Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

The six refused traits

Nine of 140 trait words were refused by the constitution writer as not naming a disposition; three Goldberg markers were re-screened on their inventory sense and accepted, leaving six refusals and 134 trainable traits.

134 = 100 + 40 - 6. This page is the arithmetic.

The constitution writer was given an escape hatch: if the word does not name a coherent human disposition it must reply NOT_A_TRAIT: <reason> rather than invent a character (Constitution generation). Nine of the 140 words were refused on the first pass. Three of those nine were Goldberg markers refused on the wrong sense of the word, were re-screened, and were accepted. Six refusals stand, all from the Lexicon draw, and those six were never trained.

The first nine

From qwen35/backups/constitutions.json.pre-generic-anchor.bak, which holds 140 entries of which 9 carry a rejected field: Verbal, Prompt, Complex, Unconformable, Frightened, Busy, Defenseless, Significant, Sleepy.

qwen35/constitutions.py:report sets a hard stop at more than 15 of 140 rejected — "the trait source is suspect" — and exits 2. Nine did not trip it.

The three that were re-screened

qwen35/regen_goldberg_senses.py exists for Verbal, Prompt and Complex. Its argument: these three are in Goldberg's published 100 unipolar markers, meaning they are administered to human subjects as personality adjectives, and the teacher screened the dictionary sense instead:

The glosses in the right-hand column are GLOSSES in qwen35/regen_goldberg_senses.py and are stored on the accepted entries as sense_disambiguated (qwen35/constitutions.json#Verbal.sense_disambiguated and the two siblings). The script's own stated stake: "Dropping them costs Intellect 2 of its 20 markers, and Intellect is the factor a congruence result is most fragile on."

Three properties of the re-screen are worth recording because they are what make it a screen rather than a forced pass:

• the model, prompt template, max_tokens and temperature are imported verbatim from constitutions.py, not retyped — "a second screen run against a paraphrased prompt is not the same screen";
• one paragraph is spliced in immediately before the exception paragraph, so the NOT_A_TRAIT escape hatch is still the last thing the teacher reads, and it is explicitly re-armed for the disambiguated sense;
• "A trait that refuses again is NOT written -- that is a real finding about the trait, and it is reported."

All three were accepted. Exactly three entries in constitutions.json carry sense_disambiguated.

The six that stand

From qwen35/constitutions.json, verbatim rejected texts:

• Unconformable — "Unconformable" is a geological term describing rock strata with a discontinuous boundary, not a human character disposition.
• Frightened — "Frightened" names a transient emotional state or condition, not a stable character disposition — it describes how someone feels in a moment, not an enduring trait that shapes personality across contexts.
• Busy — "Busy" describes a situational condition of having much to do, not a stable dispositional orientation toward the world — it is a transient state, not a character trait.
• Defenseless — Defenseless describes a situational condition or relational state (lacking protection or defense against something), not a stable character disposition a person embodies from the inside.
• Significant — "Significant" is a relational property describing magnitude or importance relative to a standard, not a stable disposition or character trait a person can embody.
• Sleepy — Sleepy describes a transient physiological state, not a stable character disposition.

All six are from the Lexicon draw; none is a Goldberg marker.

Where the six leave the pipeline

qwen35/gen_pairs.py skips any trait whose constitutions.json entry is a rejected record. Every pair-generation log opens with the same line (qwen35/genpairs.log:4-5):

traits 134 (skipped 6 without a constitution)  prompts 500  target pairs 67000
  skipped: Unconformable (...), Frightened (...), Busy (...),
           Defenseless (...), Significant (...), Sleepy (...)

The blog page states the same thing for a reader: "Forty were drawn; six were refused by the constitution writer as states or situations rather than dispositions (sleepy, busy, frightened...) and never trained" (qwen35/build_blog_page.py).

A note on the screen as method

This is the only quality screen the Lexicon draw gets. Draws 1 and 2 died of dictionary-based filtering (The two discarded secondary draws), so the replacement question is "can a competent writer render this as a person?", asked of the model that then has to render it (qwen35/constitutions.py docstring). Its observed selectivity is 9 of 140 on the first pass and 6 of 140 after the sense correction.