Document distractor discrimination; write the two supplement captions

MANIFEST.json

@@ -46,12 +46,12 @@
   },
   {
     "path": "paper-supplement/tab-by-operator-removed-2026-08-20.tex",
-    "bytes": 1505,
+    "bytes": 1509,
     "why": "per-operator CCR table (cut for space)"
   },
   {
     "path": "paper-supplement/tab-rank-metrics-removed-2026-08-20.tex",
-    "bytes": 837,
+    "bytes": 1953,
     "why": "rank-aware metrics table (cut for space)"
   },
   {
@@ -146,7 +146,7 @@
   },
   {
     "path": "v1.0.1-addendum.md",
-    "bytes": 9334,
+    "bytes": 10901,
     "why": "known issues + evaluation protocol; the interval-leak fix the paper cites"
   }
 ]

paper-supplement/tab-by-operator-removed-2026-08-20.tex

@@ -1,12 +1,22 @@
 % tab:by-operator, cut from the CIKM camera-ready 2026-08-20 for the page
 % limit. Summarised by one sentence in the retrieval-quality paragraph.
-% Belongs in the dataset card. Preserved verbatim.
+% Numbers verbatim as cut. The caption's causal clause was corrected for this
+% release on 2026-08-22: as the cut version's own TODO recorded, attributing
+% interval's stale-blind resilience solely to the same-(s,r,o) test was a real
+% but secondary effect, and the primary driver is the answer leakage now
+% disclosed in the paper.
 
 \begin{table}[t]
 \centering
 \footnotesize
-% CAMERA-READY-TODO(prose): this caption's attribution of interval's stale-blind resilience ("the metric's same-(s,r,o) test admits stale variants") is a real but SECONDARY effect. The primary driver is the interval answer leakage: t_query == gold answer year on 864/864 = 100.0% of interval questions, which is also why interval tops every retrieval slice (TRP 0.376, CCR 0.920, stale-blind CCR 0.773 vs 0.03-0.08 on all other operators). Rewrite the caption's causal clause when the leakage subsection lands; do not leave the current attribution as the sole explanation.
-\caption{Per-operator CCR breakdown on the \tempbench{} test split. Stale-blind collapses where the temporal filter is decisive (PIT/BA/sequence) but holds on \emph{interval} because the metric's same-$(s,r,o)$ test admits stale variants as covering an interval gold (operator property, not retriever skill).}
+\caption{Per-operator CCR breakdown on the \tempbench{} test split. Stale-blind
+collapses where the temporal filter is decisive (point-in-time, before/after,
+sequence) but holds at $0.77$ on \emph{interval}. That is chiefly answer
+leakage, not retriever skill: $t_{\mathrm{query}}$ equals the gold answer year
+on $864/864$ interval questions, which is also why interval tops every
+retrieval slice (TRP $0.376$, CCR $0.920$). A secondary effect compounds it ---
+the metric's same-$(s,r,o)$ test admits stale variants as covering an interval
+gold. Read no capability result off the interval row.}
 \label{tab:by-operator}
 \begin{tabular}{lrrr}
 \toprule

paper-supplement/tab-rank-metrics-removed-2026-08-20.tex

@@ -1,11 +1,25 @@
 % tab:rank-metrics, offloaded from the CIKM camera-ready 2026-08-20 for the
 % page limit. Its findings are carried by one sentence in the results text.
-% Belongs in the dataset card. Preserved verbatim.
+% Numbers verbatim as cut; the caption was written for this release on
+% 2026-08-22, since the table shipped with an unfilled placeholder.
 
 \begin{table}[t]
 \centering
 \footnotesize
-\caption{[CAMERA-READY-TODO(prose): caption]}
+\caption{Rank-aware retrieval metrics on the \tempbench{} test split, against
+the subgraph metrics they are meant to contextualise. recall@$k$ and NDCG@$k$
+use binary relevance (retrieved triple in $S^{*}$ \emph{and} valid at $\tq$),
+macro-averaged, oracle anchor. The gold chain is nearly always retrieved and
+near the top of the ranking (recall $0.921$, NDCG $0.886$), so TRP's low
+absolute value reflects $1$--$3$-triple gold chains scored against a $k$-sized
+list, not poor retrieval. TRP is \emph{not} $k$-dominated for the temporal
+retriever --- its evidence list averages $9.09$ triples at $k{=}15$, so the
+candidate pool binds rather than $k$, and a $10\times$ sweep of $k$ moves TRP
+only $0.410\!\to\!0.303$. The $k$-dominated regime is the stale-blind
+ablation, whose pool saturates $k$ and whose TRP falls fourfold over the same
+sweep. NDCG tracks recall closely here ($0.886$ vs $0.921$); it adds
+interpretability, not discriminative power. These values are not comparable to
+TimelineKGQA's MRR/Hits@K, which score a different candidate universe.}
 \label{tab:rank-metrics}
 \begin{tabular}{lrrrrr}
 \toprule

v1.0.1-addendum.md

@@ -85,6 +85,33 @@ with the release; stale-dependent evaluation should be restricted to the
 functional subset (1,429/1,743 test questions), and the interval × stale cell
 (n=12 functional) is too small to support slice-level claims.
 
+### 2b. Distractor negatives: lexically hard, temporally easy
+
+`S_dist` swaps the object of one gold hop, keeping `(s, r)`. The two axes come
+apart, and both matter when reporting on this benchmark.
+
+- **Lexical ranking cannot separate them.** On the 1,237 functional distractor
+  pairs in the test split, BM25 over `label(s) label(r) label(o)` scores gold
+  above the released distractor in 489 cases (39.5%), *below* it in 472
+  (38.2%), and exactly tied in 276 (22.3%). Excluding ties, gold wins 489 of
+  961 — **50.9%, which is chance.** A purely lexical retriever has no signal
+  here at all.
+- **A temporal filter removes most of them anyway.** 96.4% of released
+  distractors are invalid at `t_query` and are dropped by the per-hop filter;
+  only 3.6% survive it. This is a **side effect of construction** —
+  `_build_distractor` relaxes the `valid_at` constraint when selecting the
+  substitute object — not a designed temporal challenge. For the
+  temporally-filtered reference retriever, typed distractor negative-hit-rate
+  is 0.138.
+
+**Intended use.** Distractors are the residual difficulty for a time-aware
+system, and the honest framing is narrow: they show that answer-string identity
+is not evidence identity, and they defeat lexical retrieval outright. Do not
+describe them as temporally hard — the temporal filter disposes of 96.4% of
+them — and do not read the 3.6% that survive as a designed property. As with
+stale, restrict distractor-dependent evaluation to the functional subset using
+the shipped flags.
+
 ---
 
 ### 3. Reference retriever: full specification