--- title: "A Tape That Doesn't Get Erased" subtitle: "Consensus is not a sensor—and neither is a model" author: "Dietrick Hardwick, Founder & Chief Technologist, Interveil Labs" date: "2026-08-18" status: "Public journal draft" --- # A Tape That Doesn't Get Erased Paris Observatory, 1961. A team tracking artificial satellites logs an object moving retrograde—against the direction of everything then being launched. By Jacques Vallée's later account, a supervisor wipes the tape. Not because the object had been identified. Because filing the report would expose the observatory to ridicule. Notice what was destroyed. Not proof of anything. A *measurement*. The tape's value was never in what it settled. Its value was that someone, later, under a better theory than anyone possessed in 1961, might have been able to open it again. Now notice the status of the story itself. It reaches us as one person's recollection, decades on, unwitnessed and presently unverifiable. It is an anecdote about the destruction of evidence, and it is not itself evidence of the event it describes. Both statements can be held at the same time. A serious record must hold them separately without letting either impersonate the other. That capacity—carrying a claim and its support structure as different objects—is what Interveil is for. ## Consensus is not a sensor Mystery sets the mind in motion. A light appears. A photograph survives. A witness searches for a noun. One account becomes ten articles. Ten articles become a consensus. The consensus becomes a mythology. Soon, an idea repeated by thousands can feel heavier than the solitary artifact from which it descended. But consensus cannot see, hear, weigh, timestamp, or measure the original event. It measures people—or, more often, an undefined impression of what people appear to believe. Without a cohort, method, denominator, independence analysis, and date, “consensus” is not a finding. It is atmosphere. The failure has a loop: 1. An interpretation is published. 2. The interpretation is indexed. 3. The indexed interpretation is cited as corroboration. 4. Corroboration is counted as weight. 5. Weight is read as evidence. At no step did anything enter from the world. In human systems this loop can take decades, slow enough to resemble scholarship. In machine systems it closes at speed: today's inference becomes tomorrow's corpus becomes next year's apparent ground truth. A model is asked what is true, answers from recurrence, and its answer joins the recurrence it mistook for evidence. This is the echo problem. A thousand articles derived from one anonymous post do not constitute a thousand sources. They constitute one unresolved root and 999 echoes. Any system that counts them as a thousand has laundered repetition into corroboration. Language models did not invent the failure. They made it instantaneous. ## Three truths for the record Interveil begins from three governing truths. Each names a collapse; each imposes an obligation on anything built to know better. ### Reality exceeds representation Every descriptor is a projection, and every projection is lossy. A witness may report with total accuracy at the level of luminance, motion, and duration—and catastrophically at the level of the noun. *Craft* imports a builder. *Occupant* imports intent. *Landed* imports a destination. The noun is a compressed theory applied at the moment of speech, and from that moment it is typographically indistinguishable from what was seen. This is not a mystical exemption from rigor. It is the reason rigor is necessary. If every representation is partial, we must preserve what was registered, how it was captured, what disappeared in translation, and which conclusions exceed the instrument. “Unknown” is not an error to be auto-completed. ### Lineage is not veracity No datum arrives immaculate. A photograph has a camera, crop, compression history, and custody chain. A witness report has an interviewer, question sequence, vocabulary, and interval between event and account. A valid signature proves that someone signed an assertion. A hash proves that bytes have not changed. Neither proves that the originating claim was correct. A perfectly authenticated lie remains a lie. An anonymously preserved truth remains poorly sourced. The system must represent both conditions honestly. ### Meaning must remain revisable The same material may support a forensic reading, historical reading, legal reading, or cultural reading. Trouble begins when a reading hides its frame and presents itself as the source. The answer is not to pretend the interpreter has vanished. It is to preserve interpretation as an explicit, attributable, versioned layer. A new reading never overwrites the old one. It supersedes it. The predecessor remains legible beside the reason it was displaced. This is not relativism. Some claims are far better supported than others. The discipline simply refuses to achieve certainty by deleting the history of uncertainty. Together the three truths yield a practical instruction: > **Preserve the thing. Expose the chain. Version the meaning.** ## The Engine The Interveil Engine is a typed, append-only epistemic graph built for those obligations. Its basic unit is a **Truth Object**—a deliberately dangerous name carrying its own warning. A Truth Object does not contain truth. It contains a bounded question and the inspectable approaches to it: referents, artifacts, observations, atomic claims, evidence bearings, hypotheses, interpretations, dependency clusters, structured loss, provenance frontiers, assessments, and revisions. The Engine keeps two planes. The **Record Plane** holds what was captured, how it changed, which claims were extracted, what bears on them, where sources share a root, and where the chain stops. It is append-only. Correction arrives as a new state, never as amnesia. The **Lens Plane** renders the record for a person, purpose, and moment. A forensic lens and a plain-language lens may speak differently. The graph under them does not move. Between the planes is a one-way membrane: observation may inform interpretation; interpretation may never be re-ingested *as* observation. That is a type rule, not an editorial aspiration. The Engine refuses eight familiar transformations. Interpretation cannot become observation. A signature cannot become truth. Repetition cannot become corroboration. Missing provenance cannot become deception. Revision cannot erase its predecessor. Consensus cannot support the event whose belief-state it measures. Partial support cannot become full support. Absent context cannot masquerade as neutrality. And there is no truth score. Every important claim instead carries seven separate assessments: source integrity, chain completeness, claim-evidence fit, source independence, alternative coverage, context sensitivity, and unresolved loss. Each has a rationale, assessor, date, and trigger for reconsideration. The vector is harder to read than a number. That is precisely why it is more honest. The convenience of a single score comes from throwing away the information that would let someone disagree with it. ## The relationship layer A trustworthy record can still be wrapped in an extractive product. That is why the Engine is not the entire stack. **Authentic Intelligence** is the enrichment layer: the promises governing how the system meets a person. A person's material stays home by default. Capture is explicit and attributed. The interface discloses what it selected, what it omitted, which model participated, which lens governed, and whether a human reviewed the result. The person remains the author. The model proposes. A person accepts, edits, or rejects. **No autonomous canonical writes.** Above that sit **Instruments**: Inspector, Reader, Studio, forensic view, Evidence API. An instrument is a surface over a packet. None is canonical. The same record can be picked up differently without being rewritten beneath the hand. The distinctions are not branding. A system can have excellent asset provenance and still exploit its user. A system can honor consent and still launder a thousand echoes into a thousand sources. Those are different failures. They need different layers. ## What the Alpha actually proves Interveil Alpha is deliberately small. It does not identify a mysterious object, certify a truth, model all knowledge, or automate judgment away. Two independent teams applied the same doctrine to two different contested records. The first encoded a mid-century crash narrative as 114 nodes: 40 claim-family assertions, three independence clusters, five loss events, two provenance frontiers, and seven assessment dimensions. Its twelve-question evaluation returned **9 pass, 3 partial, 0 fail**. The second encoded a contested 1990 photograph as 87 nodes and 66 links: twelve atomic claims, nine observations, six evidence records, three hypotheses, five loss events, three frontiers, seven assessments, three lenses, two summary revisions, and fifteen machine-checked invariants. The figures are not blended. The agreement is the finding. Both teams independently needed a Record Plane and Lens Plane; a one-way membrane; evidence as an evaluated bearing rather than a synonym for artifact; scoped independence; explicit loss; provenance frontiers; seven assessment dimensions; supersede-not-erase history; consensus as a derived population record; and no truth score. They disagreed about whether one or two concepts deserved their own node types. That is useful data, not embarrassment. The unification rule now is severe: a concept earns a core node type only if two independent packets both needed it as a node. The ontology grows from pressure, not imagination. What neither packet proves is equally important. We do not yet know whether the vocabulary generalizes to scientific replication, procurement, litigation, or medicine. We do not yet know whether ordinary users answer better with the Engine than with a conventional summary. We do not know how to determine source independence automatically. We do not know whether seven-dimensional assessment survives contact with a reader who wants one green light. Those are the next tests. Not a larger claim. ## Search ancestry, not just strings Search finds words. Retrieval finds neighboring passages. A conventional knowledge graph connects entities. Interveil asks a different class of question: what did this sentence *inherit*? Return every assertion whose apparent support terminates in one unwitnessed testimony. Show the claims that gained citation weight without gaining sensory support. Identify every point where an interpretation entered the record as an observation. Trace this noun to the first document that used it and show what it displaced. List the conclusions that collapse if one named source is removed. The last query is the one that matters commercially. It is the structural signature of credibility laundering. It is invisible to keyword search because the repeated sentence can be perfectly accurate in every copy. The failure lives in the dependency structure beneath it. We call this **provenance grep**. ## A bounded promise Interveil is not a truth oracle. It is not a debunking engine. It is not a blockchain in search of a problem. It is not a universal ontology. It does not turn missing provenance into suspicion, credentials into evidence, or consensus into a sensor. It offers something narrower and more useful: an answer that can show exactly how far it reaches, and a record that can change its mind without lying about what it used to believe. Mystery is not the enemy of knowledge. Counterfeit certainty is. The erased tape remains an anecdote. The measurement it describes, if it existed, is gone. Nothing we build can restore it. But we can build the next record differently. We can let a frontier remain a frontier. Let an echo remain an echo. Let a noun disclose the theory smuggled inside it. Let a new interpretation arrive without murdering the old one. Not an answer to every mystery. A tape that does not get erased. --- *Interveil Engine v1.0 was developed under the direction of Dietrick Hardwick, Founder & Chief Technologist, with disclosed AI research, design, implementation, and drafting collaboration. External claims belong to their cited records. Interveil synthesis is proposal. The record speaks.*