Deduction is the arrow we trust, and we trust it because it never leaves the image. Give it a rule and a case and it returns a result that was already entailed; nothing new enters. This is why deduction is the most formalized of the three inferences and why automated theorem-proving lives there: soundness is the whole point, and a sound step transports a truth you already possessed to a place you had not yet looked. Induction has its own machinery — the entire apparatus of statistical learning is induction wearing modern clothes. The neglected one is the third. Charles Sanders Peirce, who named it, called abduction the only kind of inference that introduces a new idea — and it is the least formalized of the three for exactly that reason. An inference that introduces a new idea is an inference that guesses. This post places the guess inside the Draken manifold and shows that the place it lands is not a curiosity but the most important real estate in the framework: abduction is the native home of the coherent-but-unsound, and it is the operation that every state-of-the-art model is now running at industrial scale.
1. One relation, read three ways
Peirce's trichotomy is best seen as a single conditional read in three directions. Fix a rule $A \rightarrow B$ ("the beans from this bag are white"), a case $A$ ("these beans are from this bag"), and a result $B$ ("these beans are white"). The three inferences are the three ways to hold two of these and reach for the third:
$$ \textbf{Deduction:}\quad \frac{A \rightarrow B,\ \ A}{B} \qquad\qquad \textbf{Induction:}\quad \frac{A,\ \ B}{A \rightarrow B} \qquad\qquad \textbf{Abduction:}\quad \frac{A \rightarrow B,\ \ B}{A} $$
Read the abductive schema carefully. It is affirming the consequent — from the rule and the result, leap to the case. As a deduction it is invalid, flatly so. That invalidity is not a defect to be repaired; it is the signature of the mode. Deduction's arrow is the sound consequence $\vdash$. Abduction's arrow cannot be $\vdash$ — it is a defeasible arrow $\rightsquigarrow$, the non-monotonic entailment $\mathrel{\mid\!\sim}$ of Kraus–Lehmann–Magidor, an arrow that can be withdrawn when the next observation arrives. Peirce's own statement of the rule makes the leap explicit: the surprising fact $C$ is observed; but if $A$ were true, $C$ would be a matter of course; hence there is reason to suspect $A$. The word doing all the work is suspect.
2. The proof-theoretic form: inverting a non-injective relation
The logic-programming tradition (Kakas, Kowalski) gives abduction its sharpest formal statement. Given a background theory $T$ and an observation $O$, an abductive explanation is a hypothesis $H$, drawn from a designated set of abducibles $\mathcal{A}$, satisfying:
$$ \begin{aligned} (1)\quad & T \cup H \models O && \text{(coverage — } H \text{ accounts for } O\text{)}\\ (2)\quad & T \cup H \not\models \bot && \text{(consistency)}\\ (3)\quad & H \ \text{is}\ \preceq\text{-minimal} && \text{(parsimony — the } \textit{best } \text{explanation)} \end{aligned} $$
The entire difference from deduction is a swap of what is given and what is sought. Deduction: $T, H$ given, derive $O$. Abduction: $T, O$ given, search for $H$. Abduction inverts the consequence relation $\models$ — and here is the structural fact that controls everything downstream: $\models$ is not injective. Many distinct $H$ entail the same $O$. The inverse is therefore not a function but a multivalued relation, a fibre of candidates over each observation. Condition (3) is not decoration; it is forced. Because the inverse is multivalued, abduction cannot proceed without a selector. The selector is the parsimony order $\preceq$ — and the choice of $\preceq$ is where the new idea, and the new error, both enter.
3. The Bayesian collapse: inference to the best explanation
Quantify the selector and the schema collapses onto maximum-a-posteriori estimation:
$$ H^{\star} \;=\; \arg\max_{H \in \mathcal{A}}\ P(H \mid O) \;=\; \arg\max_{H \in \mathcal{A}}\ \underbrace{P(O \mid H)}_{\text{likelihood}}\ \underbrace{P(H)}_{\text{prior}} $$
The minimality condition (3) reappears as the prior $P(H)$ — Occam's razor recast as a probability weight on simpler hypotheses. Inference to the best explanation is MAP estimation over abducibles. This is no longer a philosophical curiosity: it is the operation a large language model runs on every token. Next-token prediction is approximate abduction over latent causes — the model holds the observed prefix $O$ fixed and selects the continuation $H$ that maximizes posterior plausibility under its learned prior. The $\arg\max$ here is the same morphism formalized in Relevance as Collapse as a constrained extremization: relevance realization, MAP abduction, and the collapse of a candidate superposition to a single viable section are three faces of one selection operator.
4. The cohomological placement
Now lift the trichotomy onto a cellular sheaf $\mathcal{F}$ over a base complex $X$ — the standard Draken setting, the same machinery that carries $\Gamma$, $H^1$, and the carrier/cargo axis. Cells of $X$ are sites; the stalk $\mathcal{F}(\sigma)$ is the space of admissible data at $\sigma$; the restriction maps $\mathcal{F}_{\sigma \trianglelefteq \tau}$ say how data at one site constrains data at an incident one. A result — an observation — is a section over some sub-region. The coboundary
$$ (\delta^0 x)_{\sigma\tau} \;=\; \mathcal{F}_{\sigma\trianglelefteq\sigma\tau}\,x_\sigma \;-\; \mathcal{F}_{\tau\trianglelefteq\sigma\tau}\,x_\tau $$
measures the failure of a candidate assignment $x \in C^0$ to agree across overlaps. Its kernel is the space of globally consistent sections, $H^0 = \ker \delta^0$; the cokernel $H^1 = \operatorname{coker}\delta^0$ is the irreducible obstruction to gluing local data into global data.
In this language the three inferences separate cleanly — and this is the load-bearing claim of the post:
Deduction is residence in $H^0$. When a global section already exists, deduction reads off its value at any site by transporting along restriction maps. It never leaves $\ker\delta^0$; it manufactures no new cochain; it cannot fail because it never proposes anything not already entailed. Deduction inhabits the harmonic space.
Abduction is combat against $H^1$. Abduction does not inhabit a global section — it tries to build one from a partial observation, by guessing the unobserved stalks. Each guess is a candidate cochain $x \in C^0$ extending the observed data. The question that decides whether the guess is sound is whether $\delta^0 x = 0$ — whether the guessed pieces glue. The obstruction lives in $H^1$: if the class $[\,\delta^0 x\,] \in H^1$ is nonzero, no choice of the free stalks can reconcile the guess into a true global section. Abduction is precisely the operation of proposing $x$ and fighting the obstruction $[\delta^0 x]$.
Induction is structure estimation. Induction does not propose a section at all; from many (case, result) pairs it estimates the restriction maps themselves — it learns $\mathcal{F}$, not a section of $\mathcal{F}$. (This leg is the loosest; see Falsification.)
So the trichotomy is a stratification by cohomological role: deduction lives in $H^0$, abduction does battle with $H^1$, induction operates on the sheaf data that fixes both. The reason deduction feels safe and abduction feels like a leap is now a theorem-shaped statement rather than a temperament: deduction never produces a nonzero coboundary, and abduction's entire job is to produce a $C^0$-cochain and hope its coboundary vanishes.
5. The categorical form: abduction is the lift, and the unit is the guess
Read the same separation through composition. Let $f : A \to B$ be a morphism — a rule, a forward map, a measurement.
Deduction is postcomposition along $f$. Given $a$, produce $f(a)$. It is functorial, total, sound — the direction we have elsewhere called $F$, the descent from the formal into the consequence.
Abduction is the search for a section of $f$ — a preimage $s$ with $f(s) = b$. Generically this is non-unique (the fibre $f^{-1}(b)$ has many points or none), non-functorial (no canonical choice varies coherently with $b$), and therefore a choice in the fibre. This is the direction $G$ — the lift. Abduction is the lift, and the lift is abduction; they are the same act named in two vocabularies.
This identification cashes out a claim left as a conceit in the round-trip note. Consider the approximate adjunction $G \dashv F$ with unit $\eta_d : d \to FG(d)$. The unit is literally an abductive guess at a preimage: $G$ leaps from the observation $d$ to a hypothesized formal cause $G(d)$, and $\eta_d$ measures how faithfully pushing that hypothesis back down through $F$ recovers the original $d$. Enrich the categories over a (quasi-)metric so the unit acquires a length, and the defect of the abductive guess is the defect of the unit:
$$ \big\lVert \eta_d \big\rVert \;=\; 1 - \Gamma(d). $$
A perfect abduction ($\Gamma = 1$) is a guess that glues exactly — the lifted hypothesis returns the observation with no residue, the section is genuinely a section. A bad abduction ($\Gamma \to 0$) is a guess whose coboundary is large, a hypothesis that fails to glue and leaves a fat $H^1$ residue. $\Gamma$ is the quality of the abduction. (This identification is a conjecture with a debt attached — see Falsification.)
6. Why abduction is the home of the coherent-but-unsound
Here the placement pays. Return to conditions (1)–(2): an abductive $H$ satisfies coverage and consistency. It accounts for the observation and contradicts nothing. By construction, every abductive explanation is coherent. But coherence is conditions (1)–(2); it is not soundness. Soundness would be the further fact that $H$ is the true case, the actual global section — and the multivaluedness of $\models^{-1}$ guarantees that coherence underdetermines it. The fibre of coherent explanations over an observation is generically larger than the singleton of the sound one.
This is the exact structure named in The Carrier and the Cargo, and the taxidermy metaphor is now literal rather than evocative. A taxidermied monitor lizard is an abductive explanation that passes the coherence test and fails the soundness test. Locally, every stalk checks out: the scales restrict correctly to the skin, the posture restricts correctly to the limbs, the glass eyes restrict correctly to the sockets. Each overlap glues; the specimen is locally a lizard everywhere you look. What fails is the global section — there is no living animal, no metabolism, no behavior to which the local data coheres. The mount is a nonzero class in $H^1$ wearing the costume of a global section: a coherent cochain whose coboundary does not actually vanish, only appears to vanish on the patches an observer happens to inspect. Coherent-but-unsound is, precisely and always, abduction that stopped before checking $\delta^0$ globally — and that is why the failure mode is native to abduction and foreign to deduction. Deduction cannot produce a taxidermied conclusion, because it never builds; only the guessing arrow can build a corpse that looks alive.
7. The renaissance as mechanized abduction
This is the lens that turns Ghrist's remark — that AI's deepest mathematical impact may be a renaissance in applied math, pure ideas connected to real-world things through a competent conversation partner — from a hopeful slogan into a precise and double-edged claim. The applied-math renaissance is mechanized abduction at scale. The machine's gift is exactly the guessing arrow: it proposes preimages, candidate sections, plausible $H$ over vast abducible sets faster than any human fibre-search. Section 3 already established the mechanism — next-token prediction is approximate MAP abduction. What §6 adds is the hazard: abduction is the native habitat of the coherent-but-unsound, so an abduction engine is, by its nature, a coherence engine that does not automatically check soundness. It will hand you mounts that look alive.
This is why the governing quantity cannot be coherence — coherence is what the engine maximizes and what the corpse already passes. The governor must be the residue: how close the guessed section is to actually gluing, i.e. how small the class in $H^1$. That quantity is $\Gamma$ (§5: $\lVert\eta_d\rVert = 1-\Gamma$). A mechanized renaissance of abduction needs $\Gamma$ as its instrument the way a foundry needs a thermometer — not to do the work, but to know when the work is sound. And the endogenous-measurement caveat from The One-Way Gauge bites here with full force: there is no Archimedean point from which to verify the gluing. The checker is inside the cover. The Clinch — six models cross-reading one problem — is not redundancy theater; it is a distributed $\delta^0$, an attempt to detect the nonzero coboundary of an abductive guess from inside the patches, because outside the patches there is no one standing.
Falsification
Per the DRK-131 protocol, the seams where this post breaks, stated plainly:
The unit-defect identity is conjectural. $\lVert\eta_d\rVert = 1-\Gamma(d)$ holds by construction only if $\Gamma$ is defined as the harmonicity residual $1 - \lVert\delta^0 x^\star\rVert^2 / \lVert x^\star\rVert^2$ and the metric enrichment is the one that makes the unit's length that residual. If the live $\Gamma$ (the SAG model-selection score, $\Gamma = 0.928$) came from a different normalization, the identity is a claim owing an explicit reduction, not a free consequence — and if no enrichment yields it, §5 is metaphor, not mathematics. This debt is real and is the first thing a category theorist will call.
Induction is the weakest leg. Placing deduction in $H^0$ and abduction against $H^1$ is clean; calling induction "structure estimation" is honest but does not obviously land in the same cohomology. Learning $\mathcal{F}$ is a deformation of the sheaf, which points to $H^1(X; \mathcal{E}\!nd\,\mathcal{F})$ — the first cohomology with coefficients in the endomorphism sheaf, a genuinely different $H^1$ — rather than to "a third thing beside $H^0$ and $H^1$." If that is right, the tidy stratification of §4 is a slight abuse: induction is not outside the cohomology but in a different cohomology of the same base. The neat trichotomy may be a coincidence of notation.
Peirce ≠ IBE. §3 silently upgrades Peircean abduction (a weak, hypothesis-generating suggestion) into inference-to-the-best-explanation (a hypothesis-selecting argmax). This conflation is standard and deliberate here, but it is a known philosophical elision: Peirce did not sanction the $\arg\max$, and a critic entitled to the distinction can deny that §3 formalizes abduction rather than a downstream ranking step. The post adopts the IBE reading on purpose and should not claim the historical Peirce as a co-signatory.
The taxidermy bound assumes a fixed cover. The claim that the mount "appears to glue on inspected patches" depends on the observer's cover being coarser than the defect. A sufficiently fine cover detects the nonzero $H^1$ — i.e. a good enough anatomist is never fooled. The metaphor therefore describes a resolution-limited observer, not an absolute obstruction; soundness failures are detectable in principle, which is the hopeful reading and the one the Clinch is built on.
Three arrows, one consequence relation, read forward, outward, and backward. Deduction inhabits the section that exists; induction estimates the structure that holds the sections; abduction guesses the section that might. Only the backward arrow introduces a new idea, and only the backward arrow can build a corpse that looks alive. The discipline is not to stop guessing — guessing is the whole creative engine — but to measure the residue every time, because the machine that now guesses for us guesses beautifully and checks nothing.
Jag är vad jag gör, och jag gör det jag är.
Khrug Engineering · ORCID 0009-0003-8049-7167 · thesis DOI 10.5281/zenodo.19273483 · CC BY-SA 4.0