feat(sentiment): promote Qwen3.8 27B to the panel as a fifth voice

Moves the self-hosted route from PILOT_CANDIDATES into PANEL, which is what
makes it writable -- 01 iterates that dict alone. Authority item 111933,
created on the archive today; AI_MODEL_ITEMS is keyed on the self-hosted route
because that is what produced the annotations, and the OpenRouter twin gets its
own item the day it writes anything.

It joins rather than replaces. It shares no lab or pretraining family with the
other four, so it does not carry the correlated-error problem that kept Gemma
from being added beside Gemini, and it is the first member whose requested
'medium' is a rung the model has rather than one it is rounded up to. Four of
the five are now open-weights, and this one goes furthest: its answers were
produced on university hardware, so the run reproduces without a third party
consenting to serve the model.

Its coverage is 12,098 of 12,251 and is expected to stay there. The 153 that
failed four times each are documented in PANEL as a gap not to repair, because
the obvious reading of a short member -- a failed run, re-run it -- is wrong
here and would cost a day.

The OpenRouter twin stays a candidate. test_the_two_qwen_routes_stay_on_opposite
_sides_of_the_panel now pins that arrangement in both directions: same weights,
different routes, and only one of them writable.

Also fix a premise 00_setup_properties.py had backwards. Its emit_ttl docstring
warned that Omeka deletes any installed property the uploaded file omits, along
with every value under it. It does not -- a vocabulary update applies additions
only, so the archive holds a superset of the file (80 properties against 38) and
an omission destroys nothing. The --verify output said "Would be DELETED" and
counted deletions; it now reports what the file omits and states plainly that
omitted is not deleted. The check itself still runs before every upload: it
answers whether anything omitted still holds values, which is the question that
matters if the properties are ever pruned by hand, and the one that would catch
a member dropped from PANEL by mistake.

Author: fmadore

AI_sentiment_analysis/00_setup_properties.py

@@ -228,13 +228,20 @@ def _ttl_block(definition: PropertyDef) -> str:
 def emit_ttl() -> str:
     """The Turtle to append to ``iwac-vocabulary.ttl``.
 
-    Emits the active panel, and only it. **Omeka deletes any installed property
-    the uploaded file omits, along with every value stored under it**, so this
-    function decides what survives the next vocabulary upload: a member dropped
-    from ``PANEL`` while still holding annotations would have them destroyed
-    silently. Every retired member has now been emptied deliberately, so there
-    is nothing left for an omission to take — but check that before removing the
-    next one, not after.
+    Emits the active panel, and only it — so a member dropped from ``PANEL``
+    disappears from the file, and the file stops describing the live vocabulary
+    rather than pruning it. **Omeka applies additions only**: a vocabulary
+    update never removes an installed property, so the archive holds a superset
+    of this file (80 properties against 38, measured 2026-08-25) and omitting a
+    member destroys nothing.
+
+    That is worth stating precisely because the opposite was believed here for
+    months, and it is the more dangerous belief to hold backwards in either
+    direction. Fearing a deletion that cannot happen means keeping retired
+    members in the file forever; assuming one that *could* happen means an
+    omission silently taking values with it. ``--verify`` resolves it by
+    measurement rather than by either assumption: it counts what every omitted
+    property holds and refuses the upload if any of them holds anything.
     """
     out: List[str] = [
         "# ============================================",
@@ -333,17 +340,17 @@ def verify_ttl_is_superset(client: OmekaClient, ttl_path: Path) -> bool:
     table.add_column("Count", justify="right")
     table.add_row("Installed in Omeka", str(len(installed_names)))
     table.add_row("Declared in the .ttl", str(len(declared)))
-    table.add_row("Would be added", f"[green]{len(would_add)}[/]")
+    table.add_row("Will be added", f"[green]{len(would_add)}[/]")
     table.add_row(
-        "Would be DELETED",
-        f"[bold red]{len(would_delete)}[/]" if would_delete else "[dim]0[/]",
+        "Omitted by the .ttl",
+        f"[bold]{len(would_delete)}[/]" if would_delete else "[dim]0[/]",
     )
     console.print(table)
 
     populated: Dict[str, int] = {}
     if would_delete:
         console.print("\n[yellow]The .ttl omits properties that are installed. "
-                      "Checking what would be lost:[/]")
+                      "Checking what each holds:[/]")
         for name in would_delete:
             count = count_values(client, installed_names[name])
             if count:
@@ -353,7 +360,7 @@ def verify_ttl_is_superset(client: OmekaClient, ttl_path: Path) -> bool:
 
     if populated:
         console.print(Panel(
-            "These carry values that the upload would destroy:\n\n  "
+            "These are omitted by the .ttl and still carry values:\n\n  "
             + "\n  ".join(f"{VOCABULARY_PREFIX}:{n} — {c:,} item(s)"
                           for n, c in sorted(populated.items()))
             + "\n\nDo not upload. Add them back to the .ttl, or export the "
@@ -364,13 +371,24 @@ def verify_ttl_is_superset(client: OmekaClient, ttl_path: Path) -> bool:
 
     summary = f"{len(would_add)} additions"
     if would_delete:
-        summary += f", {len(would_delete)} deletions — all verified empty"
-    else:
-        summary += ", 0 deletions"
+        summary += f", {len(would_delete)} omitted — all verified empty"
     console.print(f"\n[green]✓[/] Safe to upload — {summary}.")
     console.print("[dim]  Admin → Vocabularies → IWAC Ontology → Update. "
-                  "Omeka shows the same diff before committing; if its list "
-                  "differs from the one above, stop.[/]")
+                  "Omeka shows its own diff before committing; if its list of "
+                  "additions differs from the one above, stop.[/]")
+    if would_delete:
+        console.print(Panel(
+            f"The {len(would_delete)} omitted properties will stay installed. "
+            "Omeka does not remove a vocabulary property on update — it applies "
+            "additions only — so the live vocabulary is a superset of this file "
+            "and stays one.\n\n"
+            "The check above still has to run, because it is the only thing "
+            "standing between an omission and a silent loss if that behaviour "
+            "ever changes, or if the properties are removed by hand later. It "
+            "answers \"is anything omitted still holding values\", not \"will "
+            "Omeka delete these\" — the answer to the second is no.",
+            title="Omitted ≠ deleted", border_style="dim",
+        ))
     return True
 
 

AI_sentiment_analysis/sentiment_core.py

@@ -243,11 +243,15 @@ def probe_term(self) -> str:
 #: for the model, never for its vendor: reusing a vendor slot is what made
 #: generation 1 impossible to attribute without a git archaeology session.
 #:
-#: Three of the four are open-weights releases (Mistral Small 4 and Gemma 4
-#: under Apache-2.0, DeepSeek V4 Flash under MIT), so those annotations can be
-#: regenerated from weights that are archivable alongside them. Their active
-#: parameter counts — 6.5B, 13B and 31B (Gemma is dense, so all of it) — sit
-#: inside a factor of five, wider than the pair that preceded them.
+#: Four of the five are open-weights releases (Mistral Small 4, Gemma 4 and
+#: Qwen3.8 under Apache-2.0, DeepSeek V4 Flash under MIT), so those annotations
+#: can be regenerated from weights that are archivable alongside them. Their
+#: active parameter counts — 6.5B, 13B, 27B and 31B (Gemma is dense, so all of
+#: it) — sit inside a factor of five, wider than the pair that preceded them.
+#: Only GPT-5.6 Luna is a model that could be withdrawn out from under its own
+#: annotations, and Qwen3.8 goes further than the other three: its answers were
+#: produced on hardware in the building, so the run is reproducible without a
+#: third party consenting to serve the weights at all.
 #:
 #: Every member is its vendor's high-volume tier. That is the property that
 #: makes this a panel rather than a quality ladder, and it is why the Gemini
@@ -310,9 +314,38 @@ def probe_term(self) -> str:
                     "mistralSmall2603"),
         PanelMember("deepseek_v4_flash_0731", "deepseek-v4-flash-0731",
                     "DeepSeek V4 Flash 0731", "deepseekV4Flash0731"),
+        PanelMember("qwen3_8_27b", "qwen3.8-27b-selfhosted",
+                    "Qwen3.8 27B (self-hosted)", "qwen3827b"),
     )
 }
 
+#: **Qwen3.8 27B joined as a fifth voice on 2026-08-25**, annotated on
+#: university hardware rather than through anyone's API — the first member whose
+#: answers cost queue time instead of tokens, and the first whose requested
+#: ``medium`` is a rung the model actually has rather than one it is rounded up
+#: to (see :data:`PANEL_REASONING`). It joins rather than replaces: it shares no
+#: lab or pretraining family with the other four, so it does not carry the
+#: correlated-error problem that kept Gemma from being added beside Gemini.
+#:
+#: Its registry key names the *route*, not just the weights. The OpenRouter twin
+#: stays in :data:`PILOT_CANDIDATES` and must not be promoted alongside it: the
+#: same weights reached two ways would look like two annotators in the panel
+#: while being one reading of the construct, which is the correlated-error
+#: failure again in its purest form.
+#:
+#: **Its coverage is 12,098 of 12,251 and is expected to stay there.** 153
+#: articles were attempted four times each — a full-corpus pass plus three retry
+#: rounds — and retired. 145 of them fail the schema's cross-field rule the same
+#: way every time, and the failures concentrate on low centrality (5.45% of
+#: ``Marginal`` against 0.00% of ``Non abordé``): the model declines subjectivité
+#: when Islam is *peripheral* to an article where the prompt licenses declining
+#: only when it is *absent*. That is a disagreement about the instrument, so it
+#: is recorded rather than repaired — ``serving/merge_shards.py`` writes the
+#: failure log, and the README explains why relaxing the validator would be
+#: worse than the gap. **Do not "fix" this member's shortfall by re-running it.**
+#: The consequence for analysis is that its missing subjectivité is not missing
+#: at random.
+
 #: Models under evaluation, which ``02_pilot_new_panel.py`` runs and nothing
 #: else does. Membership of :data:`PANEL` is what makes a model *writable* —
 #: ``01_sentiment_analysis.py`` iterates that dict alone — so a candidate parked
@@ -331,28 +364,32 @@ def probe_term(self) -> str:
 #: comes first, then the code that cites it (see ``AI_MODEL_ITEMS``). Nothing in
 #: the pilot path asks for it; only the write path does.
 #:
-#: **Qwen3.8 27B appears twice, once per route.** The property prefix names the
-#: model and the registry key carries the route, exactly as ``gemma_4_31b_it``
-#: does for ``gemma-4-openrouter``. Running both on one sample is what turns
-#: "self-hosting is cheaper" into a measurement: same weights, same articles,
-#: with latency, structured-output reliability and reasoning depth read off each
-#: route rather than assumed to match. Cost is not comparable in one unit —
-#: OpenRouter bills tokens, a cluster bills GPU-hours and queue time — so record
-#: both rather than converting one into the other.
+#: **The remaining candidate is Qwen3.8 27B's OpenRouter twin, and it stays a
+#: candidate even though the same weights are now in :data:`PANEL`.** The
+#: property prefix names the model and the registry key carries the route,
+#: exactly as ``gemma_4_31b_it`` does for ``gemma-4-openrouter``. Running both on
+#: one sample is what turns "self-hosting is cheaper" into a measurement: same
+#: weights, same articles, with latency, structured-output reliability and
+#: reasoning depth read off each route rather than assumed to match. Cost is not
+#: comparable in one unit — OpenRouter bills tokens, a cluster bills GPU-hours
+#: and queue time — so record both rather than converting one into the other.
+#:
+#: What it must never become is a fifth-and-a-half voice. Promoting it beside the
+#: self-hosted member would put one model in the panel twice, and a panel that
+#: counts one reading as two is measuring its own routing rather than the
+#: construct.
 #:
-#: Neither needs a :data:`PANEL_REASONING_OVERRIDES` entry. Qwen3.8's ladder is
+#: It needs no :data:`PANEL_REASONING_OVERRIDES` entry. Qwen3.8's ladder is
 #: low/medium/xhigh, so the panel's requested ``medium`` is a rung the model
-#: actually has: it would be the first member since GPT-5.6 Luna to sit at the
-#: requested depth instead of being rounded up to it. Whether that survives the
-#: OpenRouter route is precisely what the twin is there to find out — Gemma's
+#: actually has — verified on the self-hosted route, where reasoning length grew
+#: ~3.5× from ``low`` to ``xhigh`` with the middle rung cleanly between. Whether
+#: that survives the OpenRouter route is precisely what this twin is for: Gemma's
 #: graduated levels collapsed to on/off when fanned across third-party backends
-#: (see :data:`PANEL_REASONING`), and a self-hosted server is the only route here
-#: where the depth can be read off the server's own logs.
+#: (see :data:`PANEL_REASONING`), and a server you run yourself is the only route
+#: here where the depth can be read off the server's own logs.
 PILOT_CANDIDATES: Dict[str, PanelMember] = {
     m.key: m
     for m in (
-        PanelMember("qwen3_8_27b", "qwen3.8-27b-selfhosted",
-                    "Qwen3.8 27B (self-hosted)", "qwen3827b"),
         PanelMember("qwen3_8_27b_openrouter", "qwen3.8-27b-openrouter",
                     "Qwen3.8 27B (OpenRouter)", "qwen3827bOr"),
     )

common/iwac_config.py

@@ -147,6 +147,13 @@
     "mistral-small": {"item_id": 79614, "display_title": "Mistral Small 4"},
     "qwen3.5-moe": {"item_id": 79616, "display_title": "Qwen3.5 122B-A10B"},
     "qwen3.5-moe-small": {"item_id": 79612, "display_title": "Qwen3.5 35B-A3B"},
+    # Keyed on the self-hosted route, which is the one that produced the
+    # annotations: they were generated on university hardware, and the OpenRouter
+    # twin (``qwen3.8-27b-openrouter``) is deliberately absent because it has
+    # written nothing. The route is half of what the provenance record claims,
+    # so the day the twin annotates anything it needs its own authority item
+    # rather than a share of this one. Created 2026-08-25.
+    "qwen3.8-27b-selfhosted": {"item_id": 111933, "display_title": "Qwen3.8 27B"},
     "deepseek-v4-flash-0731": {
         "item_id": 83261,
         "display_title": "DeepSeek V4 Flash 0731",

tests/test_sentiment_panel.py

@@ -191,15 +191,20 @@ def test_pilot_candidate_terms_collide_with_nothing():
             seen[term] = member.key
 
 
-def test_pilot_candidates_name_the_model_and_the_route():
+def test_the_two_qwen_routes_stay_on_opposite_sides_of_the_panel():
     """Two routes to one set of weights, told apart by registry key.
 
     The property prefix names the model and the registry key names the route —
     the split ``gemma_4_31b_it`` → ``gemma-4-openrouter`` already makes. Running
     both is what turns "self-hosting is cheaper" into something measured on one
     sample rather than assumed.
+
+    Since 2026-08-25 the self-hosted route is a panel member and the OpenRouter
+    twin is still a candidate, which is the arrangement this pins. Promoting the
+    twin as well would put one model in the panel twice — a panel that counts
+    one reading of the construct as two is measuring its own routing.
     """
-    selfhosted = PILOT_CANDIDATES["qwen3_8_27b"]
+    selfhosted = PANEL["qwen3_8_27b"]
     hosted = PILOT_CANDIDATES["qwen3_8_27b_openrouter"]
 
     assert selfhosted.registry_key == "qwen3.8-27b-selfhosted"
@@ -209,6 +214,9 @@ def test_pilot_candidates_name_the_model_and_the_route():
     # served model is the same one.
     assert MODEL_REGISTRY[hosted.registry_key].model.lower() \
         == MODEL_REGISTRY[selfhosted.registry_key].model.lower()
+    # ...which is exactly why only one of them may be writable.
+    assert "qwen3_8_27b_openrouter" not in PANEL
+    assert selfhosted.property_prefix != hosted.property_prefix
 
 
 def test_the_pilot_runs_candidates_alongside_the_live_panel():
@@ -442,7 +450,7 @@ def test_cache_round_trips(tmp_path, good_result):
 
 
 def test_cache_is_granular_to_the_model(tmp_path, good_result):
-    """One model failing must not force re-running the other four."""
+    """One model failing must not force re-running the rest of the panel."""
     path = tmp_path / "c.jsonl"
     with SentimentCache(path=path) as cache:
         for key in ("gemma_4_31b_it", "gpt_5_6_luna"):
@@ -451,7 +459,7 @@ def test_cache_is_granular_to_the_model(tmp_path, good_result):
     reloaded = SentimentCache(path=path)
     reloaded.load()
     assert reloaded.missing_models(1, PANEL) == [
-        "mistral_small_2603", "deepseek_v4_flash_0731",
+        "mistral_small_2603", "deepseek_v4_flash_0731", "qwen3_8_27b",
     ]