Evidence-Fidelity Decay¶
Core Idea¶
Evidence-fidelity decay is the pattern in which an evidentiary record's fidelity to its source event falls with the delay between the event and its capture, and the gap created by memory decay is silently filled by inference from remembered cues and reconstruction from prior expectations. The decisive structural fact is that observation, inference, and reconstruction are written in the same register — same prose, same voice, same surface markers — so a downstream consumer reading the record cannot recover which sentences carry which evidentiary weight. What is lost is not the record's correctness (it may be substantially right on the margin) nor its completeness (it may appear complete) but its attribution structure: which sentences are direct observation, which are inference, which are reconstruction.
Three commitments fix the pattern. First, the substrate is a human-mediated record of an event whose fidelity is delay-sensitive: the longer the gap between event and write-up, the more reconstruction is needed to fill in lost detail. Second, the backfill process is involuntary and undetected by the recorder — the recorder is not consciously fabricating but is unable to mark which sentences came from memory versus reconstruction. Third, the output is registered uniformly: there is no typographic, syntactic, or conventional signal that separates the strata, so attribution information is destroyed at capture time and cannot be recovered downstream. Because the strata are fused at the surface, the consumer mis-weights the record, treating reconstructed material as observed and observed material as reconstructed, and the design variable that decides whether this happens is the attribution-marking convention of the capture format together with the length of the delay.
How would you explain it like I'm…
Memory Filling Gaps
Guesses That Look Real
Lost Attribution Over Time
Structural Signature¶
the source event — the delay interval between event and capture — the delay-sensitive fidelity decay — the involuntary backfill from schema and expectation — the uniform-register fusion of strata — the destroyed attribution structure — the mis-weighting downstream consumer
A record exhibits evidence-fidelity decay when each of the following holds:
- A source event and a mediated capture. Some occurrence is recorded by a recorder whose rendering of it stands between the event and any downstream consumer; the record is not the event but a transformation of it.
- A delay interval. Time elapses between the event and its capture, and fidelity to the source falls as that interval grows. The interval is the primary driver and the primary design lever.
- An involuntary backfill process. As memory of detail decays, the recorder fills the gaps by inference from remembered cues and reconstruction from prior expectation — not consciously, and not stochastically: the backfill is schema-driven, so the resulting distortion is correlated and directional, not zero-mean noise.
- Uniform-register fusion. Observation, inference, and reconstruction are written in the same prose, voice, and surface markers, with no typographic or syntactic signal separating the strata.
- Destroyed attribution structure. What is lost is not correctness or completeness but the recoverable mapping from each sentence to its evidentiary stratum. Attribution is destroyed at capture time and cannot be reconstructed downstream.
- A mis-weighting consumer. A downstream reader, seeing a uniform surface, treats reconstructed material as observed and vice versa — the failure that the lost attribution makes inevitable.
The governing design variables are the delay interval and the attribution-marking convention of the capture format: together they decide how much fidelity decays and whether the strata can be told apart at all.
What It Is Not¶
- Not
provenance. Provenance tracks a record's chain of custody and origin after the record exists, assuming its contents are fixed. This prime concerns the internal evidentiary structure of a record at the moment of capture, before it ever enters a custody chain. - Not
traceability. Traceability is the follow-ability of a record back to its sources; a record can be perfectly traceable yet have its observation/inference/reconstruction strata fused beyond recovery. The defect here is lost attribution within the record, not lost links to its origin. - Not
measurement_uncertainty. Noise is zero-mean and averages out; this prime's backfill is schema-driven, correlated, and directional, so averaging late accounts or trusting inter-rater agreement certifies the bias rather than cancelling it. - Not
biasin general. Bias is a fixed systematic offset; evidence-fidelity decay is a time-dependent loss whose magnitude grows with the event-to-capture delay, so a single correction cannot remove it. - Not
narrative_construction_in_history. Narrative construction is the deliberate, conscious shaping of events into a story; here the backfill is involuntary and undetected by the recorder, who cannot mark which sentences came from memory versus reconstruction. - Common misclassification. Auditing a delayed record for correctness, finding its claims plausible, and certifying it — while never asking which sentences are observed versus reconstructed. The record can stay marginally correct while its attribution structure is wholly destroyed; correctness is the wrong test.
Broad Use¶
The same event–delay–backfill–fused-record skeleton recurs wherever a human recorder captures an event after a delay. In qualitative research, ethnographic field notes written hours or days after observation blend what the ethnographer saw with what they inferred from later conversations and reconstructed for narrative coherence; methodological doctrine prescribes contemporaneous jottings precisely to counter the decay.[1] In law, witness statements taken days after an event blend what the witness saw, what interviewers cued the witness toward, and what the witness reconstructed for coherence — and cross-examination functions partly as a recovery tool for the attribution that statement-taking destroyed.[2] After-action reports written weeks after an operation fuse what was observed, what was learned in subsequent debriefs, and what was reconstructed for narrative arc, which is why mature doctrine runs immediate hot-washes before the formal report.[3] The pattern recurs in accident-reconstruction interviews (months after a crash, direct recall mixes with the investigator's working hypothesis and the witness's exposure to media coverage), in oral history (at decades' remove, the narrative is the product of repeated retellings and exposure to canonical accounts),[4] in laboratory notebooks (a bench entry during a run is dominated by observation; an end-of-day reconstruction blends observation, rationalisation, and reconstruction),[5] in code review written hours after reading a diff, in medical chart documentation written at end of shift, in diplomatic cables and meeting minutes drafted hours after a meeting, and in machine-learning data labelling under time pressure, where labels produced after batch fatigue blend observation of the artefact with reconstruction from the labeller's developing schema.[6] Across all of these, inter-rater or surface-level checks can remain reassuring while the attribution structure is quietly gone.
Clarity¶
Evidence-fidelity decay clarifies by separating three failure modes that an unaided reader of an evidentiary record conflates: the record is incorrect, the record is incomplete, and the record's attribution is lost. The first two are visible, or at least auditable, once one looks — a wrong claim can be checked, a missing section can be noticed. The third is invisible because the surface form of the record is uniform across strata: the reconstructed sentence looks exactly like the observed one. Naming the third failure mode makes it inspectable. A reader can now ask of any sentence, "is this observation, inference, or reconstruction?" and discover that the record cannot tell them — which is itself the diagnostic finding. Designers of capture systems can ask the matched question: does our format preserve attribution, or does it destroy it at write time?
The clarifying force is that it reframes a record as the output of a known transformation rather than as a transparent window onto an event. Once the consumer sees the record as event-then-delay-then-backfill-then-fusion, the right question stops being "is this account true?" and becomes "how long was the delay, and how did the format mark the strata?" This shifts attention from the content of any single sentence to the structural properties of the capture pipeline that produced all of them, and it distinguishes the prime sharply from neighbouring concepts: it is not about chain-of-custody after a record exists, nor about the follow-ability of records to their origins, but about the internal evidentiary structure of a single record before it ever enters a custody chain — a distinction that prevents the decay from being mistaken for a downstream provenance or traceability problem with a different fix.
Manages Complexity¶
A record produced by a delay-sensitive capture process contains multiple strata of evidentiary weight bundled into one textual surface. Without the prime, an analyst must reason about each sentence individually, drawing on context to guess at provenance — a combinatorial inspection across every sentence in the record. With the prime, the analyst can treat the entire record as the output of a single known transformation and apply matched corrections at the level of the whole: downweight the record by an estimated fusion ratio, request a contemporaneous source, or schedule re-capture sooner next time. The per-sentence inspection collapses into a single audit move — how long was the delay, and how did the format mark the strata? — and the answer governs the treatment of the record as a unit.
The compression is more than convenience, because it lets the analyst reason about a class of records without inspecting any individual one. Two records of the same event written at different delays can be ranked by their expected fidelity rather than by their content. A capture pipeline can be evaluated by its delay distribution and its attribution-marking convention independently of any specific record it produces. A reform proposal can be tested by predicting its effect on fidelity — does it shrink the delay distribution, add stratum markers, or require contemporaneous co-capture of a high-fidelity referent? — before any new record is written. The prime supplies the variables (delay interval, backfill ratio, attribution-marking convention) and downstream calculation works in them, so that an intractable sentence-by-sentence reconstruction of provenance becomes a tractable reasoning over a handful of pipeline parameters.
Abstract Reasoning¶
Evidence-fidelity decay trains a reasoner to interrogate any human-mediated record through a fixed set of questions. What was the delay between the event and its capture? What backfill mechanism filled the memory gaps, and was it schema-driven rather than stochastic? Did the format mark observation, inference, and reconstruction separately, or fuse them into a uniform register? Is there a high-fidelity co-record against which the account could be checked? And how will a downstream consumer weight the record, given that the attribution structure is silently absent? Because these questions reference only the abstract roles — event, capture instrument, delay interval, backfill process, fused output, downstream consumer, attribution-marking convention — they apply to a field note, a deposition, a cable, or a label without translation.
The abstraction also licenses comparative and predictive reasoning that the content of any record cannot support. Because the contamination is prior-structured rather than stochastic — inference and reconstruction follow the recorder's schema, so the error is correlated and directional, not zero-mean noise — the reasoner can predict the direction in which a record will distort, not merely its magnitude. Two pipelines with the same nominal delay but different attribution conventions will produce records of different recoverable fidelity, and the reasoner can rank them in advance. A proposed change to a capture system can be evaluated by its predicted effect on the delay distribution and the stratum-marking convention before it is deployed. This is reasoning at the level of the measurement system rather than the measurement, and it is what distinguishes the prime from generic notions of noise or error: the failure is a determinate transformation with named variables, so it can be modelled, compared, and forecast.
Knowledge Transfer¶
The strongest evidence that the prime travels is that, across every substrate, exactly four matched interventions recur — and they are recognisable instances of one another even though the craft conventions are unrelated. The first is to shrink the delay through contemporaneous capture: bench notes during a run, hot-washes immediately after an operation, real-time labelling, immediate witness statements, in-meeting minutes. The second is to mark the strata typographically or syntactically: separate observation, inference, and reconstruction columns in field notes, explicit "I observed / I inferred" markers in after-action reports, side-margin inference annotations in oral-history transcripts.[1] The third is to preserve a high-fidelity co-record: tape the interview, record the bench instrument output, photograph the scene, keep the email thread alongside the meeting cable.[7] The fourth is to audit fidelity by re-capture: ask the witness again later, re-interview the survivor, re-listen to the meeting tape against the cable, spot-check labels against the source artefact.[8]
The transfer is deep because the four moves are not analogies but instances of one structural toolkit applied to one structural problem. A contemporaneous code-review comment and a contemporaneous police statement are doing the same structural work — collapsing the delay interval to zero — even though one is software practice and the other is criminal procedure. A field investigator interviewing a control-room operator six weeks after a near-miss, who transcribes a coherent account that silently fuses alarms the operator heard, alarms inferred from the circulated incident report, and alarms reconstructed for narrative sense, faces exactly the problem an oral historian faces with a decades-old recollection and an ML team faces with end-of-session labels.[9] The matched intervention is identical in form across all three: schedule first-pass capture sooner, structure the record with explicit heard/inferred/reconstructed strata, preserve a co-record (the alarm log, the original recording, the source artefact), and require a second-pass reconciliation against that co-record with the differences kept in the file rather than erased. Because the roles and the interventions are the same, a practitioner who has internalised the pattern in one field can design the capture discipline for another on first contact.
Examples¶
Formal/abstract¶
Model a capture pipeline by two governing parameters and the prime becomes a comparative instrument rather than a single observation. Let the delay interval \(d\) be the elapsed time between event and write-up, and let the attribution convention \(a\) be a binary property of the format: either strata are marked (observation, inference, and reconstruction kept typographically distinct) or fused (one uniform register). Treat the record as the composition event \(\to\) decay(\(d\)) \(\to\) backfill \(\to\) register(\(a\)). Two structural facts fall directly out of this composition. First, fidelity to the source falls monotonically with \(d\), but — and this is the load-bearing point — the correctness of the resulting record can stay high while its recoverable attribution drops to zero, because backfill is schema-driven and therefore produces plausible, coherent, directionally-biased sentences rather than visibly wrong ones. The damage is to the mapping from sentence to evidentiary stratum, not to the marginal truth of the sentences. Second, because the backfill is correlated rather than zero-mean, the model predicts the direction of distortion (toward the recorder's prior schema), so two pipelines with identical \(d\) but different \(a\) can be ranked in advance: the marked pipeline preserves the stratum mapping the fused one destroys at write time. This lets an analyst reason over the class of records a pipeline produces — rank two pipelines, forecast the effect of a reform on \(d\) or \(a\) — without inspecting any individual record sentence by sentence.
Mapped back: \(d\) and \(a\) are the delay interval and attribution-marking convention named in the signature; the composition event-decay-backfill-fusion is the prime's transformation, and the prediction that correctness can survive while attribution is destroyed is exactly the destroyed-attribution invariant made parametric.
Applied/industry¶
Three unrelated professional records run the identical transformation. A police witness statement taken four days after a robbery fuses what the witness saw, what the interviewer's questions cued them toward, and what the witness reconstructed for narrative coherence — all in one continuous prose paragraph with no marker separating the strata, which is precisely why cross-examination exists as a downstream recovery tool for the attribution that statement-taking destroyed.[2] A field investigator interviewing a control-room operator six weeks after a near-miss transcribes a coherent account that silently merges alarms the operator actually heard, alarms inferred from the circulated incident report, and alarms reconstructed for sense; an inter-rater check between two investigators can look reassuring while the attribution structure is wholly gone. A machine-learning team labelling images under batch fatigue produces end-of-session labels that blend direct observation of the artefact with reconstruction from the labeller's hardening schema, so a label set can pass surface QA while quietly encoding the labellers' expectations rather than the images.[6] In all three the matched four-part intervention is the same in form: shrink the delay (immediate statement, on-shift incident capture, real-time labelling), mark the strata (heard/inferred/reconstructed columns, "I observed / I inferred" flags), preserve a high-fidelity co-record (interview tape, alarm log, source artefact), and reconcile a second pass against that co-record with the differences kept in the file. A detective, a safety investigator, and an ML lead are running one capture discipline under three craft vocabularies.
Mapped back: The robbery, the near-miss, and the labelling session are each a source event; the four-day, six-week, and end-of-session delays are the interval; the uniform prose, the coherent transcript, and the clean label set are the register-fusion that destroys attribution — the same prime in criminal procedure, safety engineering, and data labelling.
Structural Tensions¶
T1 — Correctness versus Attribution (measurement). The prime's signature claim is that a record can stay marginally correct while its attribution structure is destroyed; what decays is not the truth of the sentences but the recoverable mapping from sentence to evidentiary stratum. The characteristic failure is auditing for the wrong thing — checking the record's claims for accuracy, finding them plausible, and certifying it, while the question of which sentences are observed versus reconstructed was never asked. The diagnostic is to test attribution directly: pick any sentence and ask whether the record can say if it is observation, inference, or reconstruction. If it cannot, the failure is present regardless of correctness.
T2 — Delay versus Fidelity (temporal). Fidelity falls monotonically with the event-to-capture delay, which makes delay the primary lever — but it is a lever in tension with every reason capture is deferred (the operation is still running, the shift is not over, the labeller is mid-batch). The failure is treating a delayed record as equivalent to a contemporaneous one because both read as coherent prose; coherence rises with delay as reconstruction smooths the gaps, so fluency is anti-correlated with fidelity. The diagnostic is to record and weight by the delay interval itself, treating a smooth late account as more suspect, not less, than a fragmentary immediate one.
T3 — Directional Backfill versus Zero-Mean Noise (sign/direction). The contamination is schema-driven, so the distortion is correlated and directional — toward the recorder's prior expectation — not zero-mean noise that averages out. The failure is applying noise-cancelling remedies to a biased signal: averaging multiple late accounts, or trusting inter-rater agreement, when all recorders share the schema and so converge on the same reconstruction. Agreement then certifies the bias rather than the truth. The diagnostic is to ask whether the recorders share a prior; where they do, concordance is evidence of common reconstruction, and only an independent high-fidelity co-record breaks the tie.
T4 — Single Record versus Custody Chain (scopal). The prime concerns the internal evidentiary structure of a record before it enters any custody chain, which is precisely where provenance and traceability take over — those track a record's movement after it exists and assume its contents are fixed. The failure is reaching for a provenance fix (chain-of-custody, tamper-evidence, source-linking) for an attribution problem, hardening the pipeline around a record whose strata were already fused at write time. The diagnostic is to locate the defect temporally: was attribution lost at capture (this prime) or could the record be traced but not trusted as it travelled (provenance)?
T5 — Marking Convention versus Uniform Register (coupling). Whether strata can be told apart downstream is fixed at capture by the format's attribution-marking convention, coupling a downstream consumer's epistemic options to an upstream design choice they cannot revisit. The failure is a consumer trying to recover attribution from a fused record by close reading, attempting to reconstruct at consumption time information that was destroyed at write time and is not latent in the text. The diagnostic is to inspect the capture format, not the record: if the convention fuses observation, inference, and reconstruction into one register, no downstream effort recovers the mapping, and only re-capture or a co-record helps.
T6 — Per-Record Treatment versus Pipeline Reform (scalar / local-global). The prime supports two levels of action: correct an individual record (downweight it, request a co-record) or reform the pipeline that produces a whole class (shrink the delay distribution, mandate stratum markers). Optimising locally — patching each suspect record as it surfaces — leaves the generating process intact and guarantees the next record carries the same defect. The failure is endless per-record firefighting mistaken for quality control. The diagnostic is to ask whether the intervention changes the pipeline parameters (delay, marking convention) or only the record in hand; durable fidelity gains live at the pipeline level.
Structural–Framed Character¶
Evidence-fidelity decay sits on the framed side of the structural–framed spectrum, consistent with its framed label and aggregate of 0.6. The pattern has a real transformational skeleton — event, delay, schema-driven backfill, register-fusion, destroyed attribution, mis-weighting consumer — but it cannot be stated without naming a human recorder and an evidentiary register, and that human-practice dependence is what places it past the middle.
The decisive criterion is human_practice_bound (1.0): every role presupposes a person and an institution of record-keeping. The "recorder" who involuntarily backfills, the "register" in which observation and reconstruction fuse, the "evidentiary stratum" whose attribution is lost, and the "downstream consumer" who mis-weights are all human-mediated documentary roles. There is no physical or biological substrate — a sensor log does not reconstruct from schema; only a remembering human recorder does. The remaining four criteria each read mid (0.5) rather than fully framed, which is why the aggregate lands at 0.6. The vocabulary partly travels — "delay," "backfill," "attribution-marking convention" are statable in field notes, depositions, after-action reports, and ML labels each in its own idiom — but it carries an evidentiary, documentary lexicon along with it (vocab_travels 0.5). It carries a mild negative valence: "decay" and "destroyed attribution" frame the pattern as a defect to be guarded against, though the structural diagnostic itself stays value-neutral on the record's correctness (evaluative_weight 0.5). Its origin is epistemological and documentary rather than formal, but not the creature of any single named institution — courts, labs, and ethnography all host it (institutional_origin 0.5). And invoking it imports a modest interpretive frame — reading a coherent statement as a fused composition of strata rather than as a transparent window — while still recognising a genuine transformation that is really present in the record (import_vs_recognize 0.5). The transformational skeleton is real and recurs cleanly across recorder-mediated substrates, but because it exists only where a human recorder captures into an evidentiary register, the prime belongs on the framed side.
Substrate Independence¶
Evidence-fidelity decay is a moderately substrate-independent prime — composite 3 / 5 on the substrate-independence scale. Its domain breadth is real but confined: the event–delay–backfill–fused-record skeleton recurs across qualitative research (ethnographic field notes), law (witness statements and cross-examination), military after-action reports, accident-reconstruction interviews, oral history, laboratory notebooks, code review, medical charting, diplomatic cables, and machine-learning data labelling — many genuinely distinct fields, but every one of them a setting where a human recorder captures an event after a delay, with no reach into purely physical or non-recorder-mediated substrates (a sensor log does not reconstruct from schema). Its structural abstraction is mid because the roles — the recorder who involuntarily backfills, the register where observation and reconstruction fuse, the evidentiary stratum whose attribution is lost — are portable across those fields but cannot be lifted out of human, memory-mediated record-keeping; the signature presupposes a remembering documentary agent. Its transfer evidence, though, is strong enough to earn a 4: the same defect is documented with the same recovery doctrine across substrates — contemporaneous jottings in ethnography, immediate hot-washes before formal AARs, cross-examination as an attribution-recovery tool in law, bench-time entries versus end-of-day reconstruction in lab notebooks — and the diagnostic carries cleanly between them, with the same reassuring-surface-checks-while-attribution-is-gone pathology recurring intact. The composite sits at 3 because the breadth and abstraction are pinned by the human-recorder ceiling even as transfer within that band is well evidenced.
- Composite substrate independence — 3 / 5
- Domain breadth — 3 / 5
- Structural abstraction — 3 / 5
- Transfer evidence — 4 / 5
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
-
Evidence-Fidelity Decay is a kind of, typical Temporal Decay and Degradation
The file: 'a TIME-DEPENDENT loss whose magnitude grows with the event-to-capture delay' — a specialization of temporal decay applied to an evidentiary record's attribution structure (vs a static bias).
Path to root: Evidence-Fidelity Decay → Temporal Decay and Degradation → Entropy (Thermodynamic Sense)
Neighborhood in Abstraction Space¶
Evidence-Fidelity Decay sits among the more crowded primes in the catalog (24th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.
Family — Memory, Records & Persistence (27 primes)
Nearest neighbors
- Evidence-Latency Window — 0.73
- Silence as Signal — 0.73
- Replay — 0.73
- Record-Reality Divergence — 0.72
- Memory Consolidation — 0.72
Computed from structural-signature embeddings · 2026-06-14
Not to Be Confused With¶
The nearest and most dangerous confusion is with provenance, because both are about how much one can trust an evidentiary record. Provenance is concerned with a record's custody and origin after it exists: who held it, when, through what hands, with what tamper-evidence — a chain that lets a consumer verify the record was not altered or substituted in transit. It treats the record's contents as a fixed object whose integrity is the question. Evidence-fidelity decay operates one step earlier, at capture: it concerns whether the contents themselves faithfully separate observation from inference from reconstruction, a property fixed by the event-to-capture delay and the format's attribution-marking convention before any custody chain begins. The two have different failure loci and different fixes. A record can have impeccable provenance — securely held, demonstrably unaltered — and still be epistemically corrupt because its strata were fused at write time; conversely a record with honest internal attribution can be compromised by a broken custody chain. The discriminating question is temporal: was the defect introduced at capture (this prime, fixed by shrinking delay and marking strata) or during travel (provenance, fixed by chain-of-custody and tamper-evidence)? Reaching for a provenance fix — hardening the pipeline around a record whose attribution was already destroyed — is the canonical misapplication.
A second confusion is with measurement_uncertainty, since both describe departures of a record from the truth. Noise is zero-mean and stochastic: it scatters around the true value and is removed by averaging, by more samples, or by independent re-measurement whose errors cancel. The backfill in evidence-fidelity decay is schema-driven, correlated, and directional — recorders fill memory gaps from their shared prior expectation, so the distortion points the same way across recorders and across re-tellings. This difference is not academic: the noise-cancelling remedies that work against observational noise backfire here. Averaging multiple late accounts, or trusting high inter-rater agreement, certifies the common reconstruction rather than the truth, because all the recorders share the schema and converge on the same biased fill. The discriminating question is whether independent recorders' errors are independent: where they share a prior, concordance is evidence of common reconstruction, not of fidelity, and only an independent high-fidelity co-record breaks the tie.
A third confusion, especially in historiographic and testimonial settings, is with narrative_construction_in_history. Narrative construction is the deliberate and conscious assembly of events into an emplotted account — the historian or witness knowingly selecting, ordering, and connecting. Evidence-fidelity decay is involuntary and undetected: the recorder is not consciously fabricating and cannot tell which sentences came from memory versus reconstruction, because the format gives no marker to separate them. The practical consequence is that narrative construction is, in principle, something the author can flag and the reader can discount, whereas the fusion this prime names is invisible to author and reader alike and is recoverable only by re-capture or a co-record. Treating the decay as deliberate narrative shaping mis-locates both the cause (a memory-and-format process, not an authorial choice) and the cure (shrink the delay and mark the strata, not "be more objective").
These distinctions matter because each points to a different intervention: provenance hardening for custody defects, averaging or replication for genuine noise, authorial transparency for deliberate narrative — none of which addresses the involuntary, directional, capture-time attribution loss that is this prime's signature, fixed only by contemporaneous capture, stratum marking, and a preserved high-fidelity co-record.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.
References¶
[1] Emerson, Robert M., Rachel I. Fretz, and Linda L. Shaw. Writing Ethnographic Fieldnotes. University of Chicago Press, 1995. Methodological doctrine prescribing contemporaneous jottings and separating observation from inference in field notes to counter delay-driven reconstruction. ↩
[2] Loftus, Elizabeth F. Eyewitness Testimony. Harvard University Press, 1979. Demonstrates that delayed witness recollection is reconstructive and contaminated by post-event suggestion, schema, and interviewer cueing — directional, not zero-mean, distortion. ↩
[3] U.S. Army. A Leader's Guide to After-Action Reviews (TC 25-20). Headquarters, Department of the Army, 1993. Distinguishes the immediate 'hot wash' that captures lessons at once from the more formal, later after-action review, motivating contemporaneous capture before the formal delayed report. ↩
[4] Thompson, Paul. The Voice of the Past: Oral History. Oxford University Press, 1978. Treats decades-removed recollection as the product of repeated retelling and exposure to canonical accounts, addressing the reconstructive fidelity of oral testimony. ↩
[5] U.S. Food and Drug Administration. Data Integrity and Compliance With Drug CGMP: Questions and Answers — Guidance for Industry. FDA, 2018. Codifies the ALCOA principle that records be 'contemporaneous' — recorded at the time of the activity — because entries written later from memory are a principal data-integrity failure, the lab-notebook analogue of delayed reconstruction. ↩
[6] Sambasivan, Nithya, Shivani Kapania, Hannah Highfill, Diana Akrong, Praveen Paritosh, and Lora M. Aroyo. "'Everyone wants to do the model work, not the data work': Data Cascades in High-Stakes AI". Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. Documents how under-valued, time-pressured data/labeling work produces compounding downstream 'data cascades,' so label sets pass surface checks while encoding annotators' developing schemas rather than the artifact. ↩
[7] College of Policing. "Investigative Interviewing (Authorised Professional Practice)". College of Policing, UK. Recommends audio/video recording of witness and suspect interviews as a high-fidelity co-record that preserves the original account against later reconstruction and supports attribution recovery. ↩
[8] Fisher, Ronald P., and R. Edward Geiselman. Memory-Enhancing Techniques for Investigative Interviewing: The Cognitive Interview. Charles C. Thomas, 1992. Structured re-interview protocol (context reinstatement, report-everything, varied retrieval order) that re-elicits eyewitness memory — a re-capture/fidelity-audit method for delayed recollection. ↩
[9] Dekker, Sidney. The Field Guide to Understanding 'Human Error', 3rd ed. Ashgate/Routledge, 2014. Shows that accident reconstruction works backward from the outcome, fusing what was actually perceived with hindsight-driven inference — directional reconstruction in delayed safety investigation. ↩