refactor(plausibility): grounding-based rules + R3 transcript contradiction

The plausibility prompt previously instructed the judge with "lexical
patterns" and "no judgment calls" and listed five drop rules as exact
token enumerations: R3 role-leak token list, R4 unresolved-placeholder
regex tied to the constrained {Person|City|Country|Org|Thing}
vocabulary, R5 sentinel-object token list, R6 dot-separated HA-style
identifier.

Two pivots applied here:

  * Drop R4. The constrained-prefix regex became incoherent with the
    open-vocabulary anonymizer pivot (commit d8de45c) and was already
    structurally redundant with _strip_residual_placeholders running
    inside _apply_bindings before the deanon-stage plausibility judge.
  * Add R3 transcript contradiction. The transcript was rendered into
    the prompt but no rule referenced it. The judge now drops facts
    whose object contradicts the transcript on the same
    (subject, predicate). Live probe confirms deterministic catch
    across 3 runs at temp=0.

  * Soften R3 (Role leak → Conversation-role reference), R4 (Empty /
    sentinel object → Content-free object), R5 (System entity ID →
    Namespaced system identifier) from "Subject or object is exactly
    one of: ..." token enumerations to grounding-flavoured category
    descriptions with the original tokens kept as illustrative
    parentheticals. The rule examples are retained verbatim.

The structure of the prompt is preserved exactly: KEEP / DROP / Input /
Output sections; numbered R-rules with bold name + description +
`Example: [N] {...} → DROP (Rk) — include N` line; Input section
rendering transcript and indexed facts. The _DEFAULT_PLAUSIBILITY_PROMPT
inline default is kept byte-equivalent to the file (parity-gated by
tests/test_prompts_contract.py::test_inline_default_matches_file).

Empirical baseline (probe at /tmp/probe_plausibility_grounding.py,
10-fact synthetic input on local Mistral 7B): deterministic
{"drop": [3, 6, 8, 9]} across 3 runs. Catches R1 (self-loop), R2
(name-swap pair, both indices), R3 (transcript contradiction). Misses
on this probe: R4 (Assistant), R5 (Unknown), R6 (system identifier).
The misses share a pattern - Mistral 7B's KEEP-bias on ambiguous cases
overrides the rule when the fact reads as plausible content
(Assistant could be a name; Unknown could be a noun; controls
media_player.X sounds like a smart-home claim). Strengthening these
rules without regressing the structural KEEP-default behaviour is
tracked as an open optimization in the project memory; revisit
condition is whether SOTA enrichment in production actually emits
these failure modes.

Pre-existing miss: R6 system identifier was missed by the original
lexical prompt on the same probe input. Pattern-match wording vs.
semantic-judgment wording made no difference for the specific fact
shape "Person_1 controls media_player.X" - Mistral reads it as a
legitimate smart-home claim regardless. The miss is a Mistral 7B
capability gap, not a prompt bug.

Pytest CI-parity scope (-m "not gpu", --timeout=60): 3656 passed,
26 deselected, 0 failed in 11:19.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: tiberius1701d

configs/prompts/sota_plausibility.txt

@@ -1,34 +1,32 @@
-You are filtering enriched personal facts from a voice assistant conversation. Inputs may be anonymized (placeholders like Person_1) or real-named — apply the same rules either way.
-
-You decide DROP or KEEP per fact. Output ONLY the indices of facts that match a DROP rule. Never echo, modify, or add facts.
+You are filtering enriched personal facts from a voice assistant conversation. Decide KEEP or DROP per fact, grounding each against the transcript. Output ONLY the indices of facts that match a DROP rule. Never echo, modify, or add facts.
 
 ## KEEP — leave the fact in the output
 
-Default action. Apply whenever none of the DROP rules below match the fact's lexical pattern. Do not drop on semantic judgment. When uncertain, KEEP.
+Default action. Keep the fact when its subject, object, and predicate are supported by the transcript (literally or by clear implication). When uncertain, KEEP.
 
 Example: `[0] {{"subject": "Alex", "predicate": "lives_in", "object": "Portland"}}` → KEEP (do not include 0 in the output)
 Example: `[1] {{"subject": "Alex", "predicate": "likes", "object": "Uptown Funk"}}` → KEEP (do not include 1 in the output)
 
 ## DROP — emit this fact's index
 
-Apply when any of the lexical patterns below match. No judgment calls.
+Apply when any of the rules below matches.
 
 **R1. Self-loop.** `subject` and `object` are the same string (case-insensitive), regardless of predicate.
 Example: `[2] {{"subject": "Person_1", "predicate": "has_name", "object": "Person_1"}}` → DROP (R1) — include 2 in the output.
 
 **R2. Name-swap pair.** Both `A has_name B` and `B has_name A` are in the input (also `named`, `is`, `equals`). Drop both indices.
 Example: `[3] {{"subject": "Alex", "predicate": "is", "object": "Bob"}}` AND `[4] {{"subject": "Bob", "predicate": "is", "object": "Alex"}}` → DROP both (R2) — include 3 and 4.
 
-**R3. Role leak.** Subject or object is exactly one of: "Assistant", "User", "Speaker", "the bot", "the model". Note: "Person_1" / "City_1" / "Org_1" / "Thing_1" are valid entity placeholders, NOT role leaks.
-Example: `[5] {{"subject": "Assistant", "predicate": "responded_to", "object": "Alex"}}` → DROP (R3) — include 5.
+**R3. Transcript contradiction.** The transcript states a different value for the same subject and predicate.
+Example: `[5] {{"subject": "Alex", "predicate": "lives_in", "object": "Tokyo"}}` (when the transcript says Alex lives in Portland) → DROP (R3) — include 5.
 
-**R4. Unresolved placeholder in real-name input.** When the input is real-named (no `Person_N` / `City_N` / `Country_N` / `Org_N` / `Thing_N` placeholders expected in the transcript), drop any fact whose subject or object still matches `^(Person|City|Country|Org|Thing)_\d+$`.
-Example: `[6] {{"subject": "Alex", "predicate": "owns", "object": "Person_4"}}` → DROP (R4, real-name input) — include 6.
+**R4. Conversation-role reference.** Subject or object refers to a conversation-role label (e.g. "Assistant", "User", "Speaker", "the bot", "the model") rather than a real-world entity introduced by the transcript.
+Example: `[6] {{"subject": "Assistant", "predicate": "responded_to", "object": "Alex"}}` → DROP (R4) — include 6.
 
-**R5. Empty / sentinel object.** Object is exactly one of: "", "Unknown", "None", "Various", "Something", "N/A".
+**R5. Content-free object.** Object provides no transcript-grounded claim — empty, whitespace-only, or a content-free placeholder (e.g. "Unknown", "None", "Various", "Something", "N/A").
 Example: `[7] {{"subject": "Alex", "predicate": "is_from", "object": "Unknown"}}` → DROP (R5) — include 7.
 
-**R6. System entity ID.** Subject or object contains a dot-separated HA-style identifier (e.g. `media_player.sonos_office`, `sensor.temperature_kitchen`).
+**R6. Namespaced system identifier.** Subject or object is a system / automation identifier — a dot-separated or URI-shaped namespaced token (e.g. `media_player.sonos_office`, `sensor.temperature_kitchen`, `climate.kitchen_thermostat`) referring to a device, sensor, or service rather than a personal entity.
 Example: `[8] {{"subject": "Alex", "predicate": "controls", "object": "media_player.sonos_office"}}` → DROP (R6) — include 8.
 
 ## Input

paramem/graph/extractor.py

@@ -1810,22 +1810,21 @@ def _validate_with_ha_context(graph: SessionGraph, ha_context: dict) -> SessionG
 # fmt: off
 _DEFAULT_PLAUSIBILITY_PROMPT = (
     "You are filtering enriched personal facts from a voice assistant conversation. "  # noqa: E501
-    "Inputs may be anonymized (placeholders like Person_1) or real-named — apply the same rules either way.\n"  # noqa: E501
-    "\n"
-    "You decide DROP or KEEP per fact. Output ONLY the indices of facts that match a DROP rule. "  # noqa: E501
+    "Decide KEEP or DROP per fact, grounding each against the transcript. "
+    "Output ONLY the indices of facts that match a DROP rule. "
     "Never echo, modify, or add facts.\n"
     "\n"
     "## KEEP — leave the fact in the output\n"
     "\n"
-    "Default action. Apply whenever none of the DROP rules below match the fact's lexical pattern. "  # noqa: E501
-    "Do not drop on semantic judgment. When uncertain, KEEP.\n"
+    "Default action. Keep the fact when its subject, object, and predicate are supported by the transcript "  # noqa: E501
+    "(literally or by clear implication). When uncertain, KEEP.\n"
     "\n"
     'Example: `[0] {{"subject": "Alex", "predicate": "lives_in", "object": "Portland"}}` → KEEP (do not include 0 in the output)\n'  # noqa: E501
     'Example: `[1] {{"subject": "Alex", "predicate": "likes", "object": "Uptown Funk"}}` → KEEP (do not include 1 in the output)\n'  # noqa: E501
     "\n"
     "## DROP — emit this fact's index\n"
     "\n"
-    "Apply when any of the lexical patterns below match. No judgment calls.\n"
+    "Apply when any of the rules below matches.\n"
     "\n"
     "**R1. Self-loop.** `subject` and `object` are the same string (case-insensitive), regardless of predicate.\n"  # noqa: E501
     'Example: `[2] {{"subject": "Person_1", "predicate": "has_name", "object": "Person_1"}}` → DROP (R1) — include 2 in the output.\n'  # noqa: E501
@@ -1835,24 +1834,23 @@ def _validate_with_ha_context(graph: SessionGraph, ha_context: dict) -> SessionG
     'Example: `[3] {{"subject": "Alex", "predicate": "is", "object": "Bob"}}` AND '
     '`[4] {{"subject": "Bob", "predicate": "is", "object": "Alex"}}` → DROP both (R2) — include 3 and 4.\n'  # noqa: E501
     "\n"
-    "**R3. Role leak.** Subject or object is exactly one of: "
-    '"Assistant", "User", "Speaker", "the bot", "the model". '
-    'Note: "Person_1" / "City_1" / "Org_1" / "Thing_1" are valid entity placeholders, NOT role leaks.\n'  # noqa: E501
-    'Example: `[5] {{"subject": "Assistant", "predicate": "responded_to", "object": "Alex"}}` → DROP (R3) — include 5.\n'  # noqa: E501
-    "\n"
-    "**R4. Unresolved placeholder in real-name input.** When the input is real-named "  # noqa: E501
-    "(no `Person_N` / `City_N` / `Country_N` / `Org_N` / `Thing_N` placeholders expected in the transcript), "  # noqa: E501
-    "drop any fact whose subject or object still matches "
-    r"`^(Person|City|Country|Org|Thing)_\d+$`."  # noqa: E501
+    "**R3. Transcript contradiction.** The transcript states a different value for the same subject and predicate.\n"  # noqa: E501
+    'Example: `[5] {{"subject": "Alex", "predicate": "lives_in", "object": "Tokyo"}}` '
+    "(when the transcript says Alex lives in Portland) → DROP (R3) — include 5.\n"
     "\n"
-    'Example: `[6] {{"subject": "Alex", "predicate": "owns", "object": "Person_4"}}` → DROP (R4, real-name input) — include 6.\n'  # noqa: E501
+    "**R4. Conversation-role reference.** Subject or object refers to a conversation-role label "  # noqa: E501
+    '(e.g. "Assistant", "User", "Speaker", "the bot", "the model") '
+    "rather than a real-world entity introduced by the transcript.\n"
+    'Example: `[6] {{"subject": "Assistant", "predicate": "responded_to", "object": "Alex"}}` → DROP (R4) — include 6.\n'  # noqa: E501
     "\n"
-    "**R5. Empty / sentinel object.** Object is exactly one of: "
-    '"", "Unknown", "None", "Various", "Something", "N/A".\n'
+    "**R5. Content-free object.** Object provides no transcript-grounded claim — empty, "  # noqa: E501
+    'whitespace-only, or a content-free placeholder (e.g. "Unknown", "None", "Various", "Something", "N/A").\n'  # noqa: E501
     'Example: `[7] {{"subject": "Alex", "predicate": "is_from", "object": "Unknown"}}` → DROP (R5) — include 7.\n'  # noqa: E501
     "\n"
-    "**R6. System entity ID.** Subject or object contains a dot-separated HA-style identifier "  # noqa: E501
-    "(e.g. `media_player.sonos_office`, `sensor.temperature_kitchen`).\n"
+    "**R6. Namespaced system identifier.** Subject or object is a system / automation identifier — "  # noqa: E501
+    "a dot-separated or URI-shaped namespaced token "
+    "(e.g. `media_player.sonos_office`, `sensor.temperature_kitchen`, `climate.kitchen_thermostat`) "  # noqa: E501
+    "referring to a device, sensor, or service rather than a personal entity.\n"
     'Example: `[8] {{"subject": "Alex", "predicate": "controls", "object": "media_player.sonos_office"}}` → DROP (R6) — include 8.\n'  # noqa: E501
     "\n"
     "## Input\n"

tests/test_prompts_contract.py

@@ -175,17 +175,32 @@ def test_renders_without_format_errors(self):
         assert "{facts_json}" not in rendered
 
     def test_lists_drop_rules(self):
-        """Plausibility judge relies on six numbered drop rules (R1-R6). Verify they
-        all exist so a prompt edit that removes a rule is caught at unit-test time.
+        """Plausibility judge relies on six numbered drop rules (R1-R6).
+
+        The prior R4 ("Unresolved placeholder in real-name input") was
+        tied to the constrained ``^(Person|City|Country|Org|Thing)_\\d+$``
+        regex that became incoherent with the open-vocabulary anonymizer
+        pivot.  It was also structurally redundant with
+        ``_strip_residual_placeholders`` running inside ``_apply_bindings``
+        at the deanon stage, before plausibility.
+
+        The grounding refactor revised the remaining rules: lexical
+        token lists became illustrative parentheticals, and a new R3
+        (transcript contradiction) closes the gap that the prior
+        "no judgment calls" framing left open.  The structure of the
+        prompt and of the examples is unchanged.
+
+        Verify each rule's identifying substring still exists so a
+        prompt edit that removes a rule is caught at unit-test time.
         """
         tmpl = _load_prompt("sota_plausibility.txt", _DEFAULT_PLAUSIBILITY_PROMPT)
         required_rules = [
             "self-loop",  # R1
             "name-swap",  # R2
-            "role leak",  # R3
-            "placeholder",  # R4
-            "sentinel",  # R5
-            "system entity",  # R6
+            "contradiction",  # R3 — new transcript-grounded rule
+            "conversation-role",  # R4 — was "Role leak", grounded
+            "content-free",  # R5 — was "Empty / sentinel"
+            "system identifier",  # R6 — was "System entity ID"
         ]
         for rule in required_rules:
             assert rule.lower() in tmpl.lower(), f"Plausibility prompt missing rule: {rule!r}"