High-Fidelity Maps: A Specification for Documents That Track Their Territory
High-Fidelity Maps
A specification for documents that try to track their territory: the commitments and instruments this corpus runs on, derived from a model of fidelity and applied to themselves
Meta-specification · the guidelines every other document type in this corpus is written to
Version 1 · June 2026 · written to its own rules
A document is a map of some territory, and its fidelity is how well it makes a reader's resulting model match that territory. High-fidelity territory mapping is the meta-goal the whole corpus serves; an individual document is one instance of it, and what kind of instance it is is named in a single line, its type (§2). This piece specifies the guidelines that keep fidelity high. The central claim, and the correction over earlier drafts, is that fidelity is not traded against usability and is not raised by adding apparatus. Fidelity is the single target, always correspondence between the model the map affords its target reader and the purpose-scoped territory. What changes with purpose is the fidelity profile, declared as a tolerance per dimension, which of fidelity's dimensions matter, at what level, for which reader, in which register, and therefore which instruments serve it. A set of commitments is invariant across every profile (§3), and is shown discharged in plain prose, not only in apparatus (§3.1); the apparatus itself, claim and confidence labels, citations, scope boxes, glossaries, visuals, is instrumental and applied only as the profile calls for it (§6). This is why a stripped, plain-register piece and a fully formalized research module can both be high-fidelity: they honor the same commitments with different instruments. Each guideline is tagged invariant or instrument, the specification is held to its own rules (§7), and it is deliberately scoped: §9 states what this version does not yet do.
Claim labels used for this document's own guidelines (declared per I1): definitional stipulated, settled by choice of meaning; invariant a commitment that holds for any high-fidelity map, whatever its profile; instrument apparatus applied only as a document's fidelity profile calls for it; open an honest gap or limit.
1. The model: fidelity, its dimensions, signal and noise
The guidelines below are derived from a small model of what a document is for. Stating the model first is what lets a reader see which guidelines are invariant (they hold for any high-fidelity map) and which are instruments (a given document may drop them without losing fidelity).
1.1 The objects
Four objects, defined by stipulation. definitional The territory is the part of the world a document is about. The map is the document. The target reader is the audience the purpose names. The target is the model of the territory a competent member of that audience would form on reading the map. Fidelity is the correspondence between the target and the purpose-scoped territory, that is, the correspondence the map affords its target reader.1 Whether that afforded model actually forms in real readers is fidelity's empirical test, currently proxied by register and navigation and not yet measured (§8.5). Indexing fidelity to the afforded model of the intended reader is load-bearing in two directions: a map a competent member of its audience cannot extract a correct model from is low-fidelity however rigorous it looks on the page, so reaching that reader is part of fidelity and not a rival to it; and because the target is the afforded model and not any onlooker's reaction, a document is not made low-fidelity by an audience it was never written for.
1.2 Fidelity has four invariant dimensions
Treating fidelity as a single quantity is the first error to remove, because a guideline that raises one dimension can lower another. invariant At minimum fidelity has four components, each definitional. definitional (Whether these four exhaust it is left open, §8.6.)
- Accuracy. Do the map's claims match the territory.
- Resolution. How fine-grained the map is. A true but coarse map is low-fidelity for a purpose that needs detail.
- Coverage. How much of the scoped territory is represented. Named "coverage" rather than "scope" on purpose, so "scope" is left to mean the document's declared boundary and is not asked to carry two referents (an instance of C7).
- Calibration. Whether expressed confidence tracks the actual reliability of each claim.8 An overconfident accurate map is lower fidelity than a calibrated one, because it misrepresents where it can be trusted.
1.3 The profile is what flexes
Under finite effort and finite space these dimensions trade off, and which to favor depends on purpose. invariant A document occupies a point on a frontier internal to fidelity, not a frontier of fidelity against some other goal. A survey buys coverage by spending resolution; a proof buys accuracy and calibration on a narrow coverage. There is no universal best map, for the same structural reason that all models are lossy and chosen relative to a use.23 The term "fidelity" is borrowed in its information-theoretic sense, the criterion by which a reconstruction is judged against a source.4 §2 makes the point on that frontier precise as a declared tolerance per dimension.
1.4 The apparatus is instrumental, not constitutive
An earlier version of this specification treated register as a trade of fidelity against usability, and implied that more apparatus meant more fidelity. Both were errors, and both contradicted §1.1. invariant The correction: the apparatus (labels, citations, scope boxes, glossaries, formal hedges, constructed visuals) are instruments that raise specific dimensions for specific profiles, and the same instrument can lower fidelity under a different profile. "The Manual for When You Don't Know What to Believe" drops nearly the entire apparatus and is high-fidelity, because for a frightened reader the apparatus would block the afforded model from forming; it keeps the invariant commitments and delivers them in prose (§3.1 shows how). A research module uses the full apparatus and is high-fidelity for its profile. Instrument count is not fidelity. What is invariant is the set in §3; what flexes is the instrument set in §6, selected by the tolerances in §2.
1.5 Signal and noise
definitional Signal is what raises fidelity to the purpose-scoped territory; noise is what lowers it. invariant Noise is not only padding words. A false claim, a grammar or structural error, an equivocation, and an instrument used where the profile does not call for it are all noise, because each degrades the correspondence the map affords its target reader. This is the precise reason removing apparatus can raise fidelity: apparatus a profile does not need is noise. At the document level, mis-scoping the territory or mis-mapping the scoped part is noise too, which is the failure the scope-box guardrail names. C6 removes noise at the word level; the same test, does this raise correspondence to the purpose-scoped territory or lower it, applies at every level. (A systematic catalog of noise, the failure modes of a map, is named and deferred in §9.)
2. Document types, profiles, and tolerances
2.1 Type, named in a sub-sub-title
A document's type is the one-line handle for its fidelity profile, named in a sub-sub-title directly under the subtitle. instrument The type tells the reader what the document is and therefore what is legitimately present and what is legitimately absent, so that a stripped document is read as a deliberate profile rather than an unfinished one.
The corpus keeps no closed catalog of types, because a catalog can never be complete and an incomplete one presented as complete is itself noise (§1.5). A type is instead a region in a space with three axes: temporal stance (past-looking, such as a post-mortem, a retraction, or a genealogy; or future-looking, such as a specification, an RFC, or a prolegomena); social stance (solo, adversarial such as a polemic or rebuttal, or collaborative such as an RFC or a call for counterexamples); and ontological stance (mapping the world, such as an empirical paper; mapping maps, such as a commentary or this specification; or mapping possible worlds, such as a thought experiment or an aporia). A new type is named by its coordinates and placed; it need not appear on any prior list.
Exemplars, placed: a research program (future-looking, collaborative by invitation, world-mapping: the Thermodynamic Realism modules); a diagnostic essay (present, solo, world-mapping: "The Shape of the Disagreement"); a manual or scaffold (present, solo, world-mapping, plain register: "The Manual for When You Don't Know What to Believe"); a condensation or skeleton (present, solo, world-mapping, maximal signal density: "What I Know"); a manifesto (future-looking, solo toward adversarial, position-stating); a reference such as a glossary or a map page (retrieval, not argument).
2.2 The profile, declared as tolerances
The type sets the profile, and the profile is best stated as a tolerance per dimension rather than a vague emphasis. instrument Engineering does not demand zero error; it demands error within a tolerance the purpose sets, and "high fidelity" has been doing the work of "within tolerance." A document declares, for each of the four dimensions, how much it may carry: the accuracy error it can tolerate (usually near zero), the resolution floor, the coverage boundary, the calibration tightness. The Manual runs a loose resolution tolerance and a tight accuracy tolerance over a narrow coverage; a proof runs a near-zero accuracy tolerance over a tiny coverage. Two consequences. The smallest adequate set of guidelines, the floor, is just the loosest tolerance still adequate to the purpose, which is what scales the apparatus down rather than up. And a declared tolerance is checkable against the finished document, which partly answers the bad-faith worry of §8.3: a tolerance reverse-engineered to excuse a sloppy document is at least visible, because it was stated rather than hidden.
| Document (type) | Scoped territory | Target reader | Tolerances, by dimension | Register | Instruments |
|---|---|---|---|---|---|
| A TR module (research program) | The framework's claims and their support | A motivated reader willing to check the work | Accuracy near-zero error; resolution tight; coverage broad but bounded; calibration tight | Specification | Nearly all: three-axis labels, per-claim trace and scope, falsification, spine and cross-links, glossary, references with status, labeled math |
| "The Shape of the Disagreement" (diagnostic essay) | The structure of a contested debate | An engaged general reader | Accuracy near-zero; resolution medium; coverage selective; calibration tight | Specification, restrained | Most: purpose-built chips, scope box, callouts, references; no formal trace, no math |
| "The Manual" (manual / scaffold) | Epistemic vertigo and how to act within it | A frightened layperson at a bad hour | Accuracy near-zero on a narrow coverage; resolution loose; coverage deliberately narrow; calibration plain | Plain, warm | Almost none. Calibration, framing, fencing, and honest limits are carried in prose (§3.1). |
All three are high-fidelity. The instrument count differs because the type, and so the tolerances, differ. The commitments in §3 hold in all three; only the instruments in §6 vary. A reader who knows the type knows what to expect, and reads the absence of apparatus correctly.
3. The invariant commitments, and how plain prose carries them
These hold for any high-fidelity map, the Manual included. Drop one and fidelity falls, whatever the profile. Each names the dimension it protects.
C1. Be accurate to the scoped territory. invariant A false claim within the document's own scope is the primary fidelity fault; no register excuses it. (accuracy)
C2. Calibrate. invariant Expressed confidence tracks reliability; a guess never wears the dress of a result. State credence in words that fit, reserving "certain" for tautology, mathematics, and direct entailment, and hedging only where the conditionality is real. How a credence is earned, so a label is not pulled from thin air, is §5.8 (calibration)
C3. State the limits. invariant Mark where the map stops and what it does not cover. An unstated boundary is a coverage claim you did not earn. (calibration, coverage)
C4. Frame reflexively. invariant State what the framing foregrounds, what it suppresses, and why. No framing is neutral, so expose the choice rather than perform neutrality. (accuracy about the map itself)
C5. Fence the non-derived. invariant Mark personal views, speculation, and out-of-scope commitments as not results, so they cannot inherit the weight of the derived.7 Without the fence, a reader breaks your metaphysics and believes the work fell with it. (accuracy, calibration)
C6. Pure signal. invariant Remove any word whose deletion does not change a claim, and more broadly remove noise at every level (§1.5). Transition-only and atmosphere sentences are the common offenders. (resolution, uptake)
C7. Hold semantic and linguistic precision. invariant One term, one referent per document. Choose the exact word, split an equivocation rather than trading on it (the "one word, three loads" failure is a fidelity fault, not a style quibble), define each specialized or coined term on first use, and keep the grammar and structure clean, since errors there are noise. (accuracy)
C8. Lead with the answer. invariant The conclusion or the disagreement comes first, qualifications after. (uptake)
C9. Apply rigor recursively. invariant Scrutinize the material, the response, and the checking itself, looping until a pass stops changing the answer, and stopping at loop-as-theater. Scrutinize concessions as hard as assertions; a concession survives only if you would still make it with the pressure reversed. (accuracy, calibration)
C10. Scrutinize symmetrically. invariant Hold absent parties (authors, interlocutors, sources) to the same standard as yourself. Calm prose, numbered points, and the absence of snark are not proxies for correctness; evaluate the substance. (accuracy)
C11. Distinguish the epistemic sources. invariant Keep separate what is derived, imported, an interpretive bridge, asserted, and merely reported, and mark the point where a chain changes status (for example, where a derivation gives way to a premise). (calibration)
C12. No genetic fallacy. invariant Judge a claim or a contribution on its content, not its origin, human or AI. (accuracy)
C13. Credit and license. invariant Credit collaboration specifically, and state reuse terms. This is honesty about the map's provenance rather than a fidelity dimension, but it is non-negotiable across the corpus. (provenance)
3.1 The same commitments, carried in plain prose
The claim that the commitments are invariant is earned only if they can be met without their instruments. The Manual meets them in prose, which is what makes "invariant" demonstrable rather than asserted. A few, with the Manual's own sentences:
| Commitment | Formal instrument | How the Manual carries it in prose |
|---|---|---|
| C2 Calibrate | Confidence labels (§5) | "I find it reassuring that the impulse keeps recurring. Not because recurring makes it true": it states the basis and refuses the unearned inference. |
| C3 State the limits | Scope box (I3) | "If what you are carrying is heavier than doubt ... this page is not the thing you need": it marks exactly where the map stops. |
| C4 Frame reflexively | Named framing section (I3) | The closing "Why this reads differently from the rest of this blog" names the framing choice and its cost. |
| C5 Fence the non-derived | Fence / firewall box (I3) | "The impulse behind it is not mine to claim," and "this is not a complete philosophy": it fences ownership and scope. |
| C11 Distinguish sources | Status labels (§4.2) | It separates the settled from the personal in plain words, treating "people and animals feel things" differently from its own interpretive moves. |
4. The claim-label vocabulary
Labeling is the instrument (I1) that makes calibration (C2) and source-distinction (C11) visible at the level of the individual claim. instrument It is always proper to the document: the categories and the markup are chosen to fit the document's type and declared in its legend. The corpus has two markup mechanisms, the heavier badge set and the lighter outline set; they differ in visual weight, which is itself a register choice, so both are legitimate and a document picks the one that fits its type. What must hold within any one document is that the label set is coherent, declared, and used consistently. Labels run on three independent axes.
4.1 By kind (the canonical set)
What type of move a claim is. Color tracks kind; cooler and darker are the more secure moves, warmer the more conditional.
Axiom bedrock, asserted not derived. Premise an empirical fact relied on but not derived within the work. Theorem an entailment from the base. Corollary a descriptive consequence of theorems. Definition a term fixed by stipulation. Characterisation a proposal about what a familiar word picks out, weaker than a stipulated definition. Functional what a feature does and why it exists, without claiming to exhaust its felt or intrinsic character. Analogical a reach beyond strict entailment, flagged as such. Falsification a refutation condition, not a result.
4.2 By status (reused across documents)
How secure a claim is and where its support comes from: imported result (proved in the cited literature, taken as is); derived (follows from the work's own assumptions by construction); bridge (an interpretive identification, defensible but not proved); conceptual (a formal representation or definition, not a derived result); open (not established, flagged); supported, borrowed (independent evidence is consistent but was not gathered to test this claim); specified, not run (a protocol exists, no test conducted); analogical, untested (a hypothesis by extension, not fit to data); live bet (a distinctive claim asserted with a named crux); resolved / retired / corrected (moved out of the open set and recorded, including moves against the work).
4.3 By confidence
A third axis, orthogonal to the first two: how much credence the claim earns. Status records where support comes from; confidence records how much the claim has given that support, since two "derived" claims can differ in confidence when their premises differ in security. Confidence is expressed in words proportioned to the basis rather than on a forced numeric ladder. The method that keeps a confidence label grounded rather than arbitrary is §5.
4.4 The normative label, stated precisely
4.5 A flag is not a verdict
A label recording an open status, or a diagnostic that flags a question, opens an investigation; it does not close one. A flag treated as a conclusion is self-confirming and uncorrectable, because the flagged item stops being examined. Keep a provisional label provisional until the investigation that would change it has actually been run.
5. Confidence labeling: how a credence is earned
C2 is the commitment to calibrate. This is the method that keeps a confidence label from being pulled out of thin air. The explicit labels are an instrument, used when the profile calls for them; the discipline holds even when confidence is expressed in prose, as the Manual does when it finds the recurrence of an idea reassuring "not because recurring makes it true." A tension worth stating: the rule is not a fixed five-rung ladder applied to every claim, because a forced tier misreports claims whose conditionality does not fit it. What defeats thin air is not a mandatory number, it is grounds plus sensitivity: every confidence claim names what it rests on and what would move it.
CL1. Name the basis. invariant State what the credence rests on, an entailment, an imported and replicated result, a single study, several independent lines, an analogy, or a bare guess, not just a level. A level with no basis is the thin-air failure.
CL2. Separate the axes of uncertainty. invariant Logical (is the inference valid?), empirical (are the premises true?), and definitional (is the key term doing the work?) confidence can diverge; report them separately when they do. A claim can be near-certain given its premise and only as strong as that premise.
CL3. Separate confidence in the claim from confidence in your statement of it. invariant High credence that a result exists is not high credence that you have recalled or derived it correctly. For niche facts, numbers, and citations, verify before assigning the label, because confident recall confabulates. (C11 at the level of the label.)
CL4. Count only independent support. invariant Stack confidence from lines of evidence that are actually independent; correlated sources are one line, not several, and convergence raises credence only to the extent the routes do not share a failure mode.
CL5. Use words that fit, anchored not arbitrary. invariant Proportion the language to the basis, reserve "certain" for tautology, mathematics, and direct entailment, and never attach a number you cannot ground. Anchor the vocabulary so a word means the same thing across the corpus (for example: established, well-supported, a defensible bet, a conjecture, a guess), without forcing a claim onto a rung where its conditionality does not sit. (This anchoring sits in tension with the no-ladder point above; see §7.)
CL6. Carry a sensitivity. invariant State what evidence or argument would move the credence, and in which direction. A credence with no sensitivity is unfalsifiable and is the thin-air failure wearing a number. (The flag-is-not-a-verdict discipline, §4.5, applied to confidence.)
CL7. Keep the credence and the decision separate. invariant Where being wrong is asymmetrically costly, say so as a separate statement; do not fold risk-aversion into the credence, or the label stops tracking the world and starts tracking caution.
5.1 Worked examples
Each takes a real claim from the corpus, shows a label pulled from thin air, then a grounded one.
Claim: lying costs more than truth-telling. Thin air: "certain." This violates CL5 (it is no tautology or entailment), CL2 (it fuses a secure claim with an open one), and CL6 (no sensitivity). Grounded: near-certain on the dual-map architecture, since a deceiver maintains an accurate map and a presented one and the second is extra structure carrying a description-length and dissipation cost; a live bet that every real act of deception instantiates that architecture; would move if a cost-instrumented experiment showed the cost eliminated rather than relocated.
Claim: entropy selects for truth (the persistence selection principle). Thin air: "a law." Grounded: a statistical tendency, not a law, since the selection covariance is non-positive only under recurrent non-summable hazard and vanishes in the dark-room limit; conditional on non-zero variance and finite resources; would be refuted by a lower-fidelity system reliably outpersisting a higher-fidelity one under matched, non-dark-room variance.
Claim: a cited prediction-dissipation bound has a particular form. Thin air: stating the exact relation from memory at high confidence. Grounded: confident the result exists and that the inequality runs in the relied-on direction; not confident of the precise form, instantaneous single-step versus full past-and-future, without checking the source, so the citation is marked Confirm and the indices are verified before publication. (CL3: confidence in the claim is not confidence in the recall of it.)
6. The instruments, and the commitment each serves
Apparatus, applied as the tolerances (§2) call for it. Each instrument names the commitment it implements at high resolution and the profile that warrants it. A document may drop any of these and stay high-fidelity if its tolerances do not need it; the Manual drops nearly all.
Relation to standard academic conventions. Several instruments resemble familiar moves, a limitations section, subsections, a methods statement. The difference is granularity and enforcement. The conventions operate at the document level; this apparatus labels at the level of the individual claim, so a reader can break one wall without the building falling, and it requires what the conventions leave optional: a break-condition on every load-bearing claim, a fence around every non-derived commitment, a status on every equation, and a calibrated confidence on every claim that earns one. It is also notation-agnostic, implementable in LaTeX, Markdown, or HTML, because it specifies what must be labeled and exposed, not how to typeset it. The question it answers is not which format but whether the per-claim discipline is applied.
I1. Claim labels, on the three axes of §4. instrument Serves C2 and C11. The mechanism is shared; the semantic assignment is per-document and declared in the legend, so a color never means two things across two papers. Profile: analytical and formal documents.
I2. References, with a per-citation verification status. instrument Serves C1, C11, C12. Cite where a claim depends on a source, link it, distinguish primary from secondary, and mark each entry Verified (checked against the source, dated), Standard (canonical, cited from knowledge), or Confirm (identifier not pinned, verify before publication), so the list never implies uniform confidence. Profile: any document resting on external sources.
I3. Scope box and callouts. instrument The formal carriers of C3, C4, and C5. The Manual delivers the same content in prose (§3.1).
I4. Per-claim trace and per-claim scope. instrument Serves C11 and C3: show each derived claim's parents, and attach its limits at the claim, not only at the document. Profile: derivation-heavy work.
I5. Glossary, constitutive where terms are coined. instrument Serves C7. Any document that introduces or relies on specialized terminology includes a glossary, and the glossary is part of that paper, not an optional appendix; each entry points back to where the term is grounded. A document with no special terms (the Manual) needs none.
I6. Falsification conditions. instrument Serves C3: attach to each load-bearing empirical or analytical claim the observation or argument that would break it.6 A good condition names what you would observe, under what conditions, that the claim forbids; a lazy one is unobservable or immunized against every result. The contrast is worked in the TR corpus: "a rigid monoculture surviving extreme variance" refutes nothing, because the framework's own dark-room clause already permits it, whereas the reworked condition closes the escape clauses and is comparative (a lower-fidelity system reliably outpersisting a higher-fidelity one under matched, non-dark-room variance), which is testable. Profile: documents making testable claims.
I7. Single canonical source, modular spine, cross-links. instrument Serves C1 and C7: state each claim once in a canonical place, have other parts reference rather than restate it, and declare that on contradiction the spine governs.9 Restatement drifts and grows contradictory walls. Profile: multi-document corpora.
I8. Constructed, ideally interactive, visuals. instrument Serves resolution and coverage of structure: a diagram of the dependency graph raises fidelity; a decorative image is noise under C6 and §1.5. Profile: when structure is hard to see in prose.
I9. The living edge: open-problems registry and change record. instrument Serves C2 and C3 over time. Each open item states its closing condition (an item that cannot say what would resolve it is a mood, not a problem). Keep a record of what moved, including retractions, corrections, and items that resolved against the work, so the open list is not decorative and "open" is never read as "weakly held." Profile: living research.
I10. Mathematics, loaded and labeled. instrument Serves resolution. Include the math renderer only when equations are present, and label each expression by status, imported result, derived, conceptual (a formal representation or definition, not a derived result), bridge, or open, so an equation never borrows the authority of a proof it has not earned. Unlabeled math invites reading a definition, or a decoration, as a theorem; decorative math that earns no label should be cut or demoted. Profile: formal content.
I11. Navigation at two scales. instrument Serves coverage and uptake: a table of contents within a document, and cross-links, suggested reading orders, and an at-a-glance architecture across a corpus. Profile: long or modular work.
I12. The type line and versioning. instrument Serves framing and calibration over time: a sub-sub-title naming the document's type (§2), a version line in full integers (Version 1, Version 2), and within a revised document a mark on what is new or changed. Profile: all, since both are nearly free.
7. Self-instantiation: this document against its own rules
A specification that violated its own guidelines would be self-refuting. Compliance is profile-relative (§2): this document's type is meta-specification and its tolerances call for the full analytical apparatus, so it is held to the commitments and to the instruments that profile warrants. It passes most and fails honestly on some, and reporting the failures, including substantive ones rather than only unfinished work, is itself C9 and I9.
Self-check, substantive failures and compromises. Three, none of them merely "unfinished." First, the four-dimension cut (§1.2) does load-bearing work while under-defended: it is asserted as "at minimum," not argued to be the right or complete cut, and §8.6 concedes this while §1.2 still presents it with more weight than that concession warrants, a small calibration miss inside the document. Second, CL5 anchors a confidence vocabulary one sentence after disclaiming a fixed ladder; the anchors function as a soft ladder while the text denies a ladder, and this is a live tension retained because the alternative (no anchors) invites thin air, not a clean resolution. Third, the invariant-versus-instrument binary has fuzzy members: C13 is grouped with the invariant commitments but is a provenance norm, not a fidelity dimension (and is flagged as such in place), and I12 is labeled an instrument while sitting close to invariant.
Self-check, unfinished. The dependency-graph visual (I8) is not built (deferred, §9). Uptake (§1.1) is unmeasured (§8.5). Several references are Standard or Confirm rather than Verified, marked as such.
8. Open problems and honest limits
8.1 The bootstrap. open Most fidelity checks are themselves run on maps: you verify your map of the territory using other maps, and direct territory-contact is rare. "Raise fidelity" partly reduces to "raise the rate and independence of territory-contact," which the guidelines organize but do not manufacture.
8.2 No guarantee of causation. open The specification is a falsifiable bet, not a theorem: a corpus built this way should be harder to break in any single place and easier to repair. That is only loosely measurable, and the commitments could correlate with fidelity without raising it.
8.3 The vacuity risk. open Purpose-relativity is the part most open to abuse: any document can be defended as "high-fidelity to some purpose." The scope-box guardrail, the noise definition (§1.5), and the declared tolerances (§2.2) block the worst of it, but policing whether a stated purpose is honest, rather than reverse-engineered to excuse the document, stays a judgment the specification cannot fully automate.
8.4 Profile and type selection is a judgment. open The model says the profile is purpose-relative but gives no algorithm for choosing it or its tolerances, and the type space is open, so a document may sit between named regions. Which dimensions to favor, and what to call the document, stay human calls the specification informs and does not make.
8.5 Uptake is unmeasured. open Fidelity is defined against the model the map affords its target reader, but the specification governs the map and only proxies that afforded model through register and navigation. A proxy can be satisfied while the target is missed. A lightweight uptake test is named in §9.
8.6 The dimension list may be incomplete. open Whether accuracy, resolution, coverage, and calibration exhaust fidelity, or whether a fifth dimension is needed, is not settled here.
8.7 The target itself is a choice. open Fidelity-to-territory is assumed as the goal (scope box). A document optimizing for persuasion, comfort, or beauty would reject most of this, and the specification cannot justify its own target from inside.
9. Scope of this version, and what is deferred
This is Version 1, and it is deliberately scoped. It establishes the standards layer: the fidelity model, the commitments, the instruments, the three-axis label vocabulary, confidence labeling, profiles stated as tolerances, and the type space. Five things are named here and deferred to later versions, so that each version stays small enough to be checked against itself, the discipline applied to its own release cadence. Deferring them is a scope decision, not an oversight.
1. A failure-mode catalog. A systematic enumeration of how a map loses fidelity, mis-scope, mis-map, equivocation, unearned authority, unfenced import, broken dependency, stale claim, miscalibration, each mapped to the dimension it degrades and to a detection method. §1.5 defines noise; this would make it operational.
2. The mechanically-checkable set. An explicit list of which commitments a tool could check statically (needless words, undefined-but-used terms, unlabeled equations, load-bearing claims with no break-condition, confidence labels with no sensitivity, claims labeled above the status of their support) versus which require human judgment. This is the precondition for any linter, and it draws the boundary of the whole approach: the checkable surface is small, and the core, whether a claim is true, whether the framing is right, whether the purpose is honest, stays judgment.
3. The collaboration layer. A contested status, a crux-request format (a challenge against one labeled claim, naming the mechanism that would defeat it, not a disagreement in general), and a resolution ceremony for live bets (a bet that resolves moves to resolved, retired, or corrected with a dated note and a pointer to what moved it). This is what turns a solo notation into a shared discipline; the TR counterexample protocol is its prototype.
4. An uptake test. A lightweight, honest method for checking whether a document's afforded model actually forms in its target reader, since §1.1 defines fidelity against that model and §8.5 currently only proxies it.
5. Tooling. A label-aware editor or compiler, a dependency-graph generator (the visual §7 still lacks), and the linter that item 2 would specify.
None of these changes the commitments; they extend the discipline outward from the standards layer it now has.
10. Summary
- A document is a map; the meta-goal is fidelity between the model the map affords its target reader and the purpose-scoped territory (§1.1, scope box).
- Fidelity has four dimensions, accuracy, resolution, coverage, calibration, which trade off, so a document picks a point on a frontier internal to fidelity (§1.2, §1.3). Signal raises fidelity; noise, at any level, lowers it (§1.5).
- Fidelity is the constant target; the profile flexes, named in one line by the document's type and stated as a tolerance per dimension, and the tolerances select the instruments. The apparatus is instrumental, not constitutive (§1.4, §2).
- A set of commitments is invariant across all profiles (§3) and is shown discharged in plain prose, not only in apparatus (§3.1); the apparatus is applied only as the tolerances warrant (§6); claims are labeled on three axes, kind, status, and confidence, with the normative label chained to physics rather than free-floating (§4), and confidence labels grounded by basis and sensitivity rather than pulled from thin air (§5).
- Mathematics, where present, is labeled by status so it cannot borrow the authority of proof (I10).
- The specification is held to its own rules and reports where it fails, including substantive compromises (§7); the honest limits are real (§8); and the version is deliberately scoped, with five extensions named for later (§9).
References
- Korzybski, A. (1933). Science and Sanity: An Introduction to Non-Aristotelian Systems and General Semantics. The Institute of General Semantics. Source of "the map is not the territory," and of the general-semantics emphasis on language-to-territory fidelity behind C7. Standard
- Box, G. E. P., and Draper, N. R. (1987). Empirical Model-Building and Response Surfaces. Wiley, p. 424: "Essentially, all models are wrong, but some are useful." The bare "all models are wrong" appears earlier in Box, G. E. P. (1976), Science and statistics, Journal of the American Statistical Association, 71(356), p. 792. Verified · June 2026
- Dupré, J. (1993). The Disorder of Things: Metaphysical Foundations of the Disunity of Science. Harvard University Press. On the purpose-relativity of classification, used for the claim that there is no single best map. Standard
- Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27, 379 to 423 and 623 to 656. Source of the signal-and-noise framing and of "fidelity" as a reconstruction criterion. Standard (volume and year certain; exact page ranges from bibliographic knowledge)
- Strunk, W., and White, E. B. (1959). The Elements of Style. Macmillan. "Omit needless words," in the Principles of Composition; the rule number is edition-dependent (Rule 13 in some editions, Rule 17 in others). Verified · June 2026
- Popper, K. (1959). The Logic of Scientific Discovery. Hutchinson. English edition of Logik der Forschung (1934). Falsifiability as a demarcation criterion, used for I6. Standard
- Hume, D. (1739 to 1740). A Treatise of Human Nature, Book III, Part I, Section I. The is-to-ought passage, behind the fencing commitment (C5) and the chained, non-free-floating reading of the normative label (§4.4). Standard
- Tetlock, P. E., and Gardner, D. (2015). Superforecasting: The Art and Science of Prediction. Crown. On calibration as a measurable property of judgments, behind C2 and §5. The academic source is Tetlock, P. E. (2005), Expert Political Judgment, Princeton University Press. Standard
- Hunt, A., and Thomas, D. (1999). The Pragmatic Programmer. Addison-Wesley. "Don't Repeat Yourself" (DRY), used for single-canonical-source (I7). Standard
Standing of this document. This is a specification, not a proof, and not a claim that this is the only way to map a territory well. It is a falsifiable bet: that documents built to these commitments are harder to break in any single place and easier to repair, and that a document's fidelity tracks how well it maps its purpose-scoped territory rather than how much apparatus it carries. Corrections and counterexamples are welcome and change the specification.
One-sentence collapse: An ultra-low-bandwidth condensation of the paper above:
ReplyDelete"Defining a document's fidelity by the structural correspondence between its purpose-scoped territory and the model it affords a target reader, the meta-specification establishes that high-fidelity text is achieved not by accumulating academic apparatus, but by enforcing invariant epistemic commitments—including strict accuracy, tight calibration, and recursive scrutiny—while flexing situational analytical instruments according to a declared tolerance profile to minimize communicative noise."