<?xml version="1.0" standalone="yes"?>
<Paper uid="H86-1016">
  <Title>Hypotheticals as Heuristic Device</Title>
  <Section position="2" start_page="169" end_page="173" type="metho">
    <SectionTitle>
4. Heuristics for Generating Hypotheticals
</SectionTitle>
    <Paragraph position="0"> Basically what HYPO does is to start with a given fact situation, or seed case, and generate legally relevant or plausible derivative hypotheticals by modifying the seed case. Since one cannot explore all the &amp;quot;legally&amp;quot; possible (in the sense of syntactic legal move), one needs to explore the space heuristically. Dimensions provide a handle on how to do this exploration in a legally meaningful way.</Paragraph>
    <Paragraph position="1"> The process occurs in two steps:  (1) analyze the seed case; (2) generate legally relevant derivative hypotheticals.</Paragraph>
    <Paragraph position="2">  Step one is accomplished by the CASE-ANALYSIS module and results in the case-analysis-record described in the previous section. To recall, this is like a &amp;quot;legal-diagnosis&amp;quot;. Step two is accomplished by the HYPO-GEN module which given high level argument goals (e.g., generate a slippery slope sequence to refute side l's position), uses the case-analysisrecord, and heuristics like the following to generate hypotheticals derived from the seed case:  H1. Pick a near miss dimension and modify the facts to make it applicable. H2. Pick an applicable dimension and make the case weaker or stronger along that dimension.</Paragraph>
    <Paragraph position="3"> H3. Pick a dimension related to one of the applicable dimensions and apply or 2.</Paragraph>
    <Paragraph position="4"> H4. Pick an applicable dimension and make the case extreme with respect to that dimension.</Paragraph>
    <Paragraph position="5">  H5. Pick a target case ,that is a win and, using 1 and 2, move the seed case toward it to create a near win.</Paragraph>
    <Paragraph position="6"> In order to illustrate these methods, we will use the following hypothetical case, Widget-King v. Cupcake, whose facts are as follows: Plaintiff Widget-King and defendant Cupcake are corporations that make competing products. Widget-King has confidential information concerning its own product. Cupcake gained access to Widget-King's confidential information. Cupcake saved expense developing its competing product.</Paragraph>
    <Paragraph position="7"> The parts of the case-analysis-record for Widget-King v. Cupcake that are relevant for the following sections are: app\] icable dimensions: competitive-advantage-gained near-miss dimensions : secrets-voluntarily-disclosed; vertical-kr.owledge relevant CKB cases: Telex v. IBM 4.1. Make a near miss dimension apply To make a hypothetical out of a fact situation according to this heuristic method, HYPO selects a near miss dimension and &amp;quot;fills in&amp;quot; the missing prerequisites. HYPO instantiates objects and makes appropriate cross references among objects' slots so that the missing factual predicates are satisfied. For example, secrets-voluntarily-disclosed would apply to Widget-King but for the fact that the confidential information had not been disclosed to anyone. The program instantiates, let us say, five disclosures and sets the subject of the disclosures to be the confidential information. As discussed below, the number of disclosures, five, may be derived from an actual case that the program is considering in the context of making up the hypothetical, or it may be somewhat arbitrarily chosen by the program from within the range of the dimension.</Paragraph>
    <Paragraph position="8"> 4.2. Make a case weaker or stronger HYPO generates a derivative hypothetical weaker/stronger than the seed case by using the information it knows about dimensions. It can make a case weaker or stronger in two ways: (1) independently of the &amp;quot;caselaw&amp;quot; represented by the CKB; or (2) based on the CKB using a weak form of analogy. To accomplish a CKB-independent strengthening/weakening, HYPO simply changes the values of a focal slot in the manner specified by the direction-tostrengthen slot; the amount of change is somewhat arbitrary. To accomplish a CKB-based modification, for instance to strengthen, HYPO first chooses a case that (a) shares the dimension being manipulated, and (b) is further along the dimension in the stronger direction. HYPO then adjusts the values of the focal slots of the seed in the stronger direction so that the derivative case is stronger than the &amp;quot;precedent&amp;quot; chosen from the CKB. These changes can involve numerical, symbolic or Boolean values. For symbolic values, this means using a partial ordering on values.</Paragraph>
    <Paragraph position="9"> Modifications can involve more than one focal slot, for instance a ratio. For example, given our fact situation involving Widget-King and Cupcake which involves some expenditure of money by Widget-King for product development, the Telex v. IBM case in the CKB is relevant. In Telex the ratio of paintiff's to defendant's expenditures was 2:1 (and the paintiff won). So to strengthen Widget-King's case, change ratio of Widget-King's to Cupcake's expenses to be at least 2:1. All example of such ratio manipulation can also be found in \[McCarty &amp; Sridharan, 1981\].</Paragraph>
    <Paragraph position="10"> Even a simple change in a single numerical focal slot value can have serious legal impications. Again consider our Widget-King case, as modified by the introduction of 5 disclosees, and make it weaker along the secrets-voluntarily-disclosed dimension by using cases from the CKB. HYPO increases the number of Widget-King disclosees from 5 to 150 based on :Midland-Ross which was decided for the defendant because there were too many disclosees (100) and now Widget-King has passed the 100-disciosee threshold. Note, Widget-King could  still rely on Data-General and a(gue that since the plaintiff won in that case (with 6000 disclosees), it should still win with only 150. HYPO could make the case weaker still by increasing the number of disclosees near or above 6000, the highest value in the CKB or even greater (in a CKB-independent way) to the highest value allowed by HYPO.</Paragraph>
    <Paragraph position="11"> There are pros and cons to the two methods. The CKB-independent method is easy to do, but the precedential value of the derivative hypothetical is not predictable. The CKB-based method generates a hypo with known precedents; the drawback is that it call get complicated. HYPO tries to do CKB-based strengthening/weakening first, if it can't (e.g., because no relevant case exists in the CKB), it uses the CKB-independent approach. In either case, the task of actually generating the explanation (as we did above) why the hypo is stronger or weaker belongs to HYPO's EXPLANATION module.</Paragraph>
    <Paragraph position="12"> 4.3. Generate a hypo on a related dimension The dimensions disclosures-subject-to-restriction and secrets-voluntarily-disclosed are related; in particular they conflict with one another. Dimensions conflict where there is a particular case to which the dimensions apply and the facts of the case make it strong for the plaintiff on one dimension and weak on the other. Such a case is called a conflict-example. Data-General is a conflict-example: it is weak for the plaintiff along the secrets-voluntarily-disclosed dimension (100 disclosees) and strong for the plaintiff along the disclosures-subject-to-restriction dimension (each disclosure subject to nondisclosure agreement). In Data-General, the conflict was resolved in favor of the plaintiff.</Paragraph>
    <Paragraph position="13"> A hypothetical on a related dimension can be generated by taking the seed case and adding facts sufficient to make the related dimension apply to it in a manner similar to that with heuristic H1. For example, the Widget-King case, as modified by H1 and H2 above, can be further modified so that disclosures-subject-to-restriction applies by making all of the disclosures subject to nondi~closure agreements, in this example, the related dimension is also a near miss dimension but that need not always be true.</Paragraph>
    <Paragraph position="14"> A hypothetical generated on a conflicting dimension is interesting because it is an example of a case where, at least arguably, facts associated with one dimension can override the effects of the other dimension's facts.</Paragraph>
    <Paragraph position="15"> 4.4. Examine an extreme case To generate an extreme case, HYPO simply changes the value of a focal slot to be an extreme of its range of values. This can also be done in either a CKB-based or CKB-independent manner. The former method pushes the slot value to the extreme actually existing in a case in the CKB, the latter simply pushes the slot value to its permissible extreme.</Paragraph>
    <Paragraph position="16"> For instance, the extreme case on the strongest end of the secrets-voluntarily-disclosed dimension for Widget-King is the facts as stated above with the exception that there are 0 disclosees. The other extreme is the maximum value for number of disclosees which in the CKB is 6000 and which in HYPO is 10,000,000.</Paragraph>
    <Paragraph position="17"> 4.5. Manipulating a near wi.</Paragraph>
    <Paragraph position="18"> A near win hypo is one in which a seed fact situation is weak on behalf of, let us say, the plaintiff. It can be &amp;quot;moved&amp;quot; in the direction of a real target case from the CKB that has been decided in favor of the plaintiff. Using methods H1 through H3, HYPO endows the seed situation with the facts to make the case strong for the plaintiff. As a result, the target case becomes relevant to the seed hypothetical and an argument can be made, based on the pro-plaintiff target case, that the hypo should be decided in favor of the plaintiff. Correspondingly a near win hypo can start with a pro-plaintiff fact situation and be moved ill the opposite direction away from the pro-plaintiff target case or towards a pro-defendant target case. For example, consider two cases: Telex, which we have already seen above, and Automated Systems, where court held in favor of the defendant where the confidential information that  the plaintiff wanted to protect was about a customer's business operations, that is, the knowledge was about a &amp;quot;vertical rnarket&amp;quot;. Using the Telex case as a seed, and Automated Systems as target, HYPO could make Telex a near win by making IBM's confidential information be vertical knowledge (i.e., be about a customers business operations). As a result, an argument could be based on Automated Systems that, in the hypo, defendant Telex should win.</Paragraph>
    <Paragraph position="19"> 5. Examples of Heuristic HYPO Exploration HYPO's heuristically guided generation of hypotheticals makes it possible to explore a fact situation's legal significance in a manner not unlike the sequence of hypotheticals in the creche example from the Lynch case oral argument.</Paragraph>
    <Paragraph position="20"> Suppose (a) the original Widget-King case is modified so that the confidential information is about customer business operations. Suppose on appeal to the Supreme Court, Cupcake's counsel, citing Automated Systems, has just argued to the Justices that his client should win because vertical knowledge is not protectible as a trade secret. One can imagine a Justice posing the following line of hypotheticals: .Q: What never? Suppose (b) Widget-King's alleged trade secret information, eventhough it was vertical knowledge, helped it to produce its competing product in half the time like in the Telex case? Q: Suppose (c) the vertical knowledge allowed Widget-King to bring its product to market in one fourth the time and at one fourth the expense.</Paragraph>
    <Paragraph position="21"> Q: Suppose (d) that Cupcake paid a large sum to a former employee of Widget-King to use the information to build a competing product, as Telex did. Wouldn't the information be protectible as a trade secret then?.</Paragraph>
    <Paragraph position="22"> In this example, heuristic methods 1,2,3 and 5 are at work. Near miss dimension verticalknowledge is used with 1 to create the intial hypo (a). The modification at (b) is produced by 5 and 2 using the the Telex case as the target. Method 2 is used to make the stronger hypo at (c). Methods 5, 1 and 2 are used to create the hypo at (d) where the near miss dimension is common-employee-paid-to-change-employers.</Paragraph>
    <Paragraph position="23"> It is interesting that a previous version of HYPO serendipitously generated a hypothetical very much like this. The starting point was a fact situation presenting a very strong position for the plaintiff along various dimensions: it involved alleged misappropriation of plaintiff's unique, novel technical knowledge about computer system hardware for a particular purpose, knowledge that was not learnable by an employee working for one of the plaintiff's competitor's and that conferred on the plaintiff a year's competitive advantage in bringing its product to market. Then, by accident, the hypo was changed by turning the technical knowledge about hardware into vertical knowledge about bank accounting practices. Although according to the Automated Systems case, the new hypo presented a very much weakened position for the plaintiff, it was immediately apparent to the attorney using the program that the Automated Systems case was distinguishable -- it did not involve the facts that the knowledge, though vertical, was unique, novel, not learnable elsewhere and conferred a substantial competitive advantage on its possessor -- and suggested the germ of an argument for the protectibility of vertical knowledge -- demonstrate that the vertical knowledge is unique, novel, etc.</Paragraph>
    <Paragraph position="24"> Since that accidental discovery, we have provided the system with the above heuristic methods so that, given a case, it can generate a hypo that is distinguishable from the case in a legally significant way. Starting from a real case, methods 3 and 5, in particular, are recipes for creating hypo's with facts that justify a different holding from the real case. Our goal is for the system itself to realize that the hypo is significantly distinguishable and why and to generate such hypos on purpose to make points in an argument.</Paragraph>
    <Paragraph position="25"> Having reached step (d) in the above extended example, a hypothetical has been constructed that is fairly strong for the plaintiff. But plaintiff's position can be eroded by moves along  other dimensions. One can imagine the scene at 11 p.m. in the oak-paneled library at 14 Wall Street as two first year associate attorneys, assigned to preparing an initial memorandum as to the strengths of Widget-King's claim against Cupcake, play devil's advocate with the facts: Q: Suppose (e) that Widget-King made disclosures to 100 outside persons as in the Midland-Ross case.</Paragraph>
    <Paragraph position="26"> Q: Well, maybe (f) all of the disclosees entered into nondisclosure agreements as in Data-General. Under that case, Widget-King (g) could have made restricted disclosures to as many as 6000 people.</Paragraph>
    <Paragraph position="27"> Q: What if (h) Widget-King made restricted disclosures to 10,000,000 people. Is it still a secret? (Not an idle hypothetical in this day of mass marketing of software.) Q: Are the nondisclosure agreements enforceable? What did all of these people get in exchange for agreeing not to disclose the secret? Suppose (i) that the disclosees did not receive anything of value for entering into the nondisclosure agreements? With secrets-voluntarily-disclosed as near miss dimension and the Midland-Ross case as target, the hypo at (e) can be generated from (d) using methods 5, 1 and 2. (f) represents a method 3 move to a conflict dimension, disclosures-subject-to-restriction. We assume that the Data-General case has been recognized as a conflict-example. Otherwise this could be regarded as a method 5 move with Data-General as a target. Using method 4, the hypo at (g) has been moved to the extreme value in Data-General and at (h) to the extreme of the range of the dimension. The program does not know that a secret told to 10,000,000 people is not a secret, even if they promise not to tell anyone else, but the program does know that two dimensions conflict and that moving to an extreme on one dimension may cause the conflict to be moot. Having exhausted the possibilities for weakening the case along the secrets-voluntarily-disclosed dimension, the program moves, using methods 1 and 2, to a dimension that became a near miss as soon as nondisclosure agreements came into the hypo at (f), agreement-supportedby-consideration. null One can also analyze the sequence of hypotheticals about the civic creche display from the Lynch case oral argument in terms of the dimensional model and heuristics for building hypo's. The justices make the basic fact situation weaker and stronger along a dimension that might be called focus-of-attention: they remove all of the secular images leaving only the religious one, they physically shrink the symbol to an extreme and relegate it to a corner, they remove the religious symbols and leave the secular ones. They weaken plaintiff's case along the dimension of civic-content-message by moving it to a municipal art museum or the frieze of a courtroom. They compare the case along the dimension of government-involvement to an extreme example, the Pope's mass on the Mall.</Paragraph>
  </Section>
  <Section position="3" start_page="173" end_page="174" type="metho">
    <SectionTitle>
6. Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper, we have discussed an aspect of reasoning involving the use of hypothetical cases.</Paragraph>
    <Paragraph position="1"> In particular, we have discussed how our case-based legal reasoning program HYPO currently uses case examples, dimensions, and five or so heuristic methods to compare the legal consequences of facts and to generate hypothetical fact situations to augment and explore its case base. The hypos help accomplish analysis tasks, such as testing the sensitivity of positions and relating a fact situation to significant past cases, and argument tasks, such as generating a slippery slope to refine or refute an argument and controlling the course of argument. HYPO's heuristics involve (1) strengthening/weakening of a case; (2) taking the case to extremes: (3) making a near miss case a winning one; (4) manipulating a near win; and (5) examining a case along a related dimension.</Paragraph>
    <Paragraph position="2"> As indicated earlier, one of our performance goals for HYPO is to have HYPO generate 3-ply argument exchanges which involve a heavy dose of case-based reasoning like distinguishing cases and using hypotheticals. Eventually we hope to bring together our descriptive work on argument moves and hypotheticals \[Rissland, 1985; Stucky, 1985\] with our computational 3-ply argument work. We also hope that this work on HYPO will cross-potentiate with work on the intelligent selection of examples for learning systems, a topic, we feel has been too often glossed over. The heuristic generation of hypotheticals is a step towa.rds both these goals.</Paragraph>
    <Paragraph position="3"> However even as they now stand, HYPO's current hypothetical reasoning powers can be helpful in formulating, testing, debugging, and learning in case-based tasks.</Paragraph>
  </Section>
class="xml-element"></Paper>