Reality Monitoring

Version

v1 · 2026-08-24 · History

Prime #

1113

Origin domain

Cognitive Science

Subdomain

memory and source attribution → Cognitive Science

Core Idea

Reality monitoring is the structural pattern in which a system that processes information from multiple kinds of sources must, for downstream use, attribute each piece of information to its source class — most fundamentally, internally generated versus externally perceived.^[1] The decisive structural commitment is that the judgment is made not at acquisition but at recall or use, on the basis of features that correlate with source: perceptual detail, schematic regularity, retrieval fluency, and the cognitive operations recorded with the trace. When the discrimination fails, internal content is treated as external — intrusions, hallucinations, confabulations, false attributions — or external content is treated as internal, as in the failure to recognize a borrowed idea. The prime is the existence of, and the mechanism for, this post-hoc source discrimination. It is a thin gate operating on a thick store: the items themselves commingle regardless of provenance, and what reality monitoring adds is a use-time classification that decides which source class each item belongs to and therefore how it should be weighted.

This skeleton recurs across substrates as a system whose internal stores commingle items from multiple source classes, a need at use-time to know which class each item came from, features that correlate with source class and form the basis for attribution, and characteristic failure modes when those features mislead. In cognition it is the source-monitoring framework distinguishing memories of perceived events from memories of imagined ones.^[2] In generative-AI safety it is the distinction between retrieved factual claims and parametric hallucinations, with retrieval-augmented systems tagging generations with source metadata so consumers can verify. In journalism it is tagging every claim with its source so editors and readers can apply appropriate confidence. In forensics it is eyewitness-versus-suggested-memory discrimination and chain of custody. In security it is distinguishing actions originated inside a trust boundary from those arriving outside it. Strip the substrate vocabulary and what remains is a commingled store, a use occasion, source-correlated features, an attribution mechanism, and a movable decision threshold. The pattern travels structurally, though its cognitive-source-monitoring vocabulary carries some normative residue — false memory, hallucination, plagiarism — and the judgment runs at use-time rather than acquisition.

How would you explain it like I'm…

Real Or A Dream?

Sometimes you wake up and have to figure out: did that really happen, or did I just dream it? Your brain keeps memories of things you saw and things you only imagined all mixed together. Later you check little clues to decide which ones were real.

Where Did It Come From?

Reality Monitoring is how your mind figures out where a piece of information came from — most importantly, whether you actually saw or heard it, or just thought it up yourself. The tricky part is your memory mixes real things and imagined things together, so you have to sort them out later, not when they first happened. You judge using clues: real memories usually have more vivid detail, while imagined ones feel more like a plain idea. When the clues fool you, you might 'remember' something that never happened, or forget that a cool idea was actually someone else's.

Tagging the Source

Reality Monitoring is the pattern where a system that takes in information from different kinds of sources must, when it later uses that information, label each piece by its source class — most fundamentally, internally generated versus externally perceived. The key structural point is that the judgment happens not at the moment you take the information in, but at recall or use, based on features that correlate with source: how much perceptual detail there is, how schematic or regular it feels, how easily it comes to mind, and what mental operations were recorded with it. When this discrimination fails, internal content gets treated as external (intrusions, hallucinations, confabulations) or external content gets treated as internal (failing to realize an idea was borrowed). It's a thin gate on a thick store: the items themselves commingle regardless of where they came from, and reality monitoring adds a use-time classification that decides each item's source class and therefore how much to trust it.

 

Reality Monitoring is the structural pattern in which a system that processes information from multiple kinds of sources must, for downstream use, attribute each piece of information to its source class — most fundamentally, internally generated versus externally perceived. The decisive commitment is that the judgment is made not at acquisition but at recall or use, on the basis of features that correlate with source: perceptual detail, schematic regularity, retrieval fluency, and the cognitive operations recorded with the trace. When the discrimination fails, internal content is treated as external — intrusions, hallucinations, confabulations, false attributions — or external content is treated as internal, as in failing to recognize a borrowed idea. The prime is the existence of, and the mechanism for, this post-hoc source discrimination: a thin gate operating on a thick store, where the items themselves commingle regardless of provenance and what reality monitoring adds is a use-time classification deciding which source class each item belongs to and therefore how it should be weighted. The skeleton recurs across substrates as a commingled internal store, a use-time need to know each item's source class, source-correlated features grounding attribution, and characteristic failure modes when those features mislead: the source-monitoring framework in cognition; retrieved facts versus parametric hallucinations in generative-AI safety, with retrieval-augmented systems tagging generations with source metadata; claim-by-claim source tagging in journalism; eyewitness-versus-suggested-memory discrimination and chain of custody in forensics; and inside-versus-outside-trust-boundary origin in security. Strip the vocabulary and what remains is a commingled store, a use occasion, source-correlated features, an attribution mechanism, and a movable decision threshold.

Structural Signature

the commingled store holding items from multiple source classes — the use-time occasion demanding an attribution — the source-correlated features available to the classifier — the attribution mechanism assigning each item a source class — the movable decision threshold trading off the two error types — the characteristic failures when the features mislead

A system exhibits this pattern when each of the following holds:

• A commingled store. Items from two or more source classes — most basically internally-generated versus externally-perceived — sit together in one store, undifferentiated by provenance at rest.
• A use-time attribution occasion. The source judgment is made not at acquisition but at recall or use, when a decision depends on which class an item came from.
• Source-correlated features. The classifier has access to features that statistically track source class — perceptual detail, schematic regularity, retrieval fluency, recorded operations, or attached metadata — without being infallible markers of it.
• An attribution mechanism. A classifier maps an item's features to a source class, deciding how the item should be weighted downstream.
• A movable threshold. The decision criterion can be biased toward one class or the other, trading false attributions of one kind against the other.
• Characteristic failure modes. When the features degrade or are forged, internal content is taken as external (intrusion, hallucination, confabulation) or external as internal (cryptomnesia, plagiarism).

These compose into a thin attribution gate operating on a thick store, separable into content errors (wrong fact) and attribution errors (right fact, wrong source) with distinct remedies.

What It Is Not

• Not provenance. provenance is the record of where something came from, ideally captured at origin and carried forward; reality monitoring is the use-time decision about source class made from indirect, fallible features when no clean record exists. Where provenance was tagged at intake, the reality-monitoring gate is unnecessary.
• Not traceability. traceability is the property that one can follow links from an item back to its sources; reality monitoring is the active classification of items whose links are missing or ambiguous, with characteristic errors when the cues mislead.
• Not verification. verification checks whether a claim is true against ground truth; reality monitoring checks where a claim came from. A claim can be correctly sourced and false, or mis-sourced and true — these are orthogonal channels (see signal_detection_theory).
• Not metacognition. metacognition is monitoring one's own knowledge and confidence in general; reality monitoring is the specific source-attribution operation — internally-generated versus externally-perceived — a narrow slice with its own signal-detection structure.
• Not responsibility attribution. responsibility_attribution assigns blame or credit for an outcome to an agent; reality monitoring assigns the epistemic source class of an item. One is normative and causal, the other classificatory.
• Common misclassification. Treating "externally perceived" as a trust certificate, so that correct source attribution licenses confidence. Catch it by asking the orthogonal question after the source is settled: what is this source class's base rate of correctness for this kind of claim?

Broad Use

• Cognitive science (origin) — the source-monitoring framework distinguishing memories of perceived events from imagined ones, and the systematic errors when the discrimination fails.
• Generative-AI safety — distinguishing retrieved factual claims from parametric hallucinations; retrieval-augmented architectures are essentially reality-monitoring systems that tag generations with source metadata.^[3]
• Journalism and evidence handling — tagging each claim with its source (primary witness, secondary report, AI summary, inference) so confidence can be calibrated appropriately.
• Legal and forensic — eyewitness versus suggested-memory discrimination, document provenance chains, and chain of custody, the institutional machinery for preventing source confusion under eroding conditions.^[4]
• Security and access control — distinguishing actions originated inside a trust boundary from those arriving outside, as in insider-threat monitoring and request-provenance tagging.
• Scientific replication — distinguishing observed effects from expected-and-imagined ones, which is precisely what blinding and double-blinding protect.^[5]
• Epistemology and introspection — distinguishing what one has reasoned from what one has been told, and what one has perceived from what one has inferred.

Clarity

Reality monitoring distinguishes content from source attribution. A claim can be true but mis-sourced, or false but correctly sourced, and without the prime, arguments about "where did you hear that?" or "is that real or did you make it up?" become tangled because the source-attribution machinery is invisible. With the prime, source attribution is named as a distinct operation with its own success conditions, its own failure modes, and its own training. A second clarifying move is that the operation runs at use-time, not at acquisition, which is why interventions at acquisition — better recording, cleaner logging — only partially substitute for better attribution machinery: an item that was never tagged at intake must be classified later from indirect features, and those features can mislead. The prime also separates two kinds of error with different fixes. A content error is a wrong fact; an attribution error is a right fact assigned to the wrong source, or a fabrication treated as a memory. The decisive test is whether the system maintains a store from multiple source classes and faces a use-time decision about which source class each item came from, with characteristic errors when the discrimination fails. If the system only ever has one source, or never classifies by source, the prime does not apply.

Manages Complexity

Reality monitoring reduces a vast undifferentiated mass — "things in my head," "things in our database" — to a much smaller question at decision time: for this specific item, what is its source class, and how should that affect how I use it? The attribution decision is a thin gate on top of a thick store, and the store does not need to be re-architected for the gate to add value, which is what makes the pattern cheap to deploy and broadly applicable. The schema also separates two error types that would otherwise be conflated, content error and attribution error, and assigns each its own remediation: content errors are fixed by verification against ground truth, attribution errors by improving the source-correlated features or the threshold. This separation matters because the two failures look alike from outside — a confident false statement could be either — but demand opposite responses. By naming the gate and its two error modes, the prime turns a sprawling reliability problem into a small set of standard moves: tag at source where possible, develop discriminating features where not, and set the threshold according to the relative cost of the two errors.

Abstract Reasoning

Reality monitoring connects to signal-detection theory: source discrimination is a classification problem with a measurable sensitivity and a movable criterion, so biasing the criterion toward "external" reduces false memories at the cost of more rejections of real ones, and conversely.^[6] This makes the precision-recall trade-off explicit and portable — the same trade-off appears in eyewitness lineups, hallucination filters, and document authentication. It connects to provenance graphs in data engineering: a reality-monitoring system is, structurally, a runtime that maintains a provenance graph and queries it at use-time, so the cognitive and the engineering instances are the same object viewed from two disciplines.^[7] And it connects to information theory, where source attribution is meta-information attached to data and its bandwidth cost is the cost of the trust infrastructure. Recognizing these connections licenses a set of portable moves: tag every item with provenance at acquisition because it is cheaper than reconstructing it at use; where tagging at acquisition is impossible, develop features that correlate with source for retrieval-time attribution; expect systematic failures when those features degrade, and beware fluent narratives that read as memory but were synthesized; apply blinding wherever the observer is also the generator; and anticipate that adversaries will forge source-class features, spoofing external content as internal-trusted or the reverse.

Knowledge Transfer

The inheritable structure is explicit: a content store that commingles items from multiple source classes; an attribution mechanism of features plus classifier that assigns source class at use-time; features that correlate with source class such as perceptual detail, retrieval fluency, provenance metadata, or cryptographic signing; a decision threshold trading off the two error types; characteristic failure modes under degraded features, including false memories, confabulation, hallucination, plagiarism, and spoofing; and interventions that improve attribution, including tagging at source, signing, context-recording, and training the discrimination. With these fixed, the moves transfer directly. "Tag every item with provenance at the moment of acquisition" is the same move whether the items are memories, database rows, or model generations. "If you cannot tag at acquisition, develop features that correlate with source for retrieval-time attribution" maps from a clinician learning to tell a remembered patient encounter from one read in a journal to a RAG system inferring whether a claim was retrieved or fabricated. "Blinding and double-blinding are generic interventions wherever the experimenter is also the observer" ports from science to any setting where the generator and the judge coincide. And "adversaries will forge source-class features" is the same warning in security spoofing, forensic document forgery, and cryptomnesia. A team building an enterprise assistant that tags claims "from your docs" with a citation versus "model output, please verify" in a different color, a clinician distinguishing a remembered encounter from a read-about one, and an eyewitness protocol using sequential lineups and neutral administrators to keep the source-attribution gate operating on uncontaminated features are all doing the same structural work: attach provenance to each item, surface it at the moment of use, and set the threshold knowing which of the two errors is more costly.

Examples

Formal/abstract

Treat source attribution as a signal-detection problem, the prime's classifier seen formally. The commingled store holds two source classes — say, internally-generated traces (imagined, inferred) and externally-perceived traces (observed). At use-time the classifier reads a scalar source-correlated feature \(x\), the perceptual-detail score: externally-perceived items draw \(x\) from a distribution centered high, internally-generated ones from a distribution centered lower, but the two overlap.^[1] The attribution mechanism applies a movable threshold \(c\): declare "external" when \(x>c\), "internal" otherwise. The discriminability \(d' = (\mu_{ext}-\mu_{int})/\sigma\) measures how separable the source classes are, and the two characteristic failures fall out of the geometry: a false memory (internal taken as external) is the right tail of the internal distribution crossing \(c\); a missed real memory (external dismissed as imagined) is the left tail of the external distribution.^[2] Sliding \(c\) trades one error against the other along an ROC curve — there is no setting that eliminates both unless \(d'\to\infty\).^[6] This makes the design levers precise. Raising \(d'\) (richer, more diagnostic features) shrinks both errors at once; choosing \(c\) only redistributes them. So the formal model says: invest in better source-correlated features when you can, and set the threshold by the relative cost of the two errors when you cannot.

Mapped back: the overlapping feature distributions are the commingled store viewed through one source-correlated feature, \(d'\) measures how well that feature tracks source class, the threshold \(c\) is the movable decision criterion, and the two ROC tails are the prime's two characteristic failure modes — false attribution each direction.

Applied/industry

A retrieval-augmented generation (RAG) enterprise assistant is, structurally, a reality-monitoring system. Its commingled store is the language model's output stream, where two source classes flow together indistinguishably as plain text: claims grounded in retrieved company documents, and claims confabulated from parametric weights (hallucinations).^[3] The use-time attribution occasion is the moment a user acts on an answer. The source-correlated features are retrieval metadata — whether a sentence's embedding matched a document chunk above a similarity threshold, whether a citation span can be located in the corpus. The attribution mechanism tags each sentence "grounded" or "ungrounded," and the movable threshold is the grounding-similarity cutoff: lower it and more confabulations slip through tagged as grounded (false-external); raise it and genuine grounded claims get flagged "please verify" (false-internal, eroding trust). The intervention the prime prescribes is to attach provenance at acquisition rather than reconstruct it at use — so the architecture pipes citations through generation rather than running a fallible after-the-fact classifier on opaque text. The same structure governs a newsroom: every claim is tagged primary-witness, wire-service, or reporter-inference, and an editor sets the confidence threshold for publishing an unattributed assertion. And it governs a double-blind trial, where blinding keeps the observer from being the generator so that "effect I saw" cannot be silently re-sourced from "effect I expected."^[5]

Mapped back: the LLM output stream, the news draft, and the trial record are commingled stores; grounding-score, attribution-tag, and blinding are source-correlated features and the gate built on them; the similarity cutoff and the editor's publish bar are movable thresholds; and hallucination-passed-as-grounded versus expectation-passed-as-observation are the two characteristic failures — across AI safety, journalism, and experimental science.

Structural Tensions

T1 — Acquisition-Time Tagging versus Use-Time Attribution (temporal). The prime insists the judgment runs at use-time, yet its own best intervention is to tag at acquisition. The tension is that the cheaper, more reliable fix moves the work earlier, dissolving the very use-time classification the prime is about — where provenance is captured at intake, no fallible feature-based gate is needed. The failure mode is investing in ever-better retrieval-time classifiers for a store that could simply have been tagged at the door. Diagnostic: ask whether intake tagging was feasible and skipped — if so, the use-time gate is compensating for an avoidable upstream omission.

T2 — Content Error versus Attribution Error (scopal). A confident false statement could be wrong-fact or right-fact-wrong-source, and the two demand opposite remedies — verification against ground truth versus better source features. The failure mode is treating every reliability problem as one kind: hardening fact-checking while attribution rots, or auditing provenance while the facts themselves go unverified. Diagnostic: hold source fixed and ask whether the content is true; hold content fixed and ask whether the source is right — only running both separates the channels.

T3 — Sensitivity versus Criterion (measurement). Signal-detection geometry forces a split: \(d'\) (how separable the source classes are) and the threshold \(c\) (how the two error types trade off) are independent levers, yet are routinely conflated. The failure mode is sliding the criterion to cut false memories and declaring the system "more accurate," when only \(d'\) improvements shrink both errors at once — criterion moves merely redistribute them. Diagnostic: if an intervention reduced one error while inflating the other, it moved \(c\), not \(d'\), and discriminability is unchanged.

T4 — Honest Feature Degradation versus Adversarial Forgery (sign). The prime's failures assume features degrade benignly, but an adversary forges source-class features deliberately — spoofing external content as internal-trusted, or the reverse. The boundary is with a security/trust-boundary prime where the threat model is active. The failure mode is calibrating a threshold against natural noise and leaving it trivially gamed by anyone who can manufacture perceptual detail or a plausible citation. Diagnostic: ask whether a motivated party benefits from misattribution here — if so, features must be cryptographically bound, not merely correlated.

T5 — Source Class versus Source Reliability (scopal). Reality monitoring attributes an item to a class (internal vs external), but downstream use needs reliability, and the two diverge: a correctly-sourced external claim can still be false, and a correctly-flagged internal inference can be sound. The failure mode is treating "externally perceived" as a trust certificate, so that correct attribution licenses misplaced confidence. Diagnostic: after the source class is settled, ask separately what that class's base rate of correctness is for this kind of claim.

T6 — Thin Gate versus Thick Store Coupling (coupling). The prime sells the gate as cheap because it sits atop an unmodified store — but accurate attribution often requires features the store never recorded, forcing store re-architecture after all. The failure mode is promising lightweight provenance on a system whose items carry no source-correlated signal, then bolting on a classifier that hallucinates attributions from whatever is at hand. Diagnostic: check whether the store actually carries features that track source — if attribution must be invented rather than read, the gate is not thin and the store is the real project.

Structural–Framed Character

Reality Monitoring sits just on the structural side of the middle of the structural–framed spectrum — mixed-structural, aggregate 0.4 — a genuine relational skeleton overlaid with a thin film of cognitive and normative vocabulary. The skeleton is bare: a commingled store, a use-time occasion, source-correlated features, an attribution mechanism, and a movable threshold. Half of the five diagnostics read zero or low, half register modest pull toward framed, and none reaches full framed weight.

Three diagnostics carry half-points. vocab_travels (0.5) reflects that the home lexicon — "source monitoring," "false memory," "internally-generated versus externally-perceived" — is cognitive-scientific, and a reader meeting a RAG grounding filter, a chain-of-custody log, or an insider-threat boundary must translate; but the underlying object is plainly signal-detection geometry (\(d'\) and a criterion), which each domain can and does state in its own terms. evaluative_weight (0.5) is the honest concession that the failure-mode vocabulary carries normative residue: "hallucination," "confabulation," "plagiarism," and "false memory" are not value-neutral the way "intermediate accumulation" is — they name something gone wrong. Yet the core operation, attributing an item to a source class, is itself neutral classification until a cost is attached to each error. import_vs_recognize (0.5) sits in between because invoking the prime partly RECOGNIZES a pattern already present (a store does commingle sources whether or not anyone monitors them) and partly IMPORTS the source-monitoring lens that decides the discrimination is worth making.

The two diagnostics that hold it on the structural side are institutional_origin (0) and human_practice_bound (0.5, not 1.0). Its origin is formal — a classification problem with signal-detection structure — not an institution like contracting or law. And while its paradigm cases are cognitive and its most natural homes are human (memory, journalism, forensics), the pattern does run in non-introspective substrates: a RAG pipeline tagging generations "grounded" versus "ungrounded," a security system distinguishing inside-boundary from outside-boundary requests, an automated provenance graph queried at use-time. Because it can operate in systems that have no "self" at all, it is only half human-practice-bound, not fully so — which is exactly why the aggregate lands at 0.4 rather than tipping framed. The relational core is real; the cognitive frame and normative coloring are genuine but thin.

Substrate Independence

Reality Monitoring is strongly substrate-independent — composite 4 / 5 on the substrate-independence scale. Its domain breadth is broad (4): the post-hoc source-attribution pattern operates with the same structural force in cognition (the source-monitoring framework distinguishing perceived from imagined memories), generative-AI safety (retrieval-augmented systems separating grounded claims from parametric hallucinations), journalism (tagging each claim primary-witness vs. inference), forensics (eyewitness-versus-suggested-memory and chain of custody), security access control (inside-versus-outside the trust boundary), and scientific blinding (observed versus expected-and-imagined effects). The structural abstraction is high (4): the bare skeleton — a commingled store, a use-time occasion, source-correlated features, an attribution mechanism, and a movable threshold — is medium-neutral and reduces formally to signal-detection geometry (\(d'\) and a criterion), which each field can state in its own terms; what keeps it just below the ceiling is a thin film of cognitive-and-normative vocabulary ("false memory," "hallucination," "plagiarism") that travels with light translation and carries some residue of something-gone-wrong. The transfer evidence is concrete (4): the same provenance-graph object appears in cognition and data engineering viewed from two disciplines, and documented moves — "tag at acquisition," "develop source-correlated features for retrieval-time attribution," "apply blinding where the generator is also the judge" — carry intact across memory, RAG pipelines, newsrooms, and double-blind trials. Because the gate can run in non-introspective substrates that have no "self" at all, the prime is recognized across domains rather than confined to human cognition.

• Composite substrate independence — 4 / 5
• Domain breadth — 4 / 5
• Structural abstraction — 4 / 5
• Transfer evidence — 4 / 5

Relationships to Other Abstractions

Current abstraction Reality Monitoring Prime

Parents (1) — more general patterns this builds on

• Reality Monitoring is a kind of, typical Classification Prime

  Reality monitoring is the use-time CLASSIFICATION of each stored item by source class (internally-generated vs externally-perceived) from source-correlated features — a specialized classification with its own signal-detection (d'/criterion) structure.

Children (1) — more specific cases that build on this

• False Fame Effect Domain-specific presupposes Reality Monitoring

  False Fame Effect presupposes Reality Monitoring because the error occurs when use-time source attribution cannot recover prior exposure and assigns familiarity to the wrong origin.

Hierarchy path (1) — routes to 1 parentless root

• Reality Monitoring → Classification

Neighborhood in Abstraction Space

Reality Monitoring sits in a sparse region of abstraction space (61^st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely rather than landing on a neighbor.

Family — Hidden Correlation & Shared Drivers (14 primes)

Nearest neighbors

• Correlated-Source Attribution Failure — 0.72
• Confidence Annotation — 0.71
• Responsibility Attribution — 0.71
• Evidence-Fidelity Decay — 0.70
• Signature-Borne Provenance — 0.70

Computed from structural-signature embeddings · 2026-09-10

Not to Be Confused With

The most insistent confusion is with provenance. The two are intimately related — reality monitoring's best intervention is literally "attach provenance at acquisition" — but they are structurally opposite ends of the same problem. Provenance is a stored record: a chain of metadata, captured at or near origin, that travels with the item and can simply be read off at use-time. Reality monitoring is what you are forced to do when that record does not exist: a use-time inference of source class from indirect features (perceptual detail, retrieval fluency, schematic regularity) that merely correlate with source and can mislead. Provenance, where present, dissolves reality monitoring — there is nothing to classify. The prime earns its keep precisely in the gap provenance leaves: human memory carries no provenance tags, and a raw LLM output stream commingles grounded and confabulated text with no native record. Conflating them leads to the error of assuming a system "has provenance" when all it has is a fallible after-the-fact classifier.

A second genuine confusion is with verification (and the broader signal_detection_theory machinery). Verification answers is this claim true? — it compares content against ground truth. Reality monitoring answers where did this claim come from? — it classifies source. These are orthogonal: a correctly-sourced eyewitness report can be false, and a flagged internal inference can be sound. The danger is collapsing the two channels, hardening fact-checking while source attribution rots, or auditing provenance while the facts themselves go unchecked. A confident false statement is ambiguous between a content error and an attribution error, and only running both tests separately — hold source fixed and ask if content is true; hold content fixed and ask if source is right — disentangles them. Reality monitoring contributes the attribution axis specifically, and borrows signal-detection's \(d'\)/criterion split to characterize its own errors, but it is not itself a truth check.

A third confusion is with metacognition. Reality monitoring is often described as "knowing whether you really saw something or imagined it," which sounds like a metacognitive self-monitoring act. But metacognition is the broad faculty of monitoring one's own cognitive states — confidence, comprehension, knowing-that-you-know — across all content. Reality monitoring is a specific, structured sub-problem: attributing each item to a source class, with a definite classifier, a movable threshold, and two named error modes (false memory, cryptomnesia). It applies equally to non-introspective systems — RAG pipelines, newsrooms, chain-of-custody — where "metacognition" would be a category error. The source-attribution gate has signal-detection geometry that generic metacognitive monitoring lacks.

For a practitioner the discriminations are decisive about where to invest. If you can capture provenance at intake, do so and skip the fallible gate entirely — the cheapest reality monitoring is the kind you never have to perform. If your reliability problem is truth, build verification; if it is mis-sourcing, build attribution features — and never let one substitute for the other. And recognize that the source-attribution gate is a narrow, well-characterized object with its own ROC curve, not a diffuse self-awareness, which is exactly why it can be engineered into systems that have no "self" at all.

Solution Archetypes

Solution archetypes in the catalog that build on this prime — directly (this prime is a source ingredient) or as a related prime.

Built directly on this prime (1)

• Use-Time Source Attribution Calibration: Before using a commingled memory, note, claim, trace, or generated output, classify where it came from and how certain that attribution is.
  ▸ Mechanisms (12)

Also a related prime in 2 archetypes

• Object-Centered Feature Binding: Bind separately detected features to the right object, event, entity, or record by using shared context, co-occurrence cues, exclusivity constraints, and explicit ambiguity states instead of fusing channels blindly.
• Source Distortion Modeling: Treat a report from a systematically distorted source as a biased channel to be modeled, not as either transparent truth or useless noise.

References

[1] Johnson, Marcia K., and Carol L. Raye. "Reality Monitoring". Psychological Review, vol. 88, no. 1 (1981): 67–85. Originates the reality-monitoring framework distinguishing internally-generated from externally-perceived memories by source-correlated features. registry ↩a ↩b

[2] Johnson, Marcia K., Shahin Hashtroudi, and D. Stephen Lindsay. "Source Monitoring". Psychological Bulletin, vol. 114, no. 1 (1993): 3–28. The source-monitoring framework: use-time attribution of memories to source classes and the systematic errors when discriminating features mislead. registry ↩a ↩b

[3] Lewis, Patrick, et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks". Advances in Neural Information Processing Systems 33 (2020). Introduces retrieval-augmented generation, separating retrieved grounded claims from parametric model output. registry ↩a ↩b

[4] Loftus, Elizabeth F. Eyewitness Testimony. Cambridge, MA: Harvard University Press, 1979. Documents how suggestion contaminates source attribution in eyewitness memory, the basis for forensic safeguards against source confusion. registry ↩

[5] Schulz, Kenneth F., Iain Chalmers, Richard J. Hayes, and Douglas G. Altman. "Empirical Evidence of Bias: Dimensions of Methodological Quality Associated with Estimates of Treatment Effects in Controlled Trials". JAMA, vol. 273, no. 5 (1995): 408–412. Shows that inadequate blinding inflates effect estimates, evidencing the need to keep observer separate from generator. registry ↩a ↩b

[6] Green, David M., and John A. Swets. Signal Detection Theory and Psychophysics. New York: Wiley, 1966. Foundational signal-detection theory: discriminability (d') and movable criterion © as orthogonal levers tracing the ROC, framing source attribution as detection. registry ↩a ↩b

[7] Moreau, Luc, and Paolo Missier, eds. PROV-DM: The PROV Data Model. W3C Recommendation, 30 April 2013. Standard model for representing provenance as a graph of entities, activities, and agents that downstream systems record at origin and query at use-time, establishing the cognitive–engineering correspondence for source attribution. registry ↩