Domain Lenses — the bibliographic metabolism of each field

Framing. Philosophy is the discipline that reads its own history as live conversation, not settled record. A physicist does not read Newton to learn mechanics; a philosopher still reads Plato, Aristotle, Kant, and Hume in the original to do the work itself. The field has no textbook that makes the sources obsolete, because in philosophy the argument is the primary text. This makes it, bibliographically, closer to scripture than to science: cumulative, not superseding. The governing image is Whitehead's — European philosophy as "a series of footnotes to Plato" — and the practice is a 2,500-year conversation with the dead.

Philosophy has one of the longest citation half-lives of any field — humanities monographs are used on average ~40 years after publication, versus ~2–3 years in the health sciences. But the aggregate understates the split; the literature has two metabolisms bolted together:

• Primary texts: near-zero obsolescence. A great primary work never ages out. It can be argued against, but it cannot be refuted into worthlessness — a defeated position remains a live option someone will revive. The canon is cumulative, not a moving frontier.
• Secondary literature: ordinary academic decay. Commentary, "companions," survey articles, and textbooks date like any scholarship — interpretations get superseded, fashions turn, field-specific empirical claims rot fastest.

The mechanism is aporia: philosophy makes progress by deepening productive puzzlement, not by closing questions. Commentary accretes in layers around the primary text rather than replacing it.

1. The primary text — the source voice, read directly. The permanent spine of the shelf.
2. The interlocutor / counter-argument — the book that argues against one you own (Nietzsche against Kant, Hume against the rationalists). A gap here is an unanswered objection sitting on your shelf with no reply beside it.
3. The commentary / apparatus — the great secondary reading that unlocks a hard primary text. Lowest permanence, highest replaceability.

A gap is not "an old edition" — it is a missing school (no analytic counterweight to your continental shelf; Stoics but no Epicureans) or a dangling objection with no interlocutor.

This is the field where the subtract prior should be lowest of any domain. The Talmudic principle eilu v'eilu — "these and these are the words of the living God" — is the exact rule: refuted minority positions are deliberately kept on record so the argument is not lost. Primary texts: almost never subtract (duplicate editions yes, the work no). Secondary literature: the only real subtract surface — superseded interpretations, dated surveys — and even there, subtract is demotion, not deletion.

Philosophers re-read across a lifetime (the same Critique at 25 and 55 is a different book), annotate heavily — the marked-up primary text argued-with in the margins is the working instrument — and treat the reading list itself as a personal canon. The book is a sparring partner, not an information source.

• Subtract prior: very low (≈0.05–0.1). Default a philosophy book to keep forever. Never auto-subtract a primary text; confine subtracts to superseded secondary material, and demote rather than remove.
• Add rule: privilege the primary voice and the missing interlocutor over the recent title; recency is nearly irrelevant.
• Gap = a silenced side of an argument, not a stale edition.
• Serendipity: cross-tradition, cross-era jumps (a Stoic beside a Buddhist) — the field rewards the long, unexpected conversation.

Sources: Elu ve-elu (Eruvin 13b) · primary vs. secondary sources in philosophy

Framing. Most domains have a metabolism: books are born, cited, and die as knowledge advances. Religion is the domain where the primary sources never die. It is the clearest case of a two-speed library — a frozen core and a boiling shell. Treat it as the archetype for "high-canon, high-commentary" fields; other humanities (law, classics, philosophy) are weaker versions of it.

The half-life of scripture is ∞; the half-life of theology-of-the-moment is short. The canon does not merely resist subtraction — it is structurally re-consumed on a clock. The Torah is read cover-to-cover on a fixed annual cycle (Simchat Torah rolls the scroll back to Genesis the same day it ends Deuteronomy). Daf Yomi puts the entire 2,711-page Babylonian Talmud on one global schedule — a page a day, ~7.5 years per lap, unbroken since 1923. The book is never finished; finishing is the trigger to restart. Around this frozen core accretes an unbounded commentary tradition (midrash, tafsīr, bhāṣya, patristics). The field grows by lamination, not replacement.

• A new commentary / interpretive layer — the normal churn. Cheap, infinite, low-stakes.
• A rival tradition — a whole other core (the Sutras next to the Gospels next to the Qur'an). The highest-value add: not more of the same, a different frozen center.
• A translation — a near-original event, because the untranslated source was inaccessible scripture.
• A primary source recovered — Nag Hammadi, the Dead Sea Scrolls, a Geniza fragment: an add that retroactively enlarges the canon.

A gap is specific: an untranslated source, a tradition present only as its critics describe it, a contemplative practice with no text in the collection.

The sacred core is never subtracted; often it may not even be physically destroyed. The genizah is the proof: a worn book bearing the name of God is retired to a depository and buried like a body. Subtraction here is reverent retirement, not deletion. The dark twin is the forced subtract — heresy, censorship, the Index, book-burning. The legitimate everyday subtract is narrow: devotional ephemera (this season's sermon collection, the popular apologetics of a decade ago). The core is inviolable; only the outermost devotional skin flakes off.

Reading is practice, not acquisition: lectio divina, the annual Torah cycle, svādhyāya, dhikr, the memorization of the entire Qur'an (ḥifẓ — the text carried inside the body). The book is a sacred object. Knowledge is demonstrated by depth of re-reading, not breadth of titles. The antilibrary instinct inverts here: the prized book is the one read a hundred times.

• Subtract prior: near-zero on the core. Any primary scripture, canonical commentary, or rival tradition's core = permanent keep. The strongest keep-prior of any domain.
• Subtract target: the devotional skin only. Churn interpretive/popular ephemera; retire, don't delete — mark as "genizah," never destroy.
• Add priority: rival cores > recovered primary sources > new translations >> more commentary.
• Serendipity: adjacency across traditions — the next tradition's frozen core, not a new take on your own.

Sources: Burying the Genizah · Daf Yomi · Cairo Genizah (Cambridge)

Framing. Every domain has a metabolism: a rate at which its literature decays into obsolescence. Mathematics sits at the extreme slow end — the most Lindy field that exists. A physics result can be overturned; a drug trial gets retracted; a business book rots in five years. A theorem, once proved, is true forever. This single fact should invert Khoj's default subtract behavior for this domain.

Euclid's Elements (c. 300 BC) is the world's oldest continuously-used textbook; its proofs still hold. What ages is never the result: it is notation (Newton's fluxions vs. Leibniz's dx), rigor standards (ε-δ retro-fitted under Cauchy/Weierstrass), and pedagogy. The content is immortal; only its packaging dates. Hardy, Rudin's Principles, Spivak's Calculus, Munkres' Topology are the permanent canon. Erdős captured the metaphysics: he imagined "The Book" in which God keeps the single most beautiful proof of every theorem — "you don't have to believe in God, but you should believe in The Book." Aigner & Ziegler published an earthly shadow (Proofs from THE BOOK); its proofs will never need updating.

1. A missing prerequisite — the sharpest add. Math is a strict dependency graph; a gap is a broken edge in the prerequisite DAG blocking everything downstream.
2. A new area/theory to enter (category theory, a branch you've never touched).
3. A better exposition of something you already own (Axler's Linear Algebra Done Right over a determinant-first slog).
4. A foundational text you've been working around instead of through.

You essentially never subtract a result. Obsolescence ≈ 0, and that emptiness is itself the lens's finding. The only legitimate subtracts are exposition-level, never content-level: replace a clumsy exposition with a cleaner proof of the identical theorem (a swap, not a deletion); a genuinely superseded edition (housekeeping); a duplicate covering the same material worse. There is no "this theorem was disproved" subtract. Compare a tech shelf where half the books are dead in three years — different physical laws of decay.

Mathematicians do not read books; they work through them, pencil in hand, doing every exercise — a book is a set of problems to be earned. They re-read the same slim classic for years. Coverage is worthless; understanding is everything.

• Subtract prior ≈ 0 (maximal Lindy). Default to keep. Age is a positive signal — invert the recency penalty other domains use.
• Only subtract at the exposition layer, and only as a swap (cleaner proof of the same result). Flag "superseded edition" and "duplicate coverage" as the sole automatic subtracts.
• Adds are gap-driven, not novelty-driven — fill the prerequisite DAG bottom-up.
• Serendipity = a beautiful proof, not a new fact.

Sources: In Search of God's Perfect Proofs (Quanta) · Proofs from THE BOOK · Euclid's Elements

Framing. The natural sciences are the field that most explicitly disowns its own past. Their epistemology is disposable-by-design: a claim earns its place by surviving replication and is meant to be superseded. Popper's "graveyard of dead theories" is not a lament here — it is the operating principle. Their literature splits cleanly into two metabolisms that decay at opposite rates; get the split right and the subtract policy almost writes itself.

The primary literature ages fast: a 20-year-old bench paper is often superseded or quietly falsified. The durable layer decays slowly or not at all: the review article, the definitive textbook, the rare classic. Citation half-life (surprisingly) is rising, not falling — the science-and-technology average grew ~7.0 → ~7.4 years over 1997–2013 as digitization made old work findable. But field variance is the real signal: biology's half-life is lengthening (foundations compound), chemistry/engineering shortening (faster frontier). Physics runs on the preprint pulse — arXiv is the frontier's nervous system, in symbiosis with reviews that consolidate the churn.

Four tiers, ascending durability: (1) the frontier preprint/paper — high-signal, short-lived; (2) the review article — the single highest-leverage add, cited more than primaries precisely for curating a field and flagging its gaps; (3) the definitive current-edition textbook; (4) the classic (Darwin's Origin, Feynman's Lectures) — durable as thought, not reference. A gap is a subfield whose reviews you cannot yet read.

Primary perishes, synthesis and classics endure. A superseded textbook edition is dead weight (the cleanest, highest-confidence subtract in any domain). A primary paper absorbed into a later review is subtractable as reference but retains historical value if foundational (Watson–Crick 1953 kept for what it founded). The unread frontier preprint that never became canonical is pure subtract. Never subtract a review or a classic on age alone.

The working rule: "read the reviews, skim the primaries." The review is the map; the primary is terrain you visit only when the map sends you. Textbooks are treated as versioned software — the edition number matters, the old edition is discarded without sentiment. Preprints are read as news. Practitioners keep the syntheses, not the primaries.

• ADD priority: current-edition textbook > review of a gap subfield > the field's classic > frontier preprint (only for domains the reader actively works).
• KEEP: all reviews, all classics, current textbook editions — indefinitely; these appreciate.
• SUBTRACT: superseded textbook editions (hard); stale unread frontier papers now in a review (soft); absorbed primaries — unless flagged foundational.
• SERENDIPITY: a review from an adjacent field — the cheapest, densest way to buy into a new literature.

Sources: citation half-life rising (arXiv) · arXiv–review symbiosis · preprints (Ithaka S+R)

Framing. Medicine is the domain where the antilibrary's usual bias — prize the unread — collides with a hard constraint: in clinical knowledge, the unread that is out of date is not neutral latency, it is a hazard. This is the domain where SUBTRACT dominates and does moral work. Books here are not to be accumulated; they are to be kept current or removed.

The classic estimate puts the half-life of medical literature at ~45 years, but actionable knowledge decays far faster — commonly cited at 18–24 months. Librarians codify this as the "five-year rule" in the MUSTIE weeding method, where publication date is vital for medicine (versus irrelevant for literature or poetry). Layered on top is the evidence hierarchy: systematic review / meta-analysis > RCT > cohort > case report > expert opinion. A newer study outranks and retires an older one — the metabolism is self-cannibalizing. Yet a second clock barely moves: the humanistic spine (Osler's 1892 Principles, narrative medicine, Nuland/Gawande/Verghese) ages like philosophy — it encodes judgment and humanism, not fact.

The current guideline, the recent systematic review, the latest edition of the definitive textbook (Harrison's, Robbins — edition matters more than title), and the durable humanistic classic. A gap is a live clinical area where the shelf holds only a superseded edition, or where the humanistic layer is absent.

This maps to MUSTIE (Misleading, Ugly, Superseded, Trivial, Irrelevant, Elsewhere), activating M = Misleading and S = Superseded with unique force. In most domains a superseded book is merely useless clutter; in medicine it is actively dangerous — a prior-edition dosage, a reversed guideline can injure whoever trusts it. Health-sciences libraries weed hardest of all. The subtract prior is aggressive, safety-driven, non-negotiable: an out-of-date clinical edition earns removal simply by being out of date.

They keep the current edition only and discard the previous on arrival. Increasingly they don't keep the shelf at all — point-of-care tools (UpToDate, DynaMed) have displaced the textbook, precisely because they solve the currency problem. Three registers, three metabolisms: textbook (slow framework), guideline (medium, superseded constantly), journal/point-of-care (fast frontier).

• ADD: current-edition references, recent systematic reviews/guidelines, enduring humanistic classics.
• KEEP: humanistic/narrative works indefinitely; clinical works only while current.
• SUBTRACT — aggressive, safety-weighted: flag any superseded clinical edition as Misleading, not merely stale. Weight the subtract score by harm potential.
• SERENDIPITY: confine surprise to the durable layer (history of medicine, physician-memoir). Never serendipitously surface stale clinical fact.

Sources: MUSTIE / 5-year rule · Briscoe health-sciences weeding · The half-life of the truth

Framing. Law is the field with the most split bibliographic personality. It runs two clocks at once. The practical layer — statutes, regulations, cases as authority — is the most perishable literature in any professional library: a book can go dead overnight. Yet the same discipline reveres a canon that outlives nations: Blackstone (1765), the foundational constitutional cases, the jurisprudential greats. Any subtract policy that treats it as one metabolism will be wrong half the time.

• Supersession. Statutes/regulations are current-edition-only; "old law is dead law." An amended section is a live liability if cited. Codes are kept current by the pocket part (West, 1916) and the loose-leaf service.
• Overruling. A case can cease to be "good law" — its truth-value flips without the physical book changing.
• Jurisdiction. Law is territorial; a superb California treatise is near-useless in New York. Current, correct, and still worthless across a border.
• The exception that dominates. Constitutional canon, landmark cases, and jurisprudence (Hart, Dworkin, Holmes) endure for centuries — read for reasoning, not authority, so overruling doesn't retire them.

The current statute/code edition for your jurisdiction; the leading treatise (Wright & Miller, Corbin); a landmark case; a foundational jurisprudence text. A gap is a missing jurisdiction, an unread doctrine, or the canonical author behind a field you know only operationally.

Aggressive on the practical layer — superseded code editions, prior-edition treatises, overruled cases are actively dangerous to keep as reference. Preserve the jurisprudential/landmark canon — never subtract Blackstone for being old; age is its credential. The dividing question is never "how old?" but "is this cited as authority, or read as reasoning?" Authority perishes; reasoning endures. And law is the one domain with a native, industrial citator: Shepard's / KeyCite tells you per-point whether a case is still good law — a red flag = bad law. A rare gift: an external, machine-readable "should this be subtracted?" oracle (a strong prior, not an auto-purge).

They don't "own" a library so much as maintain currency: check the pocket part before trusting a code section; file loose-leaf updates; Shepardize every case before citing. Verification-before-use, not accumulation. Sharp distinction between reporter (raw cases), treatise (durable doctrine), and casebook (disposable pedagogy).

• Domain-keyed split prior: run two subtract engines, routed by the authority-vs-reasoning test.
• Citator-as-subtract-signal: where an external "still valid?" oracle exists, use it rather than inferring from age. Model Khoj's general subtract signal on Shepard's.
• Edition-supersession rule: when a newer edition of the same work enters, flag the prior for subtract (holding the canonical exceptions).
• Jurisdiction as a gap dimension: serendipity = an adjacent jurisdiction or unread doctrine.

Sources: Shepardizing / good law · Pocket part · Legal treatises (Harvard)

Framing. History is the two-metabolism field. Every other domain has one clock; history runs two, in opposite directions. The primary layer — documents, letters, memoirs, the archive — appreciates with age: finite, non-reproducible, scarcer and more precious as time recedes. The secondary layer — interpretation, the monograph, the synthesis — depreciates, because each generation must write its own history. A discipline that treats its own past outputs as revisable is the domain that most refuses to throw them away.

A 1950s account of the French Revolution is dated as truth; the 1790s pamphlet it quoted is more valuable in 2026 than in 1950. Aging does two opposite things depending on layer — and the depreciating layer never hits zero. A superseded interpretation transmutes: 19th-century historiography, once wrong-as-analysis, is now a primary source on the 19th-century mind. The field's memory of its own mistakes is load-bearing.

Four species: (1) a primary source — always accretive, the highest-value add; (2) a new interpretation; (3) a synthesis; (4) a revisionist counter-account. A gap is specific and moral: an unheard perspective, a neglected archive, the losers' side, "history from below" — absent voices and unread documents, not missing topics.

• NEVER subtract a primary source. Permanent, appreciating, irreplaceable — removal is destruction, not decluttering.
• Secondary interpretations are layered — but you demote, not delete. A superseded synthesis changes shelf: it stops being "truth about the past" and becomes "historiography" — evidence of how the field once thought. Subtracting is really re-classification (truth-shelf → record-shelf). Even "debunked" is suspect — a discredited account documents the era that believed it.

Historians are archive-people. They footnote-chase (a book's value is partly its apparatus, the trail into the primary record), hold primary vs. secondary as a first-class distinction, and treat the object as evidence — provenance, marginalia, a specific edition are data, not damage. A working archive, curated for traceability, not tidiness.

• Strongest "keep" bias so far: primary = permanent, never a subtract candidate.
• Two-tier metabolism flag: tag each book primary/secondary — aging raises the first, lowers the second.
• Demote, don't delete: the subtract action for a superseded secondary work is re-shelving to historiography. Model a "record shelf" distinct from removal.
• Serendipity = footnote-chasing; provenance as an ADD-weight input.

Sources: Historical revisionism · Primary & secondary sources · Primary source

Framing. Every other domain treats a book as a container for information that can spoil. Literature is the one domain where the book is the thing — not a delivery vehicle for a finding a newer paper supersedes. A novel is not made obsolete by a later novel. This is also, emotionally, the antilibrary's home turf: tsundoku (積ん読) and the TBR pile live here, and the whole Taleb/Eco premise was born in a literary study. The domain-keyed prior here must be the gentlest of all: release almost nothing.

• The canon (near-permanent). Bloom's one durable test: "unless it demands rereading, the work does not qualify." Calvino: "a classic is a book that has never finished saying what it has to say," and "every reading of a classic is in fact a rereading." Half-life in centuries.
• Genre / disposable fiction (fast churn). The airport thriller, the beach read — consumed once, half-life measured in a single flight.

Two refreshers cut across both: re-reading (the same object yields new returns as the reader ages) and translation (a new translation genuinely refreshes an old work — Pevear/Volokhonsky's Tolstoy, Wilson's Odyssey). And the decisive distinction: the personal canon (what you re-read) is not the institutional canon — Khoj must key off the reader's.

(1) the canonical gap (the unread Middlemarch); (2) the buzzed new novel (live conversation, high serendipity, high churn risk); (3) the newly-translated masterpiece; (4) the unrepresented tradition — the gap that is a shape, not a title (no African fiction, all Anglophone). The richest gaps are tradition-shaped.

Fiction resists utility-based subtraction; the science move is unavailable. The prior splits hard: never subtract the canon or the personal re-read shelf, regardless of read/unread status; subtract freely the disposable once-read genre paperback. The signal is not "is it unread?" but "is it re-readable to you?" A read-and-finished thriller is a stronger subtract candidate than an unread Tolstoy. The load-bearing rule: "unread" must NEVER trigger subtract in this domain — turning that signal off is the single most important thing this lens does.

The TBR pile is a feature, kept without guilt. Shelves self-sort: a small personal-canon shelf (re-read, permanent, the keepsake) and a large churn shelf (read-once, freely lent-away). Lending is itself a signal — you lend from the churn shelf, refuse to lend from the canon.

• Subtract prior: minimal. Default to keep; only the read-once disposable is eligible.
• Unread ≠ subtract. Hard override of the cross-domain aging rule. Tsundoku is protected.
• Two-shelf model: ask "re-readable to this reader?" not "consumed?"
• Gaps are tradition-shaped; serendipity runs hot (the buzzed novel is low-risk — churn is cheap here).

Sources: Calvino, Why Read the Classics? · Bloom, The Western Canon

Framing. Every domain metabolizes its literature differently, and poetry sits at the far pole: the smallest shelf, the densest per-page, the most re-read, the most memorized, the least disposable body of writing that exists. Where a computing or medical shelf is a depreciating asset, a poetry shelf is a reliquary. It appreciates. This is the domain that most violates Khoj's default subtract instinct, and it should be hard-coded to know that.

Poetry does not age out. Sappho (7th c. BCE), Rumi (13th c.), Bashō (17th c.), Dickinson are not "still relevant" — they are load-bearing, read with no discount for age. A new translation does not obsolete Sappho; it adds a facet. The anthology tradition proves the metabolism: the Greek/Palatine Anthology preserved ~4,100 poems across 1,600 years in a single curated manuscript, and many survive only because they were anthologized. The field's native act is gathering and keeping, not discarding-and-updating. And the thinnest books hold the most — a 60-page collection can outweigh a 600-page monograph in re-reads per shelf-inch.

Adds are additive facets, never replacements: a single poet's collected works; an anthology (a whole tradition compressed — the highest-leverage add for a gap); a translation (Rumi via Barks vs. Nicholson vs. Arberry are different books); a chapbook (how living poetry enters). A gap = an unheard voice or an untranslated tradition — traditions, not titles.

This is the shelf you carry out of a burning house. Obsolescence ~0; density and re-reading make every slim volume high-value even unopened (the antilibrary principle is strongest here). The only legitimate subtract is the weak contemporary volume that genuinely didn't land after real exposure — and even then, prefer demotion. Anthologies, translations, and any pre-20th-century poet: effectively unsubtractable.

Poetry is the one domain where the book is not the primary storage medium — the body is. Readers memorize: the poem lives in you, portable, un-confiscatable (Akhmatova's Requiem survived only in friends' memories; Ratushinskaya composed on a soap bar in a camp and washed it away once by heart). Beyond that: re-reading a single poem for years, the bedside anthology, reading aloud, a reverence for slimness.

• Subtract prior ≈ 0. Override the global metabolism model; default is keep. Route subtracts only to un-landed contemporary chapbooks (demote, not remove).
• Adds are facets, not upgrades — a new translation/anthology stacks.
• The memorized poems are a collection of their own — model an invisible shelf of poems known by heart as owned inventory no other domain has.
• Serendipity is high-value — the slim, cheap, dense unit makes wandering adds nearly free.

Sources: Poetry Under Stalin: memorization · Palatine Anthology

Framing. Psychology is the largest field in this library, so its metabolism sets the house policy for the whole social-science shelf. And it has a metabolism unlike any other: it publishes findings that are later withdrawn by the field itself. No physics book gets less true; a huge fraction of social-psychology books get less true after publication. Khoj must treat this shelf as the one place where owning a book can make you actively wrong.

• The framework endures. William James's Principles (1890), Weber, Durkheim, Freud-as-literature, Goffman, Kahneman's synthesis — read for the conceptual apparatus, aging like philosophy. Social science "cumulates softly," accreting frameworks rather than converging on settled fact.
• The finding perishes — fast, and treacherously. The 2015 Reproducibility Project re-ran 100 studies; only ~36% replicated. And not obscure results: ego depletion, power posing, the Stanford Prison Experiment, the marshmallow test all collapsed or shrank. The single-study bestseller is the most perishable object in the entire library.

(1) a foundational framework text (James, Weber, Durkheim, Goffman) — a durable classic; (2) a rigorous synthesis (Kahneman, a meta-analytic review) — synthesis diversifies risk across many studies; (3) a replicated, robust result; (4) a good popular synthesis reporting the state of a field. A gap: a major framework you own nothing on.

Finding perishes, framework endures. The single-study trade paperback built on one sexy effect — when it fails to replicate, the book becomes a MUSTIE "M" — Misleading. This is the sharpest subtract category in the whole antilibrary: a debunked pop-psych bestseller is worse than an empty shelf slot, because it actively teaches false things with a scientist's authority. Never subtract the framework for looking old — James looking Victorian is the durable layer doing its job.

Two shelves. The classics shelf (James, Weber, the handbooks) is permanent, cited, re-read. The trade-paperback shelf is disposable and held lightly — scholars know the half-life of a headline finding is short and are wary of the single-study bestseller, having watched their own field's marquee results collapse. They cite the framework and are skeptical of the anecdote.

• High churn, careful keep: the fastest churn and the most durable core — route by layer, not by date.
• The replication-fragility flag: tag any book whose thesis rests on a single named effect as replication-fragile; hard-flag MUSTIE-M if that effect is on the known failed-to-replicate list.
• Framework-vs-finding classifier: foundational/synthesis → permanent keep, ignore age; single-study popular finding → perishable.
• Serendipity: bias discovery toward foundational framework gaps (the Weber you don't own).

Sources: Reproducibility Project · Ten famous findings hard to replicate

Framing. This is the domain where the literature actively lies to you, in proportion to how confidently it sells. Its central pathology is survivorship bias industrialized into a genre: take the winners, reverse-engineer their "traits," sell the traits as a recipe. The recipe fails because the graveyard companies had the same traits. Khoj must treat this shelf not as a knowledge base but as hype sediment — a thin durable core buried under fast-rotting chaff. The SUBTRACT list's single richest hunting ground.

Here recency is a negative signal — the anti-Lindy inversion. The Lindy effect says a book in print 40 years will likely last another 40; business/self-help runs the opposite metabolism. Its flagship products self-destruct on contact with time: In Search of Excellence (1982) — most of the "excellent" companies later underperformed; Good to Great (2001) — Circuit City and Fannie Mae went bankrupt; the 11-stock portfolio would have underperformed the S&P 500. Rosenzweig's The Halo Effect names the mechanism: these "scientific" bestsellers are storytelling — observers retro-fit brilliance onto shiny companies (regression to the mean). The lesson: the more celebrated the case study, the more suspect the book.

The durable core survives because it does not peddle recipes: the Lindy classics (Drucker, Deming, Taleb, Kahneman, Christensen) that teach how to think; the operator, not the guru — the practitioner memoir with skin in the game (Grove's High Output Management, Knight's Shoe Dog, Sloan) over the consultant's framework; the counter-hype book (Rosenzweig, Pfeffer). A gap = a durable primary source or a foundational discipline (real accounting, real statistics, operations) the shelf lacks under its pile of paraphrase.

Subtract on a compound recency-trend-durability signal: recent publication + trend-riding title (AI-for-X, the current productivity fad) + no track record = prime subtract. Cull the debunked case-study book (keep one as a cautionary specimen); the guru's Nth repackaging of one idea; the ghost-written CEO hagiography; the "habits/hacks" ephemera. The 95% ephemeral layer is default-subtract unless a title has already survived a decade of reprints.

They don't read them — they mine them. The genre is built for the one-idea skim (a single concept padded to 250 pages), which is why the summary economy (Blinkist, "the 4 takeaways") is a rational adaptation. Readers ride the guru cycle (fresh → saturated → stale in ~3–5 years) and trust the operator's memoir over the consultant's 2×2.

• Anti-Lindy prior (the domain-keyed rule): everywhere else, age = keep. Here, invert it — recency is a demotion, survived-a-decade-of-reprints is the only reliable ADD signal.
• SUBTRACT (aggressive): recency + trend + no-durability; guru repackaging; hagiography; hacks ephemera; debunked case studies held as gospel.
• KEEP: the tiny Lindy core; one debunked bestseller as a specimen.
• SERENDIPITY: cross a durable classic into an adjacent unread domain (Deming → statistics) — pull the reader out of the airport bookstore.

Sources: Lindy effect · Good to Great, follow-up · The Halo Effect

Framing. Computing is the domain where the antilibrary logic is most violent, because the shelf splits into two populations that age at opposite rates. There is no smooth curve — there is a bimodal distribution. One mode rots in ~18–24 months; the other is among the most Lindy literature in any field. Any subtract policy that treats "a computing book" as a single object is wrong. Khoj must read the layer, not the domain.

• The tool/framework/version layer — landfill-fast. Framework books are "obsolete the moment they are printed"; a React/Angular title's code won't compile by launch. The knowledge didn't just age — it moved: the current, correct version lives in the official web docs (MDN, Stack Overflow). The purest library-weeding E = Elsewhere case — the book is a stale snapshot of a living web resource.
• The foundational/conceptual layer — astonishingly eternal. Knuth's TAOCP, Abelson & Sussman's SICP, Kernighan & Ritchie's C, Brooks's Mythical Man-Month (1975, still quoted), Hunt & Thomas's Pragmatic Programmer. They survive because they teach thinking — abstraction, algorithms, the sociology of software — not syntax. Their expected remaining life grows with age.

(1) the timeless foundation — the "changed how I think" book; (2) the genuine frontier / paradigm shift (right now AI/ML-native thinking, LLM systems), where being early has option value; (3) sparingly, a current tool manual only when it's a deep conceptual treatment that outlives the version. A gap is a missing rung on the abstraction ladder — ten framework cookbooks but no algorithms text, no OS-internals.

Aggressively cull the version/tool/manual layer: any title bound to a specific framework version, deprecated language edition, or superseded API — especially when the authoritative version is free online. "Available elsewhere" is a first-class subtract signal here in a way it is not for literature or history. Simultaneously fence the foundational canon as un-cullable regardless of date — 1975 is a feature. The trap: a naive recency prior would subtract Brooks (old) and keep the React book (new). Exactly backwards.

Two objects, two relationships. The O'Reilly animal manual is a consumable reference — obsolete after the project, guiltlessly discarded; developers increasingly skip the print buy and live in web docs. The classic is a possession — reread, quoted, gifted, kept for a career. The tell: does the owner reach for it, or Google? If Google always wins, it's a subtract.

• Domain-keyed bimodal subtract prior: classify the layer first. Tool/version/manual → high subtract weight; foundational/conceptual → subtract weight near zero (protected canon).
• "Available elsewhere" as an active subtract signal — unusually strong here (docs migrated to the web).
• Recency inversion guard: never let newness raise keep-value in the tool layer; age-as-survival applies only to the canon.
• ADD bias: foundations you lack + the live frontier. Serendipity: the old classic the collector never bought.

Sources: Programmers Don't Read Books · Timeless programming books

Framing. This is the one domain where the book is not merely a container for knowledge — the book is partly the artwork itself. When the content is the image, the physical object (paper, ink fidelity, plate size, binding) is load-bearing. This inverts the utility logic of text-heavy fields: you cannot cull an art book the way you cull a superseded textbook. Simultaneously, this is the one domain where a newer edition can genuinely be better — not because knowledge advanced, but because the reproduction did. Both facts pull against ordinary subtract logic, in opposite directions.

The interpretation clock is glacial (art-historical readings revise across generations; old readings retain value as period documents). The reproduction clock ticks upward: photomechanical → offset → high-fidelity digital, each raising plate quality. So a 1960s monograph's argument may still stand while its muddy plates are objectively obsolete. At the top sits the catalogue raisonné — the complete scholarly inventory of an artist's oeuvre, "the supreme arbiter of the genuine and fake," decades to compile, near-permanent; a later CR supplements rather than discards. Below it: the definitive monograph, the exhibition catalogue (often the only record of a show's scholarship), the theory/criticism text, the coffee-table book (permanence by beauty).

(1) the catalogue raisonné of an artist you study — the anchor; (2) the definitive monograph; (3) the exhibition catalogue for a landmark show (scholarship published nowhere else); (4) the theory text that reframes seeing (Berger, Panofsky, Gombrich, Fried). A gap = an unseen artist, movement, medium, or non-Western tradition — the blind spot the collection never had the plate for. Serendipity here is strongly visual: an ADD justified by a single reproduction never encountered.

• Near-permanent keep: the image-canon (great monographs, definitive plates) and every catalogue raisonné.
• Quality-based supersession (unique to this domain): a poor-reproduction old edition can be subtracted when a better-printed edition of the same content exists. The rare case where Khoj recommends replacing a book with a newer edition on the merits — the image improved, not the text. Flag it replace, not discard.
• Genuine churn: design/decor trend books, "styling" fads, tool manuals, market-price guides — date fast, no object-value, subtract freely.

The anti-cull guardrail: never subtract on pure utility/duplication grounds — the art book is partly furniture, partly totem. A book kept for the object is a valid keep.

The plate over the text — artists and designers look, rarely read cover-to-cover. The book is a visual reference and mood source; the studio shelf is browsed by spine and by flipping, not searched by index. The catalogue is consulted like a dictionary. The book-as-object earns its space by presence, not throughput.

• A domain flag arts: raise the keep-threshold sharply; disable utility/duplication-based subtracts.
• Protected class: catalogue raisonné + definitive monographs = near-permanent, exempt from subtract.
• A quality-supersession subtract type unique to this domain — surface "better-printed edition exists" as a replace suggestion, never an outright cull. The one place Khoj says "buy the new edition of a book you own."
• A narrow churn list (trend/decor/manuals) that is freely subtractable; serendipity weighted toward the visual gap.

Sources: Catalogue raisonné · A brief history of the catalogue raisonné · Edition (printmaking)