TCU Metadata Format

TrigFuzz release metadata should use explicit distance expressions.

JSON

[
  {
    "cond": "v < 2",
    "distance_expr": "tf_lt((double)v, 2.0)",
    "loc": "line 23 in main.c",
    "seq": 0,
    "conj": 0,
    "w": 1.0,
    "kind": "numeric"
  }
]

Fields:

• distance_expr: C expression compiled into distance_instrument(). It must evaluate to 0.0 when the TCU is satisfied and a positive distance otherwise.
• cond: readable label for humans and reports. It is not the release path for instrumentation.
• loc: source insertion point in the form line N in file.c.
• seq: execution-order group. 0 is first.
• conj: DNF conjunct id. Same (seq, conj) entries are ANDed; different conjuncts at the same seq are ORed by taking the best conjunct distance.
• w: positive normalization weight.
• kind: numeric or binary.

Distance Helpers

Both runtime headers provide:

tf_bool(condition)
tf_lt(a, b)
tf_le(a, b)
tf_gt(a, b)
tf_ge(a, b)
tf_eq(a, b)
tf_add_num(a, b)
tf_sub_num(a, b)
tf_mul_num(a, b)
tf_min2(a, b)

Prefer these helpers over open-coded arithmetic. The common bug pattern is an expression that overflows or underflows before it is cast:

/* Risky when nstrips is unsigned and can be 0. */
tf_gt((double)(nstrips - 1), (double)limit)

/* Preferred. */
tf_gt(tf_sub_num((double)nstrips, 1.0), (double)limit)

LLM Prompting Rule

Ask the model to output distance_expr directly. Do not ask for only cond and then lower it automatically.

Good:

<tf_gt(tf_sub_num((double)nstrips, 1.0), (double)limit), line 5973 in tif_dirread.c, 0, 0, 1.0>

Avoid:

<nstrips - 1 > limit, line 5973 in tif_dirread.c, 0, 0, 1.0>

The Python code still accepts legacy cond-only metadata as a fallback, but that path is not recommended for generated TCUs.