In classic Optimality Theory (Prince and Smolensky, 1993), EVAL is seen as a function that makes only a two-way distinction in the candidate set - between the best candidate and the set of losers. I argue that we should think differently about EVAL. EVAL should be seen as a function that imposes a harmonic rank-ordering on the complete candidate set. In this way an OT grammar provides enough information that any two candidates can be compared for their relative well-formedness. We can compare even two losers (two ungrammatical forms) and determine which is more and which is less well-formed.

These finer distinctions made by EVAL are linguistically relevant. I show that these distinctions are used by language users inter alia when they make well-formedness judgments and when they perform lexical decision tasks.

(i) Well-formedness judgments. When language users have to rate the well-formedness of non-words, they rate non-words that are more well-formed in terms of the grammar better than non-words that are less well-formed.

(ii) Lexical decision. Language users use inter alia the information supplied by grammar when they make lexical decisions. A non-word that is more well-formed in terms of the grammar is considered more seriously as a possible word. Such a non-word is therefore rejected more slowly than a non-word that is less well-formed.

In well-formedness judgments and in lexical decision tasks, language users rely on information about the relative well-formedness of tokens. This implies that language users can make finer distinctions than simply between grammatical and ungrammatical forms. We therefore also need a grammar that can make these finer distinctions.

I illustrate this using a restriction on possible words in English. English allows words of the form [sCvC] where both C's are [t], but not where both C's are [p] or [k] - i.e. "state" but *"spape", *"skake". We therefore know that [sTvT] is more well-formed that both [sPvP] and [sKvK] in English. Based on cross-linguistic evidence, I argue that [sKvK] is also more well-formed than [sPvP]. I then develop an OT account of this restriction in English that imposes the following harmonic rank-ordering: [sTvT] > [sKvK] > [sPvP].

In order to test whether this account is correct, I conducted two sets of experiments with English listeners.

(i) Well-formedness judgments. I presented English listeners with non-words of the form [sTvT], [sKvK] and [sPvP]. Their task was to rate the well-formedness of the non-words. On average, they rated [sTvT]-tokens the best, then [sKvK]-tokens, and then [sPvP]-tokens.

(ii) Lexical decision. I presented English listeners with a list of tokens containing both non-words and words. Their task was to discriminate between words and non-words. On average, they detected non-words of the form [sTvT] the slowest, then non-words of the form [sKvK] faster, and finally non-words of the form [sPvP] the fastest.

These experiments therefore provide support for the OT account of this restriction that I developed, and also for the rank-ordering model of EVAL.