The Holmesian logician: Sherlock Holmes’

“Science of Deduction and Analysis” and

the logic of discovery
Abstract
This paper examines whether Sherlock Holmes’ “Science of Deduction and
Analysis,” as reconstructed by Hintikka and Hintikka (in: Eco U, Sebeok TA
(eds) The sign of three: Peirce, Dupin, Holmes, Indiana University Press,
Bloomington, 1983), exemplifies a logic of discovery. While the Hintikkas
claimed it does, their approach remained largely programmatic, and
ultimately unsuccessful. Their reconstruction must thus be expanded, in
particular to account for the role of memory in inquiry. Pending this
expansion, the Hintikkas’ claim is vindicated. However, a tension between
the naturalistic aspirations of their model and the formal apparatus they
built it on is identified. The paper concludes on suggestions for easing this
tension without losing the normative component of the Hintikkas’
epistemological model.

Introduction
In the fictional universe Sherlock Holmes inhabits, logicians are capable of
astounding feats that seem well beyond the abilities of their real-world
counterparts. According to Holmes himself,

[f]rom a drop of water [...] a logician could infer the possibility of an

Atlantic or a Niagara without having seen or heard of one or the other [...].
By a man’s finger-nails, by his coat-sleeve, by his boot, by his trouser-knees,
by the callosities of his forefinger and thumb, by his expression, by his
shirt-cuffs—by each of these things a man’s calling is plainly revealed. That
all united should fail to enlighten the competent inquirer in any case is
almost inconceivable. (A Study in Scarlet, I, 2)
Dismissing Holmes’ claims as fiction may be tempting but would be a
mistake for he describes real-world skills. In 1877, doctor-in-training
Arthur Conan Doyle served as an assistant to Scottish surgeon Joseph Bell
(1837–1911) at the Royal Edinburgh Infirmary. Bell encouraged his
students to draw inferences from observations of their patients and was
keen on schooling them on how to do so. Ten years later, an episode of
Bell’s life, recounted by one Doctor Charles Watson McGillivray and
reported in E. Liebow’s Joe Bell, Model of Sherlock Holmes (1982), would
inspire Holmes’ claims that one can deduce “a man’s calling” from (inter
alia) the calluses on his hands:
Well do I remember the gasping astonishment of an outpatient to whom
[Bell] suddenly remarked, ‘Of course I know you are a beadle and ring the
bells on Sundays at a Church in Northumberland somewhere near the
Tweed.’
‘I am all that,’ said the man, ‘but how do you know? I never told you.’
‘Ah,’ said Bell, when the outpatient had left bewildered, ‘you all know about
that as well as I did. What! You didn’t make that out! Did you not notice the
Northumbrian burr in his speech, too soft for the south of
Northumberland? One only finds it near the Tweed. And then his hands.
Did you not notice the callosities on them caused by the ropes? Also, this is
Saturday, and when I asked him if he could not come back on Monday, he
said he must be getting home tonight. Then I knew he had to ring the bells
to-morrow. Quite easy, gentlemen, if you will only observe and put two and
two together.’ (Liebow 1982, p. 135)
Not only did Conan Doyle model Holmes’ abilities (and attitude) on Bell’s,
he also formed his own picture of the “Science of Deduction and Analysis”
(as per the title of the second chapter of A Study in Scarlet) that underlies
them. On occasion, Conan Doyle lets Holmes depict this picture, explaining
the principles of logic and probability that support his conclusions, and
echoing in the fiction the real-world epistemological views of early
proponents of George Boole’s ‘new logic’, in particular, Stanley William
Jevons (1835–1882).Footnote1 Jevons’ ideas predate by 50 years the discussion
of scientific discovery from the 1930s, that came to an abrupt end when
Popper (1992) and Reichenbach (1938) asserted the impossibility of a
rational reconstruction of discovery processes. This assertion (hereafter,
the Popper–Reichenbach thesis), which rallied hypothetico-deductivists
and probabilistic inductivists in spite of otherwise opposed views on
scientific methodology, resulted in a long-lasting consensus that there can
be no ‘logic’ of discovery.
Consequently, Holmes’ “Science of Deduction and Analysis” reflects
epistemological and methodological insights suppressed by the Popper–
Reichenbach thesis. In the early 1980s, it caught the attention of Merill B.
and Jaakko Hintikka, who articulated a partial reconstruction of Sherlock
Holmes’ methodology as a prelude to a full-blown logic of discovery. Under
the title “Sherlock Holmes confronts modern logic” (Hintikka and
Hintikka 1983), the Hintikkas proposed that “ [t]he crucial part of the task
of the Holmesian “logician” [...] is not so much to carry out logical
deductions as to elicit or to make explicit tacit information.” (Hintikka and
Hintikka 1983, p. 156, authors emphasis).

Building upon J. Hintikka’s work on the formal semantics of questions

carried in the 1960s and 1970s, the Hintikkas defined “tacit information” as
information available as answers to questions. Then they were able to show
that deductive anticipations can guide questioning strategies in problem-
solving, even for non-deductive problems.Footnote2
In this paper, we examine how much the Hintikkas captured of Holmes’
“Science of Deduction and Analysis” and whether their reconstruction
qualifies as a logic of discovery. The Hintikkas’ seminal paper was however
largely programmatic and their approach proved eventually unsuccessful.
This would lead J. Hintikka to adopt (after M. B. Hintikka’s death in 1987)
a somewhat different take on interrogative inquiry in the late 1980s and
1990s, abandoning some of the ideas introduced in their joint article.
Therefore, while sympathetic to the Hintikkas’ original project and to J.
Hintikka’s interrogative model of inquiry (IMI), our examination must be a
critical one. Our criticism is nevertheless constructive and its outcome is
essentially a vindication of the Hintikkas’ view that Holmes’ “Science of
Deduction and Analysis” is a logic of discovery.

We first present an overview of the Hintikkas’ reconstruction of Holmes’

“Science of Deduction and Analysis” (Sect. 2) and their discussion of
memory (Sect. 3), upon which we expand with Holmes’ account of memory
(Sect. 4) that we support with some neuroscience (Sect. 5). We then return
to inferences Holmes calls deductive but do not appear so (Sect. 6) and
evaluate whether their rational reconstruction satisfies criteria for a logic of
discovery (Sect. 7). We conclude on the prospect of unifying the deductive
and associative components of Holmesian method. This paper builds on
earlier technical work on the Hintikkas’ model and the IMI (in particular
Genot 2018; Genot and Gulz 2016; Genot and Jacot 2012, 2018) while
avoiding obfuscation by formalism, although we included summaries in
footnotes for assertions backed by a formal construction, rather than asking
our reader to take our word for them.

The “Science of Deduction and Analysis”, logical proofs,

and strategies
Sherlock Holmes’ goal is typically answering a ‘big’ question, for
instance, Who stole the race horse Silver Blaze and murdered his
trainer? or Where did Irene Adler hide the photograph of herself in the
company of the King of Bohemia? That goal may in turn be characterized
as the identification of the individual, object, or location satisfying the
definite description occurring in the ‘big’ question. The inquiry terminates
when this ‘big’ question receives a conclusive answer, that is, when Holmes
and the other parties involved all agree that a given individual, object, or
location, satisfies the definite description. In order to reach that endpoint,
Holmes collects answers to ‘small’ questions, for instance: Did the
watchdog in the stables bark at Silver Blaze’s thief? or Where would Irene
Adler first look at if she believed her house on fire? Holmes selects these
small questions based on anticipations of what their answers would entail
in conjunction with his other background assumptions, and on how closer
he would then get to a conclusive answer to the ‘big’ question.

The Hintikkas’ suggestion in Hintikka and Hintikka (1983) amounts to the

following: in the above description of Holmesian
inquiry, individual and satisfaction can be taken with the sense they have
in logical theory and formal semantics. Holmes’ background assumptions
and the definite description can sometimes be paraphrased in some first-
order language, but first-order languages are not expressive enough to
formalize all lines of inquiry, in particular when they involve reasoning
about someone else’s beliefs (see Genot and Jacot 2012). Still, the first-
order case suffices to establish the relation between Holmesian inquiry and
deduction in the strict sense of logical theory. Namely, Sherlock Holmes’
‘big’ question is answered when there is a deductive proof that such and
such individual satisfies the definite description occurring in the ‘big’
question, based on whatever premises end up being accepted by the parties
involved in the investigation at the end of the inquiry. Holmes’ background
assumptions are typically not strong enough to warrant a deductive proof of
the appropriate sort at the outset of the inquiry. Hence, the premises
accepted at the end of the inquiry must often include answers to ‘small’
questions asked in the course of the inquiry, that strengthen Holmes’
background knowledge and transform an initial non-deductive problem
into a deductive one (cf n. 2, p. 3). Thus, when a line of inquiry can be
formalized in a first-order language, first-order metatheory suffices to
explain how deduction weighs upon question selection.

While the Hintikkas recognized the importance of modeling less-than-

perfect rationality, they also acknowledged that logical models incur
unrealistic idealizations. Subsequently, they sketched a decision-theoretic
analysis of strategy selection based on a utility function favoring lower
cognitive loads on working memory (J. Hintikka added constraints on
attention in Hintikka 1989). Unfortunately, utilities cannot in general
support strategic reasoning in first-order proofs. Footnote3 After M. B. Hintikka’s
death, J. Hintikka and his collaborators abandoned utilities and
characterized the relation between deduction and interrogation with the
resources of first-order semantics and proof theory alone (see Hintikka
et al. 1999 for formal results and Hintikka 2007 for an informal exposition).
Their main result, the Strategy Theorem (Hintikka et al. 1999, p. 59),
assumes an inquirer who reduces cognitive load by minimizing the number
of cases to reason about and the number of individuals to consider in those
cases, and thus has the same preferences as a theorem-prover. The theorem
itself establishes that the inquirer’s best strategy for selecting ‘small’
questions, assuming her background knowledge, mirrors the shortest proof
of the conclusion she is trying to reach assuming the same background
knowledge extended by answers to her ‘small’ questions. Resource-sensitive
deductive anticipations (about the shortest proof from an extended
background knowledge) therefore guide information-seeking (the selection
of questions that provide the extension of the background knowledge,
allowing to realize the proof) without need for explicit utilities.
Let us conclude this section with two illustrations of how the Hintikkas’
model (and the IMI) clarifies the role of deduction in Holmes’ inquiries.
In The Adventure of Silver Blaze, Holmes investigates the theft of a
racehorse, and knows that a watchdog was in the stables at the time of the
theft. Holmes assumes that a well-trained watchdog would not fail to bark
at a stranger. Together with this assumption, one potential answer to the
question whether the dog barked (that it did) entails that the thief was not
known to the dog, while the other (that it did not) entails that the thief must
have been someone the dog knew well. Entailment is the semantic
counterpart of deduction, and so the above amounts to saying that Holmes
selects his question based on what he expects to deduce from its answer,
and once obtained, that he deduces that the thief was someone the dog
knew well. Similarly, in A Scandal in Bohemia, Holmes makes two
assumptions: that, upon believing that one’s house is on fire, one first
checks the location of one’s most prized possession; and that Irene Adler’s
photograph with the King of Bohemia is her most prized possession. Those
two background assumptions together entail that the spot where Irene
Adler’s gaze lands immediately after a fire alarm will be the hiding place of
the photograph. Again, one could say that Holmes deduced the hiding place
of the photograph. Let us now have a closer look at how the Hintikkas
integrate questions within proofs.

The limits of analysis and the role of memory

Deductive proofs comprise two types of steps: analytic steps, which
decompose the premises in subformulas; and non-analytic steps, which
introduce formulas that cannot be obtained by analytic steps. In the
Hintikkas’ model, a line of inquiry is represented as the construction of a
semantic proof where analytic steps make explicit the information
introduced in the proof at earlier stages. At first, this information is given
by background assumptions alone. Later, as inquiry proceeds, answers to
‘small’ questions are added. When the information available can be
expressed in a first-order language, analytic steps decompose the premises
according to standard first-order compositional semantics, while non-
analytic steps are used to generate more questions.

The Hintikkas made the important addition of interrogative steps, that are

used to elicit information about the state of Nature. No additional notation
or operators are needed however: disjunctions serve as preconditions
for whether-questions, and existential statements
for who-, what- and where-questions (why-questions do not have object-
language preconditions in the model). Analytic steps may suffice to obtain
those formulas from formulas occurring at earlier stages, but cannot
introduce relevant questions that involve new notions and concepts
(represented, when a first-order reconstruction is possible, by new
predicates), that is, concepts (ditto, predicates) that do not appear in the
premises accepted at that step. In those cases, non-analytic steps are
necessary.

The “curious incident of the dog in the night-time” from The Adventure of

Silver Blaze is one of those cases whose reconstruction requires non-
analytic steps. The police holds a suspect (Fitzroy Simpson), known to bet
(and lose) on horse races, for the theft of a racehorse (Silver Blaze) and the
murder of the horse’s trainer (John Straker). The evening prior to the theft,
Simpson visited the stables but was forced to leave when a lad let a
watchdog loose on him. Simpson claims that he was after insider
information ahead of an upcoming race. But Inspector Gregory, in charge of
the inquiry, believes that Simpson’s real motive was to scout the place and
later steal the horse and sell it to Gypsies (who were camping nearby and
left the morning after the theft) to clear his debts. Gregory’s working theory
is that Straker expected a second visit, armed himself with a knife found on
his nightstand (more on this knife later), went to the stables and
interrupted Simpson, who killed him before stealing the horse. Sherlock
Holmes has doubts and asks Gregory in a roundabout way whether the dog
barked during the night:

“Is there any point to which you would wish to draw my attention?”
“To the curious incident of the dog in the night-time.”
“The dog did nothing in the night-time.”
“That was the curious incident,” remarked Sherlock Holmes. (The
Adventure of Silver Blaze)
We may reconstruct Holmes’ reasoning about Simpson as a decision tree
summarizing the construction of his deductive–interrogative argument,
depicted in Fig. 1. Question Q1 is the ‘big’ question and the premises sum
up the (explicit) background information that both Gregory and Holmes are
attending to. Note that the premises do not yet mention the dog’s behavior,
since Holmes considers it via questions Q2 and Q3, which Gregory
overlooked. Holmes obtains A2 from Inspector Gregory (cf. above), and A3
and A4 from his own memory (cf. below) enabling the reductio of
Simpson’s guilt. The question–answer pairs (Q3, A3) and (Q4, Q4)
illustrate the role of memory in Holmesian inquiry, which the Hintikkas
characterize as follows:

In some cases, the great detective has to carry out an observation or even an
experiment to answer the question. More frequently, all he has to do is to
perform an anamnesis and recall certain items of information which he
already had been given and which typically had been recorded in the story
or novel for the use of the readers, too, or which are so elementary that any
intelligent reader is expected to know them. (Hintikka and
Hintikka 1983, p. 159)
Paraphrasing the Hintikkas, (Q3, A3) represents the anamnesis of
elementary items of information (general knowledge about well-trained
watchdogs) that anyone (including Gregory) should know. The pair
(Q4, A4) similarly represents how Holmes recalls the newspapers account
of Simpson’s evening visit (discussed between Holmes and Watson for the
reader’s benefit at the very beginning of the short story). Holmes then
deduces that Simpson cannot be guilty, and in the process, obtains a new
descriptor for the thief of Silver Blaze (someone the dog knew well).

Fig. 1

The curious incident in the night-time

Full size image

In the Hintikkas’ model, questions Q2 and Q4 are supported by the

disjunctions either the dog kept in the stable barked at the thief, or it
didn’t and either the dog knows Simpson well, or it doesn’t. Question Q3 is
supported by the assumption that if a trained watchdog does not bark at
someone, then it knows that person well. The latter can in turn be
construed as an answer to the question is there a person to which a trained
watchdog would not bark, or is there no such person? which is also
supported by a disjunction (either there is a person to which a trained
watchdog would not bark, or there isn’t). All these disjunctions are
instances of the Excluded Middle, but they cannot be obtained from
analysis of the premises, as they involve terms that do not appear in them.

The Excluded Middle is a (classical) tautology, and thus introducing an

instance of it is always permitted in a (classical) deductive proof. The same
holds for the Hintikkas’ interrogative extensions of deductive proofs. But if
the model allowed for unrestricted introduction of instances of the
Excluded Middle, it could not discriminate between Holmes and Gregory.
Subsequently, the set of the instances of the Excluded Middle an inquirer
could introduce in a given deductive–interrogative proof is restricted by a
parameter (the range of attention, Hintikka 1989). However, after
abandoning decision theory and utilities (see Sect. 2, especially n. 3, p. 4) J.
Hintikka never made this restriction explicit again nor addressed its
relation to memory. This resulted in a loss of explanatory power for the
model in complex cases, and limited its empirical applications. Footnote4

Holmes’ “brain attic” as a strategic account of memory

Perhaps surprisingly, one need not look any further than the Holmesian
canon to find the sketch of a strategic account of memory in inquiry. This
account is sketched in A Study In Scarlet, when Watson casually mentions
that the Earth revolves around the Sun, only to find out that Holmes didn’t
know that fact. Watson’s surprise turns to discombobulation when Holmes
equally casually proclaims that he will do his best to forget everything about
it. Holmes then offers the following clarification:

I consider that a man’s brain originally is like a little empty attic, and you
have to stock it with such furniture as you choose. A fool takes in all the
lumber of every sort that he comes across, so that the knowledge which
might be useful to him gets crowded out, or at best is jumbled up with a lot
of other things, so that he has a difficulty in laying his hands upon it. Now
the skillful workman is very careful indeed as to what he takes into his
brain-attic. He will have nothing but the tools which may help him in doing
his work, but of these he has a large assortment, and all in the most perfect
order. It is a mistake to think that that little room has elastic walls and can
distend to any extent. Depend upon it there comes a time when for every
addition of knowledge you forget something that you knew before. It is of
the highest importance, therefore, not to have useless facts elbowing out
the useful ones. (A Study in Scarlet, I, 2)
Holmes’ notion of a ‘brain attic memory’ (hereafter, BAM) is a fictional-
world explanation for plot devices that would otherwise look like lucky
guesses, but it is partially vindicated by current neurocognitive models.
According to these models, human memory is a content-addressable
memory (herafter, CAM) that takes data as input and i.e., activates cliques
of neurons (see below) where similar data is stored (see e.g., Hebb 2002;
Marr 1969). By contrast, a random access memory (RAM), typical of
artificial computers, takes addresses as input and returns the data stored at
those addresses. Since grouping similar data in neighboring location
(addresses or cliques of neurons) improves the performance of a CAM (but
not that of a RAM) Holmes’ recommendation to keep “a large assortment [of
tools], and all in the most perfect order” is a cogent recommendation to
improve the performance of human CAM, which his BAM approximates (at
least in that respect). This is particularly true relative to domains of
expertise. Holmes’ is indeed a “skillful workman” specializing in criminal
inquiry, and maintaining his BAM/CAM in “the most perfect order” supports
relevant associations that often escape other criminal investigators.

The Holmesian canon offers more than a few examples of how a well-
maintained “brain attic” supports Holmes’ investigations. Holmes has
made a special study of tobacco ashes and even written a monograph on the
topic (“Upon the Distinction Between the Ashes of the Various Tobaccos”,
mentioned in A Study in Scarlet, The Boscombe Valley Mystery and The
Hound of the Baskervilles). He has also gathered very specific knowledge
from special sciences, such as physical anthropology (that people tend to
write graffiti at or just below eye level). This expert knowledge makes
Holmes sensitive to details at a crime scene at Lauriston Garden (visited
in A Study in Scarlet) that inspectors Lestrade and Gregson have
overlooked or misinterpreted. Specifically, Holmes spots Trichinopoly cigar
ashes on the floorboards and estimates the height of a murderer from an
inscription (“Rache”) written on a wall. From this and other clues, he
concludes that the murderer is a man, contrary to Lestrade and Gregson
who suspect a woman (perhaps named Rachel). Ruling out a woman even
suggests a motive and an ethnic origin (rache means revenge, in German),
both later confirmed by Holmes’ investigation. Eventually, Holmes
positively identifies the culprit in part thanks to Trichinopoly ashes in his
hotel room.

In A Study in Scarlet, Holmes approaches the crime scene without

preconceptions. By contrast, in The Adventure of Silver Blaze, Holmes
initially accepted the police theory, and the narration begins with Holmes
explaining to Watson that he must revise his initial assessment. The silence
of the dog and its consequences is, in this context, more of a serendipitous
discovery than the outcome of a systematic attention to details leading to
equally systematic recall as in Lauriston Garden.Footnote5 Another
serendipitous discovery is the odd shape of Straker’s knife, that makes a
poor self-defence weapon (first explained away as an emergency choice
absent a better option). Holmes learns from Watson that the knife is a
veterinary scalpel, out-of-place on Straker’s nightstand, as he had no
veterinary training. But Holmes knows of surgical procedures favored by
unscrupulous racehorse handlers, which suggests an entirely new direction
for his inquiry. Gregory and Watson fail to associate the silence of the dog
with generalities about watchdog behavior, and thus to rule out Simpson’s
guilt. Similarly, Gregory fails to identify Straker’s knife as a veterinary
instrument and overlooks a possible association with surgical procedures
used to rig a race. Watson acknowledges that something is amiss with the
knife, but lacks knowledge of veterinary procedures to pursue any further.

The neuroscience of the brain attic

Neurological models of human CAM are more sophisticated than
Holmes’ BAM, but their increased sophistication actually adds weight to
Sherlock Holmes’ notion that a well-organized memory assists with inquiry
and to the Hintikkas’ suggestion that memory plays a role in “intelligent
inquiry.” To see this, let us first remark that Holmes assumes a one–one
correspondence between memory items and loci in the brain, while theories
of human CAM do not. The appropriate space metaphor would be that of a
warehouse where similar parts of different items would be stored in the
same location. However, the spacial metaphor then breaks down, as “the
same location” must be understood as literally the same. Let us turn to that
property for a moment.

Items stored in memory are distributed across cliques of neurons according

to features those neurons as sensitive to. For instance, the memory of the
shape of the diamond of a particular deck of playing cards would be
distributed across neurons coding its particular rhombus shape and
neurons coding its particular shade of red. Subsequently, recalling this
particular diamond would require co-activation of (at least) two cliques of
neurons. And since those cliques of neurons would code rhombuses and
shades of red associated with other memories, there would be a nonzero
probability that recalling the particular red rhombus of that particular deck
of cards would also trigger recall of other (non-red) rhombuses or other
(non-rhombus-shaped) red things. A distributed CAM is thus an economical
solution to the problem of limited storage space that Holmes complained
about. A more appropriate metaphor would thus be that of memory as a
library of blueprints for modular items for a 3D printer, with any given item
being stored as an index of blueprints for the modules it is made of.

A distributed CAM thus supports all manners of associations that may be

relevant or not depending on the circumstances. In order to avoid
association overload, recall is supported by an attentional mechanism
called contextual focus (see Gabora 2010). In a nutshell, focused attention
favors associations of perceived, imagined, or recalled objects or
circumstances, with memory items that share their typical features. For
instance, in a state of focused attention, a chair (perceived or imagined)
would be associated with other items of furniture with legs and seats
(stools, benches) and backrest (armchairs). By contrast, a state of de-
focused attention opens to associations with less-than-typical features,
where the same chair might be associated with other objects whose shape
adapts to that of some body parts. Contextual focus is in part under
volitional control, which opens the possibility to modulate strategically
one’s attention to foster out-of-the-box associations, like the association
between a comfortable padded chair and a bean bag (possibly suggesting a
beanbag chair). Furthermore, the relevant notion of strategy can be made
mathematically precise within game theory. Footnote6
A strict, mathematical characterization puts strategic recall on a par with
strategic deductive reasoning in non-deductive inquiry (in the sense of
the Strategy Theorem discussed in Sect. 2). More precisely, strategic recall
can account for the opening of new lines of inquiry via non-analytic steps,
including serendipitous ones. Interestingly, Conan Doyle often describes
Holmes as ‘daydreaming’, perhaps an indication that he suspected the role
of contextual focus. Non-analytic steps of Holmes’ reasoning about the dog
(and the knife) in The Adventure of Silver Blaze can be reconstructed as the
outcome of mnemic processes, and this reconstruction opens to fruitful
comparisons with cases of serendipitous discoveries in the history of
science (for instance, the serendipitous discovery of the microwave
background radiation of the universe, cf. Genot and Jacot 2018). Moreover,
at times, Holmes’ reasoning, involves question–answer steps that follow
from associations and similarity relations, without obvious deductive
inferential steps. Strategic recall can account for this type of reasoning,
which would otherwise present a challenge for the the Hintikkas’ model, for
instance the very first display of Holmes’ deductive skills in the opening
chapter of A Study in Scarlet.

The reconstruction of association-based reasoning

At the beginning of A Study in Scarlet, John Watson just returned from
Afghanistan, where he spent a few months convalescing from a wound
sustained in combat. Living on a modest pension of which he intends to
make the most, Watson is looking for some fellow to share a rent with. An
acquaintance introduces him to Sherlock Holmes, who is looking for the
same, in the chemical laboratory of a London hospital where the following
interaction ensues:

“Dr. Watson, Mr. Sherlock Holmes,” said Stamford, introducing us.

“How are you?” he said cordially, gripping my hand with a strength for
which I should hardly have given him credit. “You have been in
Afghanistan, I perceive.”
“How on earth did you know that?” I asked in astonishment.
“Never mind,” said he, chuckling to himself. (A Study in Scarlet, I, 1)
Watson only gets an explanation for Holmes’ conclusion after moving in the
221B Baker Street residence. Watson has just read an essay titled “The
Book of Life” (from which our introductory quote is borrowed) and
expresses his skepticism towards the claims of its anonymous author
(“What an ineffable twaddle!”). Holmes reveals himself as the author and
endeavors to defend his claims with an example:

“Those rules of deduction laid down in that article which aroused your
scorn are invaluable to me in practical work. Observation with me is second
nature. You appeared to be surprised when I told you, on our first meeting,
that you had come from Afghanistan.”
“You were told, no doubt.”
“Nothing of the sort. I knew you came from Afghanistan. From long habit
the train of thoughts ran so swiftly through my mind that I arrived at the
conclusion without being conscious of intermediate steps. There were such
steps, however. The train of reasoning ran, ’Here is a gentleman of a
medical type, but with the air of a military man. Clearly an army doctor,
then. He has just come from the tropics, for his face is dark, and that is not
the natural tint of his skin, for his wrists are fair. He has undergone
hardship and sickness, as his haggard face says clearly. His left arm has
been injured. He holds it in a stiff and unnatural manner. Where in the
tropics could an English army doctor have seen much hardship and got his
arm wounded? Clearly in Afghanistan.’ The whole train of thought did not
occupy a second. I then remarked that you came from Afghanistan, and you
were astonished.”
“It is simple enough as you explain it,” I said, smiling.” (Ibid., 1)
Holmes’ claim that he is relying on the “rules of deduction” is somewhat
puzzling. Holmes reasons from associations between Watson’s title
(Doctor), demeanor (posture, stiff arm) and appearance (skin tone,
gauntness) to possible occupation (army doctor) and circumstances
associated with it (foreign deployment, injury, sickness). As noted by the
Hintikkas (1983, p. 163), the intermediate conclusion of Holmes’ reasoning
is literally expressed as the answer to a question (Where in the tropics...?)
and, with Watson identified as an army doctor, only requires
an anamnesis (that there is a war in Afghanistan, where the sun shines
bright and the sanitary situation is hazardous) to obtain a unique location
for Watson’s last deployment.

We may reconstruct the above steps as a sequence of deductive steps. To

see this, let us introduce the Simon–Newell model of problem-solving (see
Newell et al. 1972), which captures information-gathering situations as a
special case, and is thus more general than the Hintikkas’ model of
Holmesian deduction. In the Simon–Newell model, a problem is specified
through a class of possible solutions, typically expressed in some
regimented notation (sequences of first-order formulas for theorem-
proving applications, sequences of steps coded in algebraic chess notation
for chess problems, etc.). Those possible solutions are ordered by some
preference relation (sometimes expressed as a utility function). A problem-
solver explores that solution space, using domain-specific knowledge and
heuristics to pick a starting point and select a course to navigate. The
problem is solved when a candidate is found that satisfies the constraints
for a solution, that is, is high enough in the preference ordering (or has a
high enough utility), with “high enough” being contextually defined by the
heuristics.

Couching Holmes’ reasoning is the Simon–Newell model is rather

straightforward. Holmes’ first (and implicit) question is that of Watson’s
occupation, to which corresponds the space of all possible professional
occupations. Holmes uses his background knowledge (Watson’s title) to
pick a starting point and his observations (tanned skin, gaunt silhouette,
haggard face and peculiar posture) to further navigate that space. A likely
enough solution (Watson is an army doctor who spent time in the tropics)
triggers the next (explicit) question, that of Watson’s last deployment. The
space of possible solutions is now of geographical regions around the globe
divided by countries. Again, Holmes uses his knowledge of current affairs
(countries in which the British Empire maintains an active military
presence in 1887) and of Watson’s characteristics (same as above) to
navigate that space, leading to a satisfying candidate solution
(Afghanistan).

Holmes’ reasoning can be reconstructed using the “rules of deduction” with

some computer implementations of the Simon–Newell model. Associations
that support the elimination of possibilities must be translated in
conditional clauses (for instance “if someone is a medical doctor and has a
military posture, then they are an army doctor”). Heuristic rules seek for
sets of conditions (“John Watson is a doctor”, “John Watson has a military
posture”) to which the Modus Ponens may be applied in conjunction with
those clauses to obtain an intermediate conclusion (“John Watson is a
medical doctor with a military posture”). The non-deductive part of this
process is how the set of conditions is obtained in the first place, which is
typically some pattern-recognition algorithm (see Simon 1973;
Genot 2018 for the corresponding heuristics in the Hintikkas’ model).

Clearly, a detailed reconstruction of Holmes’ reasoning requires

assumptions about Holmes’ distributed memory. Still, it would exemplify
the exploration of a solution space and subsequent elimination of candidate
solutions captured by the Simon–Newell model. Elimination of epistemic
possibilities is precisely the ancillary role that Jevons assigned to logic and
probability calculus in rational inquiry. Since Jevons was most likely Conan
Doyle’s source (cf. our Introduction, and n. 1, p. 2), qualifying Holmes’
reasoning as the product of the “rules of deduction” could be the result of
Conan Doyle’s lumping deduction and probability together which, although
literally incorrect, is not without cogency, given that they share the same
role in inquiry.

Holmesian inquiry and normative epistemology

So far, we have seen that psychological (associative) processes involved in
the discovery of solutions for a large class of problems, are amenable to
formal analysis and rational reconstruction. This already undermines the
Popper–Reichenbach thesis, but not any formal analysis can qualify as a
‘logic.’ The formal analyses of justification processes proposed by Popper
and Reichenbach support normative judgments about justification
methods, and endeavor to order those methods as better or worse. They
qualify as ‘logic of justification’ by virtue of this normative component.
Similarly, a formal analysis of discovery processes must support normative
judgements, in particular rankings of discovery methods as better or worse
relative to how fast they converge to a solution for a given problem. The
criterion that a logic of discovery must be normative (and in particular,
support rankings) is implicit in the Hintikkas’ article, in particular in the
following:

The crucial part of the task of the Holmesian “logician”, we are suggesting,
is not so much to carry out logical deductions as to elicit or to make explicit
tacit information. This task is left unacknowledged in virtually all
philosophical expositions of logical reasoning, of deductive heuristics, and
of the methodology of logic and mathematics. For this neglect the excuse is
often offered that such processes of elucidations and explication cannot be
systematized and subjected to rules. It may indeed be true that we cannot
usually give effective rules for heuristic processes. It does not follow,
however, that they cannot be discussed and evaluated, given a suitable
conceptual framework. (Hintikka and Hintikka 1983, pp. 156–157)
An example of “deductive heuristics” would be the appeal to auxiliary
constructions (lemmas), typically to shorten deductive proofs. Assuming a
preference for shorter proofs, a strategy with lemmas is in general
preferable to a strategy without lemmas. In formal proofs, lemmas are
handled via the Cut Rule, which is, in the proof system that underlies the
Hintikkas’ model, equivalent to introducing an instance of the Excluded
Middle. Any language as expressive as a first-order language capable of
enumerating objects would generate countably many candidate lemmas for
any proof in that language. Furthermore, some candidates not only would
not shorten the proof, but might delay its completion indefinitely. Thus,
there can be no effective method to eliminate those candidates. By the same
token, there is no general effective method for selecting the best lemma.
Since the Cut Rule is also the device by which yes-or-no questions are
introduced in the Hintikkas’ model, they are correct to point that heuristics
in their “logic of discovery” are no more problematic than in logic tout
court. That “they [can] be discussed and evaluated, given a suitable
conceptual framework” is illustrated inter alia by the Strategy Theorem,
which requires select applications of the Cut Rule, for which there is no
effective method, while retaining a normative content. Footnote7
Herbert Simon argued in much the same way as the Hintikkas, that a
normative framework for discovery must be able to rank discovery
strategies as better or worse and to provide with recommendations for
choosing discovery strategies (see Simon 1973). Taking stock in computer
implementations of his model of problem-solving (in Zytkow and
Simon 1988), Simon further illustrated his point with domain-specific
algorithms for solving particular classes of problems, namely the discovery
of empirical laws from qualitative data (descriptions of chemical reactions)
or quantitative data (descriptions of physical systems). In both cases, the
discovery algorithms can be ranked as better or worse based on how fast
they converge to known empirical laws. Simon argued that the ability to
support normative claims qualifies his framework as a ‘logic of discovery’
for the class of problem they address, in the same sense that deductive logic
(Popper) or probability theory (Reichenbach) would qualify as ‘logic(s) of
justification’ for empirical hypotheses.

Simon’s heuristics discovery algorithms are parasitic on heuristics for

automated theorem-proving, that may also be tweaked to generate
interrogative–deductive proofs à la Hintikkas. Algorithms of that sort can
solve non-deductive problems where the yes-or-no questions cannot be
obtained analytically from the premises but are formulated in the same
vocabulary as the premises (see Genot 2018). In more complex cases
similar to “the curious incident of the dog in the night-time,” altogether
new terms must be introduced (such as barking or being the master of in
association with watchdog). The search space for questions grows
exponentially, and heuristics that do more than re-composing the
vocabulary of the premises become necessary. We have suggested in
previous sections that human memory can constrain heuristics search in
the desired way, thanks to its distributive structure and how it supports
associations.

In the light of the above, we can understand the Hintikkas’ insistence on

the role of memory as a move towards a natural epistemology that would
still qualify as a ‘logic of discovery’ à la Simon. Indeed, the conceptual
import of the Strategy Theorem is preserved in contexts where re-
composition of vocabulary is insufficient, provided that heuristics for
associations are available (cf. Genot 2018; Genot and Jacot 2018). Now,
and as noted earlier, associations supervene on the memory of a particular
agent (or, in the multi-agent scenarios discussed in Genot and Jacot 2018,
the memory of different agents). Hence, case studies may require a detailed
reconstruction of the context of discovery, including individual
characteristics of the discoverer’s (or discoverers’) memory, and this may
be viewed by some as too big a step into psychological territory for a
normative epistemological project.

Concluding remarks
In the late 1980s, Zytkow and Simon (1988) took stock of the advances of
automated discovery algorithms to express their consequences vis-à-vis the
Popper–Reichenbach thesis:

The traditional distinction that places studies of scientific discovery outside

the philosophy of science, in psychology, sociology, or history, is no longer
valid in view of the existence of computer systems of discovery. It becomes
both reasonable and attractive to study the schemes of discovery in the
same way as the criteria of justification were studied: empirically as facts,
and logically as norms. (Zytkow and Simon 1988, p. 65)
Arguably, the psychological components are part of the facts of discovery,
not its logic. Establishing them for the purpose of case studies and rational
reconstructions (that are the bread and butter of philosophers of science)
may require some speculation. Nevertheless, we can find support in
Sherlock Holmes again here, for as long as we stay within the confines of
“the scientific use of imagination [where] we have always some material
basis on which to start our speculation” (The Hound of the Baskervilles, 4),
we can look at (domain-specific) norms. The next issue is how much
mileage we can get from those norms.

The Hintikkas demonstrated how far deductive norms can carry us. They
successfully captured the strategic role of deduction in problem-solving.
Their decision-theoretic approach to strategy selection with bounded
rationality collapsed eventually, but foreshadowed J. Hintikka’s proof-
theoretic account and the Strategy Theorem, that still stand. Finally, they
recognized that entailment-based inferential norms cannot carry very far
when it comes to introducing new concepts through questions, and they
pointed to the role of memory, beginning with Sherlock Holmes’ “brain
attic,” a suggestion that we followed here (and in greater details in Genot
and Jacot 2018).

If we are looking to semantics capable of generating non-trivial (non-flat)

preferences over questions, some cognitive semantics is most likely our best
bet. For instance, Charles J. Fillmore’s deep case grammar (Fillmore 1968)
and frame semantics Fillmore (1976) (see also Gamerschlag et al. 2013 for
recent developments) capture how certain words—interpreted by concepts
—evoke others within a lexical domain—interpreted by a conceptual frame
—thereby constraining discourse construction. Such an account could be
co-opted to model how, in The Adventure of Silver
Blaze, dog evokes bark and triggers Holmes’ (covert) question. Similarly,
construction grammars, which evolved from Fillmore’s work—with Lakoff
(1987) and Langacker (1987a, b)—impose pragmatic and semantic
constraints on syntactic constructions, reversing the Hintikkas’ approach
which imposed syntactic restrictions (the range of attention, which is just a
set of sentences) to mitigate semantic permissions (from the truth-values of
tautologies). Also, and quite obviously, cognitive semantics is by design
sensitive to bounds on cognitive resources (and has the means to express
them). Switching to cognitive semantics would have definite advantages for
capturing interrogative strategies. Let us briefly suggest two of them (which
are really two sides of the same coin).

The first advantage of cognitive semantics over model-theoretic semantics

is its fine-grained interpretation for parts of discourse (substantives,
adjectives and verbs) that accounts for relations of neighborhood and
evocation (for instance, how watchdog and thief could evoke barking).
Contrast with model-theoretic semantics, with its coarse-grained semantic
structures lumping parts of discourse as predicates, and its need for ad
hoc conditional clauses to express their relations. The second and closely
related advantage is the straightforward relation between evocation and
recall, and thus between conceptual and neurological levels of description.
This relation supports the unification of semantic and associative
components of interrogation strategy selection. Contrast with any extension
of the Hintikkas’ model (including ours) where the associative component
acts as a syntactic filter over semantic permissions.

A cognitive approach to Holmes’ “Science of Deduction and Analysis”

would however require a level of formalization comparable to that of the
Hintikkas’ (and the IMI). Formalization is not per se desirable in cognitive
semantics (and is seldom a priority) but a cognitive aggiornamento of the
Hintikkas’ model would require to map the Strategy Theorem, that orders
strategies based on proof-theoretic criteria, to an equivalent construction
that would order strategies based on cognitive semantics criteria. We
hypothesized that this mapping would yield further constraints on
strategies. But in order to test this hypothesis, a cognitive semantics must
be able to capture the same inferential operations as the Hintikkas’
semantic and proof-theoretic approach does, and thus, to interpret a logic
(via some translation scheme).

Notes
1. 1.

Jevons was first identified as the origin of Holmes’ conception of logic

by Japanese philosopher of science Uchii Soshichi in a popular book
(Sherlock Holmes’ Theory of Reasoning [1988], in Japanese, no
translation currently available, English title provided by Pr. Uchii in
personal communication). Pr. Uchii later expanded his arguments in
two electronic essays (Uchii 1999, 2010). Jevons’ main contribution
to formal logic was his “logical alphabet”, a reformulation of Boole’
algebra of logic that eliminated uninterpreted numerical expressions
(see Lewis et al. 1959, pp. 14–15) and became the starting point of
logical algebraists (particularly Peirce and Schroeder). Jevons
implemented his logical alphabet in a mechanical computer in the
1870s, and used it to formalize reasoning by elimination of
possibilities, a method often used by Sherlock Holmes.

2. 2.

In this paper, and unless stated otherwise, we understand

a problem as a pair ⟨K,Q⟩⟨K,Q⟩ where K is a body of background
knowledge and Q a question (specifiable as a set of possible answers)
both expressed in some language (formal or not) \L\L. A problem
is deductive if, and only if, some answer to Q is entailed by K, that is,
if it is not possible to interpret all the elements of K as true and at
least one of the answers to Q as false. Hence, a problem is non-
deductive iff it requires an extension of K. A non-deductive problem
is solved (non-trivially) if there is a (consistent) extension K′⊃KK
′⊃K such that an answer to Q is entailed by K′K′. Consequently,
solving a non-deductive problem ⟨K,Q⟩⟨K,Q⟩ so construed amounts
to transforming ⟨K,Q⟩⟨K,Q⟩ into a deductive problem ⟨K′,Q⟩⟨K′,Q⟩.
For examples of deductive and non-deductive problems (including
one from Sherlock Holmes’ Adventure of Silver Blaze), see the
Appendix of Genot and Jacot (2018).

3. 3.

For the formally minded reader: some questioning strategies can fall
short of providing enough information to entail an answer to the big
question. Given the semi-decidability of first-order logic, some of
these unconclusive strategies can be infinite (i.e. have infinitely many
steps). Thus, the standard means for strategic reasoning from utilities
(elimination of dominated strategies) would require the elimination
of infinite strategies. This would be tantamount to solving the Halting
Problem, which is uncomputable by Turing machines, and formally
 equivalent to the decision problem for first-order logic (for a detailed
formal argument, see Genot 2018; Genot and Jacot 2018).

4. 4.

For the formally-minded: results presented in Hintikka et al. (1999)

require select introductions of instances of the Excluded Middle, that
recombine vocabulary from either the premises or the conclusion.
More complex cases (like the “curious incident of the dog in the
night-time”) and empirical application of the IMI (e.g. to collaborative
inquiry learning, in Hakkarainen and Sintonen 2002) require
questions featuring vocabulary absent from the premises. The
absence of an explicit parameter for the range of attention is shown
(in Genot and Gulz 2016) to result in the ultimate failure of the IMI to
account for empirical subjects’ question selection. For alternative
proofs for the results in Hintikka et al. (1999) with explicit
restrictions on the range of attention, see Genot (2018).

5. 5.

The notions of serendipity and serendipity pattern were coined by

Merton (1948) to characterize “the fairly common experience of
observing an unanticipated, anomalous, and strategic datum which
becomes the occasion for developing a new or extending an existing
theory” (p. 506). Merton credited Charles S. Peirce and his notion of
abduction for influencing his definition (for a detailed history of the
term before and after Merton’s paper, see Barber and Merton 2004).
For “the curious incident of the dog in the night-time’ as a serendipity
pattern, see Genot and Gulz (2016); Genot and Jacot (2018).

6. 6.

For the formally-minded: assuming the game-theoretic form of

sequential decision problems (as a sequential 1-player game vs
Nature), contextual focus under volitional control can be
mathematically described as strategic in the game-theoretic sense.
Specifically, recall from a distributed CAM is a probabilistic process
(recalling an item carries a probability for recalling others). In a
sequential decision problem where some steps are ‘recall steps’, the
outcome of such steps would be characterized by a lottery over
possible associations. A strategy assigning to positions where a
decision-maker has to make a move, a lottery between possible
moves, is called a behavioral strategy. Since switching contextual
focus modifies the probability of associating the current content of
one’s attention or imagination with items stored in memory, it affects
the lottery over the possible outcomes of a recall step. And since
 different lotteries characterize different behavioral strategies,
switching contextual focus is tantamount to a change in strategy
selection. A more detailed strategic analysis (complete with an
explicit description of the probabilistic model of recall) is presented
in Genot and Jacot (2018), and argued to support the notion of
“strategic use of memory” in problem-solving consistent with the
Hintikkas’ project. The role of contextual focus in creative thinking
and problem-solving was first identified by Liane Gabora, presented
with its supporting empirical evidence in Gabora (2010), and shown
to better organize the data than Darwinian theories of creative
thinking (like the “blind variation and selective retention” theory) in
Gabora (2011). With J. Jacot, we adapted Gabora’s rejoinder to
Darwinian theories of creative thought in Genot and Jacot (2018) to
show that the strategic analysis of recall offers a better account than
Darwinian theories of serendipity (such as P. van Andel’s widely
accepted account exposed in Van Andel 1994).

7. 7.

For the formally-minded: the Strategy Theorem expresses a

normative claim, namely, that the best discovery strategies for a class
of non-deductive problems (Holmesian problems expressible in a
first-order language) parallel the best deductive strategies for a class
of deductive problems (the deductive transformation of the non-
deductive problems, cf. n. 2, p. 3). However, due to the semi-
decidability of first-order logic, there can be no effective method to
transform a non-deductive problem into a deductive one (some
candidate transformations where the premise set is still too weak to
entail the desired conclusion may generate infinite proof attempts,
and eliminating them would be tantamount to solving the Halting
Problem). Hence, we are in a situation where normative claims are
possible even in the absence of effective rules.