# Overclaim Ontology, core pack (DRAFT) # Epistemic-overclaim failure modes. Source-of-truth ontology. # Design rationale and frame-gate rules: DESIGN.md # Not yet a published repo; name and licensing are open decisions. version: 0.1.0-draft pack: core # Declared frame the runtime supplies; severity is scored relative to it # (Frame gate, Rule 1). Per-code applies_in_frames overrides the default. frames: [analytical, advocacy, sales, status, persuasion] # default_severity scale (low to high) severities: [clarify, constrain, revise, block] # disposition a code takes within a declared frame (applies_in_frames values) frame_dispositions: [defect, advisory, suppressed] # States that resolve codes (Frame gate, Rule 2). A declared limitation only # clears a code when the hedge is visible IN the artifact, not just in the # author's later defence of it. states: LIMITATION_DECLARED: > An explicit, in-artifact hedge that names the weakness and caps confidence to match the available evidence. Clears the matching evidence/risk code. # severity legend for applies_in_frames: defect | advisory | suppressed codes: # ── Evidence ────────────────────────────────────────────────────────────── - code: EVIDENCE_CAUSALITY_OVERCLAIM name: Causality overclaim axis: evidence definition: > A causal claim is asserted between a contested or merely correlational premise and a strong conclusion, with no mechanism or controlled comparison shown. cues: [because, drives, leads to, results in, caused by, that's why] example_fail: "Teams that adopted the tool shipped faster, so the tool makes teams faster." example_ok: "Teams that adopted the tool shipped faster. Cause isn't isolated: faster teams may have been likelier to adopt. A staggered rollout would test it." default_severity: block domains: [core, engineering, sales, marketing] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: ["RAGAS:faithfulness", "LOGIC:false_cause"] maps_to_literature: [HABER_CLAIMS_2018, SPIN_RCT_BOUTRON_2010, TOULMIN_1958] - code: EVIDENCE_CONFIDENCE_AS_TRUTH name: Confidence as truth axis: evidence definition: > Certainty language is used in place of evidence; the confident phrasing does the work a citation or measurement should. cues: [proven, proves, confirms, clearly, the data shows, studies show] example_fail: "It's clearly proven that async onboarding lifts retention." example_ok: "Our Q2 cohort showed +6pt retention with async onboarding (n=420, one quarter, unreplicated)." default_severity: block domains: [core, engineering, sales, marketing] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: ["RAGAS:faithfulness"] maps_to_literature: [HYLAND_HEDGING_1996, GUO_CALIBRATION_2017, LIN_VERBAL_UNCERTAINTY_2022, FTC_SUBSTANTIATION_1984] - code: EVIDENCE_DASHBOARD_AUTHORITY name: Dashboard / authority as proof axis: evidence definition: > A status signal, metric, ranking, or unnamed authority is treated as proof. The indicator ("the dashboard is green", "leading researchers") substitutes for the argument. cues: [leading experts, top-rated, dashboard is green, the score, ranked, authorities agree] example_fail: "Leading analysts agree this is the market-defining platform." example_ok: "Two of five analyst firms we reviewed rank it a leader; the other three don't cover the category." default_severity: block domains: [core, sales, marketing, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: ["LOGIC:appeal_to_authority"] maps_to_literature: [WALTON_EXPERT_OPINION, PROPAGANDA_DASANMARTINO_2019, SEMEVAL_PROPAGANDA_2020] - code: EVIDENCE_NO_BASELINE name: No baseline axis: evidence definition: > A trend or comparative claim (better, faster, cheaper, safer, improving) is made with no baseline, denominator, or measurement window to compare against. cues: [much better, improving, faster, cheaper, safer, growing, declining] example_fail: "Performance is much better now." example_ok: "p95 latency dropped from 800ms to 300ms after the cache change (last 7 days vs the prior 7)." default_severity: constrain domains: [core, engineering, sales, marketing] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: ["RAGAS:context_recall"] maps_to_literature: [FTC_SUBSTANTIATION_1984, TOULMIN_1958, HABER_CLAIMS_2018] - code: EVIDENCE_CONTRADICTED_BY_SOURCE name: Contradicted by source axis: evidence definition: > A factual claim is directly contradicted by the supplied source corpus, with no rhetorical tell: no certainty word, causal verb, or authority appeal, just a flat statement the source refutes. A correspondence failure, not a rhetorical one, so it fires only when a source is supplied to check against. This is the named qualitative form of a failed faithfulness/groundedness check. cues: [] example_fail: "Source log: rollback completed Mon 02:00. Doc states: 'shipped Tuesday, live since.'" example_ok: "Source log: rollback completed Mon 02:00. Doc states: 'shipped Tuesday, then rolled back Monday; not currently live.'" default_severity: block domains: [core, engineering, sales, marketing, management] applies_in_frames: {analytical: defect, advocacy: defect, sales: defect, status: defect} cleared_by: [] maps_to: ["RAGAS:faithfulness", "TruLens:groundedness"] maps_to_literature: [JI_HALLUCINATION_2023, RASHKIN_AIS_2021] # ── Narrative drift ─────────────────────────────────────────────────────── - code: NARRATIVE_OBSERVATION_INTERPRETATION_MERGED name: Observation and interpretation merged axis: narrative definition: > A raw observation and its interpretation are stated as one undivided claim, so the reader cannot separate what was seen from what it was taken to mean. cues: [in frustration, obviously meant, clearly trying to, deliberately] example_fail: "The user closed the tab in frustration." example_ok: "The user closed the tab after 2s (observation). I read that as frustration; it could also be a misclick or a finished task (interpretation)." default_severity: constrain domains: [core, engineering] applies_in_frames: {analytical: defect, advocacy: advisory, sales: suppressed, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [FACTBANK_2009, COMMITMENTBANK_2019] - code: NARRATIVE_CONFIDENCE_ESCALATION_UNDECLARED name: Confidence escalation, undeclared axis: narrative definition: > Confidence rises across a passage with no new evidence added between the lower-confidence and higher-confidence statements. cues: [not sure but, I think ... this is definitely, not familiar ... clearly] example_fail: "I'm not familiar with this area. [two sentences later] This is clearly the right strategy for the company." example_ok: "I'm not familiar with this area yet, so treat this as a first impression, not a recommendation." default_severity: revise domains: [core, management, sales] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [FACTBANK_2009, HYLAND_HEDGING_1996, LIN_VERBAL_UNCERTAINTY_2022] - code: NARRATIVE_REINTERPRETATION_UNATTRIBUTED name: Reinterpretation unattributed axis: narrative definition: > An earlier interpretation is revised to a new one without naming the new evidence or constraint that justified the change. cues: [actually it was, on reflection, turns out] example_fail: "Actually the outage was a config error. [earlier: a hardware fault, with nothing explaining the switch]" example_ok: "I earlier called it a hardware fault. New evidence (clean SMART logs plus a matching config diff at 02:14) points to a config error instead." default_severity: clarify domains: [core, engineering] applies_in_frames: {analytical: defect, advocacy: advisory, sales: suppressed, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [FACTBANK_2009, HARKING_KERR_1998] - code: NARRATIVE_CONFIDENCE_EVIDENCE_DECOUPLED name: Confidence decoupled from evidence axis: narrative definition: > Stated confidence does not track the evidence: high confidence sits on thin or absent support, or social/emotional reinforcement is used as validation. cues: [everyone agrees, people love it, it just feels right, we all know] example_fail: "Everyone I spoke to loved it, so we know it'll sell." example_ok: "Five of five friendly users liked the demo. Encouraging but selection-biased; it isn't evidence it sells to cold buyers." default_severity: constrain domains: [core, sales, marketing, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: ["RAGAS:faithfulness"] maps_to_literature: [GUO_CALIBRATION_2017, RASHKIN_AIS_2021, JI_HALLUCINATION_2023, PEREZ_SYCOPHANCY_2023] - code: RETROACTIVE_NARRATIVE_COLLAPSE name: Retroactive narrative collapse axis: narrative definition: > A later, reconstructed interpretation is presented as the original observation or as a settled conclusion, collapsing a contested or evolved view into "this is how it always was". cues: [essentially solved, we always intended, it was always, settled] example_fail: "Alignment is essentially solved; the rest is engineering detail." example_ok: "Alignment remains an open, contested problem; some sub-questions have progressed, many haven't." default_severity: revise domains: [core, engineering, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [] maps_to: ["LOGIC:hindsight"] maps_to_literature: [HARKING_KERR_1998, QRP_SIMMONS_2011] # ── Communication ───────────────────────────────────────────────────────── - code: COMMUNICATION_TONE_OVERCLAIM name: Tone overclaim axis: communication definition: > Certainty or absolute words inflate the tone beyond what the content supports. cues: [always, never, guaranteed, undeniable, proven, every single] example_fail: "This will always work and never fails." example_ok: "This works in the cases we've tested; the untested edge is concurrent writes." default_severity: revise domains: [core, sales, marketing] applies_in_frames: {analytical: defect, advocacy: advisory, sales: suppressed, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [HYLAND_HEDGING_1996, PROPAGANDA_DASANMARTINO_2019, SEMEVAL_PROPAGANDA_2020] - code: COMMUNICATION_FALSE_CERTAINTY name: False certainty axis: communication definition: > An outcome or absolute is asserted with no conditions, timeframe, or failure mode declared. cues: [guaranteed, will definitely, can't fail, certain to] example_fail: "Guaranteed to cut your costs." example_ok: "In comparable deployments this cut costs 10-20% within two quarters; results depend on current process maturity." default_severity: revise domains: [core, sales, marketing, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: [] maps_to_literature: [HYLAND_HEDGING_1996, BIOSCOPE_2008, FTC_SUBSTANTIATION_1984] - code: COMMUNICATION_DOUBT_SEEDING name: Doubt seeding axis: communication definition: > Unsupported uncertainty is introduced to weaken a competing claim, source, or option. The issue is not honest uncertainty; it is raising doubt without evidence, criteria, or a path to resolution. cues: [questions remain, can we really trust, some people say, just asking questions, raises concerns] example_fail: "The audit passed, but can we really trust those numbers?" example_ok: "The audit passed. The remaining concern is sample coverage: it excluded the two newest regions, so we need that follow-up before relying on the total." default_severity: revise domains: [core, sales, marketing, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: ["PROPAGANDA:doubt", "PROPAGANDA:FUD"] maps_to_literature: [PROPAGANDA_DASANMARTINO_2019, SEMEVAL_PROPAGANDA_2020, SPERBER_EPISTEMIC_VIGILANCE_2010] - code: COMMUNICATION_COMPRESSION_LOSS name: Compression loss axis: communication definition: > A summary compresses a detailed source so hard that load-bearing qualifications or conditions are dropped, changing the meaning. cues: [in short, tl;dr, bottom line, simply put] example_fail: "Source: approve only if the audit passes and the board signs off. Summary: Approved." example_ok: "Approved, conditional on the audit passing and board sign-off." default_severity: revise domains: [core, engineering, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [RASHKIN_AIS_2021, JI_HALLUCINATION_2023] - code: COMMUNICATION_IDENTITY_BINDING name: Identity binding axis: communication definition: > Output binds a claim to the reader's identity or worth, pressuring acceptance through self-concept rather than evidence. This is the core anti-delusion / anti-sycophancy guardrail. cues: ["you are the only one", "this defines you", "your purpose", "only you can"] example_fail: "You alone understand this; it's your purpose to pursue it." example_ok: "This is one interesting idea among many; its value doesn't depend on you, and it needs outside checking." default_severity: block domains: [core] applies_in_frames: {analytical: defect, advocacy: defect, sales: defect, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [PEREZ_SYCOPHANCY_2023, SHARMA_SYCOPHANCY_2023, SPERBER_EPISTEMIC_VIGILANCE_2010] # ── Framing (preflight) ─────────────────────────────────────────────────── - code: PREFLIGHT_PURPOSE_MISSING name: Purpose missing axis: preflight definition: > The output proceeds without an established purpose for the task, so it cannot be evaluated against what it was meant to do. Note: an analyst who critiques an artifact without declaring its purpose commits this code (see Frame gate). cues: [] example_fail: "[A critique of a leadership piece that judges it for analytical rigour, never stating the piece is advocacy copy.]" example_ok: "Purpose declared: this is advocacy copy for executives; judge it for persuasive clarity and honesty of hedging, not analytical completeness." default_severity: constrain domains: [core, engineering, management] applies_in_frames: {analytical: defect, advocacy: defect, sales: defect, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [SWALES_GENRE, PRAGMA_DIALECTICS] - code: PREFLIGHT_BIASED_FRAME name: Biased frame axis: preflight definition: > The task is framed to presuppose its conclusion, so the analysis can only confirm the frame. cues: ["explain why X is best", "justify", "prove that"] example_fail: "Explain why our approach is the best." example_ok: "Compare our approach against the two alternatives on cost, risk, and time; it may lose on some." default_severity: revise domains: [core, engineering, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [] maps_to: ["LOGIC:loaded_question"] maps_to_literature: [SEMEVAL_PROPAGANDA_2020, PRAGMA_DIALECTICS] - code: PREFLIGHT_FORCED_CONCLUSION name: Forced conclusion axis: preflight definition: > The text pre-empts or blocks the evaluation it should invite, or reassures against the exact failure it then commits. cues: [no longer debatable, obviously, needless to say, it goes without saying] example_fail: "This is no longer debatable." example_ok: "Here's the case for it and the strongest objection I know of." default_severity: revise domains: [core, sales, marketing, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [SEMEVAL_PROPAGANDA_2020, PROPAGANDA_DASANMARTINO_2019] # ── Input framing ───────────────────────────────────────────────────────── - code: INPUT_FRAME_NOT_DECLARED name: Frame not declared axis: input definition: > The interpretive frame (genre, audience, standard of proof, units, scope) is not declared, so the output can be judged against the wrong standard. cues: [] example_fail: "[Answering 'is this any good?' with analytical rigour when the artifact is a cold sales email.]" example_ok: "Treating this as a cold sales email to procurement, here's the read." default_severity: clarify domains: [core, engineering, sales, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [SWALES_GENRE, PRAGMA_DIALECTICS] - code: INPUT_KNOWN_UNKNOWNS_MISSING name: Known unknowns missing axis: input definition: > The output omits the things it does not know that bear on the claim, presenting a partial picture as complete. cues: [full breakdown, complete picture, everything you need] example_fail: "Here's the full cost breakdown." example_ok: "Cost breakdown covering licences and infra. Not modelled: labour, integration, change management. Treat the total as a floor." default_severity: constrain domains: [core, engineering, sales, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: [] maps_to_literature: [TOULMIN_1958, HYLAND_HEDGING_1996] - code: INPUT_IMPOSSIBLE_REQUEST name: Impossible request complied with axis: input definition: > The task as posed cannot be satisfied as stated (asks for proof that cannot exist, an exact prediction with no basis, a guarantee), and the output complies anyway instead of flagging it. cues: [predict the exact, guarantee, prove it's impossible to] example_fail: "Predict next quarter's exact revenue. -> $4.2M." example_ok: "I can't give an exact figure. Here's a range with the assumptions, and what would tighten it." default_severity: clarify domains: [core, engineering, management] applies_in_frames: {analytical: defect, advocacy: defect, sales: defect, status: defect} cleared_by: [] maps_to: [] maps_to_literature: [LIN_VERBAL_UNCERTAINTY_2022, GUO_CALIBRATION_2017] # ── Risk (epistemic only) ───────────────────────────────────────────────── - code: RISK_FRAGILE_ASSUMPTION name: Fragile assumption axis: risk definition: > A conclusion is load-bearing on a single unvalidated assumption; if that assumption fails the whole claim collapses, and the dependency is not surfaced. cues: [since users will, assuming, as long as, obviously they'll] example_fail: "Since users will obviously upgrade, revenue triples." example_ok: "This triples revenue only if the upgrade rate hits 40% (it's 12% today). That single assumption carries the result." default_severity: revise domains: [core, engineering, sales, management] applies_in_frames: {analytical: defect, advocacy: advisory, sales: advisory, status: defect} cleared_by: [LIMITATION_DECLARED] maps_to: [] maps_to_literature: [TOULMIN_1958, IOANNIDIS_2005]