psiloceyeben

/

oracle7-slim

Like 1

At current slim main 0e0e858238cb028b751a85a23f4690a0b0d48483, relation_spec_v2.answer checks uniqueness among subjects for only the first linked target, not across the plausible targets. Two equally ranked identities for the slot can have different unique subjects; reversing their order changes the answer rather than producing ambiguity.

Freshly downloaded source SHA256: 8c712b9af41aafb3945ca72ddbcd22b019ba4340cb212787b085789b0a3596ac (root/wander-train/relation_spec_v2.py). The same blob is in pending PR #2. No existing discussion covers this target-identity problem.

This small fixture runs the actual reader, substituting only its linker/row/type dependencies. It does not claim a particular live natural-language question or full Engine produces this result:

import importlib.util, re, sys, types
from unittest.mock import patch
x = types.SimpleNamespace(DOMINANCE=10, TIME_VARIANT=set(),
    PAST=re.compile(r'\bwas\b'), _show=lambda v, *a: v)
store = types.SimpleNamespace(label=lambda qid: {102:'First spouse',103:'Second spouse'}[qid])
seen = []
def query(sql, args):
    seen.append(args)
    return {101:[102],201:[103]}[args[1]]
v1 = types.SimpleNamespace(X=x, kind_conflict=lambda *a: False,
    _restates=lambda *a: False, _rs=lambda: store, _q=query)
spec = importlib.util.spec_from_file_location('reader', 'root/wander-train/relation_spec_v2.py')
r = importlib.util.module_from_spec(spec)
with patch.dict(sys.modules, {'relation_spec_v1':v1,
    'relation_readers_v1':types.SimpleNamespace(PROPS={})}):
    spec.loader.exec_module(r)
r.compiled = lambda key: {'prop':26,'fwd':[],
    'inv':[re.compile(r'^who is married to (?P<X>.+?)$')]}
r.in_domain = lambda qid, domain: True
for targets in ([101,201], [201,101]):
    r.slot_candidates = lambda slot: (targets,[10,10])
    seen.clear()
    print(r.answer('Who is married to Target?', 'actors', [26]), seen)

Observed: the first order answers subject 102 and queries only (26,101); reverse order answers subject 103 and queries only (26,201). Both target identities are still equally plausible. One subject per target does not establish one answer for the question. The earlier popularity-dominance reduction also does not prove that a discarded sense is impossible.

A bounded proposed repair is published as inverse-target-appliance.patch, with six focused consumer tests and assumptions/limits. It refuses recognized multi-target inverse requests before dominance, carries a specific refusal reason through the 161 copied relation organs, and preserves it through the tested Combo arbitration instead of permitting a rival answer. A fresh git apply --check against exact current-main blobs for all 163 changed files succeeded.

This is an issue rather than a release-ready PR because the full head can answer before invoking that Family. The proposed patch's recorded six tests exercise actual organ/Combo consumers with a controlled source, not complete Engine arbitration, shipped panels or source completeness. Missing type facts, temporal qualifiers, six-result linker truncation and intended-identity coverage remain separate assumptions; this repair must not claim to prove them. Integration should preserve the ambiguity refusal through all relevant consumers, then qualify the actual release route.