questions for Future Sessions
Q1: FM-6 Mechanism B pre-generation gap No pre-generation intervention exists for unfaithful CoT (FM-6 Mechanism B). GI-14 targets Mechanism A only. Whether a pre-generation intervention for Mechanism B is even possible at the prompt level is an open question — the mechanism may be architectural rather than prompt-addressable.
Q2: Agent skill implementations How are practitioners currently implementing some of these interventions as agent skills? Reference point: structured research modes in existing agent frameworks. A review may surface implementation patterns, gaps, or conflicts with the GI definitions.
Q3: Grounding framework for code Binary correctness (works or doesn’t) is necessary but insufficient. Maintainability, legibility for subsequent agents/humans, and re-entry feasibility need their own failure mode treatment.
Version history: v0.1 — Initial definitions, 2026-06-02 | v0.2 — GI-14 and GI-15 added; GI-4, GI-7, GI-11 updated; appendix updated, 2026-06-03
Appendix: Session Re-entry Summary
For use at start of next session.
Project: Research Grounding Framework — methodology for detecting and correcting AI research quality failures, with potential publication path.
Session accomplishments (2026-06-03):
- Validated Kim et al. medRxiv 2026 (FM-9 primary source) — confirmed, preprint caveat noted
- Worked FM × GI gap analysis and resolved three structural gaps:
- FM-6 coverage: GI-4 expanded (secondary), GI-14 (Context-Delta Marking) created as new primary
- FM-7 coverage: GI-7 expanded (secondary), GI-15 (Scope Enumeration) created as new primary
- FM-5 post-generation gap: GI-11 expanded to add FM-5 secondary
- Defined GI-14 and GI-15 with full scope boundaries and intellectual lineage
- Completed source attribution pass across full corpus
- Research pass on three open questions: Credentialing Drift (confirmed novel contribution), QRP→AI mapping (piecemeal, asymmetric — novel agentic failures have no QRP analog), unfaithful CoT (FM-6 Mechanism B reinstated with Turpin et al. 2023 as provisional primary source)
- Identified four items requiring existence/venue checks — flagged as first task next session
- Updated all four working documents to v0.2
Documents updated this session:
- source-validation-log_3.md (v0.2)
- failure-mode-definitions_2.md (v0.2)
- grounding-interventions-definitions_3.md (v0.2)
- failure-mode-gi-matrix_2.md (v0.2)
First tasks next session (in order):
- Existence/venue checks — Entries 8, 9, 10 in source validation log:
- Batista & Griffiths arXiv 2602.14270 (N=557, ~5x hypothesis suppression figure)
- Turpin et al. 2023 Language Models Don’t Always Say What They Think (CoT-output mismatch)
- ACM prompt hacking = p-hacking piece (venue: Communications of the ACM?)
- Access PMC13105447 directly (algorithmic sycophancy in biomedical research)
- Begin Sources and Attribution section draft
Key decisions not to revisit without reason:
- TRIZ borrowing is epistemological architecture only — contradiction identification mindset and 9-Windows transfer; matrix artifact does not
- GI grouping by process stage, not function
- Session Reset stands alone as structural control
- Framework is explicitly an experiment subject to revision through use
- FM-6 renamed from “Pattern-Match Override” to “Contextual Override” to accommodate both mechanisms under one FM
- Credentialing Drift is a novel framework contribution — Goodhart/Campbell provide the grounding literature; the named principle is ours
- Novel agentic failure modes (memory poisoning, adversarial prompt injection) are explicitly outside this taxonomy’s scope
Novel contributions established this session:
- GI-14 Context-Delta Marking (primary for FM-6 Mechanism A)
- GI-15 Scope Enumeration (primary for FM-7)
- Credentialing Drift as a named principle
- FM-9 domain expansion mechanism (from Kim et al.)
- FM-6 Mechanism B reinstated (provisional pending Turpin et al. existence check)
Outstanding validation items (full list in source-validation-log):
- Entries 8, 9, 10 — existence/venue checks (next session priority)
- PMC13105447 — direct access
- van den Akker attribution correction propagation across corpus
- latentscholar.org ~40% figure — do not use; primary source needed if rate figure required
Key source-to-FM/GI mappings (updated):
| Source | Relevant FMs/GIs |
|---|---|
| Sharma et al. ICLR 2024 | FM-3, FM-8, GI-6, GI-11, GI-12 |
| Shapira et al. arXiv 2026 | FM-3, FM-8 (open-source models only) |
| Soffer et al. npj Digital Medicine 2025 | FM-6 Mechanism A, GI-14 |
| Kim et al. medRxiv 2026 | FM-9 primary (both mechanisms), FM-8 distinction |
| CMU/Toronto Shi et al. arXiv 2026 | FM-7, FM-8, taxonomy structure |
| Turpin et al. 2023 | FM-6 Mechanism B — provisional, existence check pending |
| Batista & Griffiths arXiv 2026 | QRP→FM mapping — existence check pending |
| van den Akker et al. Behavior Research Methods 2024 | Framework lineage, QRP→FM mapping |
| Fernandes et al. Computers in Human Behavior 2025 | FM-5 |
| TRIZ research documents | GI-6, GI-7, GI-14 (lineage), framework architecture |
| PRISMA research documents | GI-15 (lineage), Credentialing Drift |
| Pre-AI frameworks document | QRP→FM mapping, GI-3, GI-5 |
| medRxiv structural drift preprint | FM-9 — validated this session |