Re-render the 1,085 single-turn longcontext rows as four messages (passage + ack, filler, preamble + question, completion)

Only messages_json changes, only on longcontext rows with one user turn (v3_bookqa, lc3flip, lc5flip, v5a_onpolicy, v5c3_intrace, v5c5_novels). Every other row and column is byte-identical to 1596f182. Variants are drawn by sha256 of the original messages_json. Provenance and builder under provenance/. Agreed with Oskar 2026-10-02.

README.md

@@ -237,12 +237,29 @@ Under this rule **79,580 of 163,016 rows (48.8%)** are contextual:
 | wikihow | 1,760 | 0 | 0 | 0 |
 | wikihownest | 7,960 | 0 | 1,882 | 1,882 |
 
+## Longcontext single-turn rows re-rendered as four messages (2026-10-02)
+
+The 1,085 `longcontext` rows that held the passage and the question in one user turn (sources
+`v3_bookqa`, `lc3flip`, `lc5flip`, `v5a_onpolicy`, `v5c3_intrace`, `v5c5_novels`) now carry four messages:
+the passage followed by a short request for an acknowledgement, a one-line assistant acknowledgement, the
+original question with a short preamble, and the original completion. The split point is the last
+horizontal-rule line (`---` or `***`) of the original user turn, so the passage and the question text are
+unchanged; the acknowledgement request, the acknowledgement and the preamble are drawn from small fixed pools
+by a hash of the original `messages_json`. Only `messages_json` changed, only on these rows; every other row
+and every other column is byte-identical to revision `1596f18267bc45e14f9a08d0f9949015ce0b1d2b`. The 679
+longcontext rows that already asked their question in a later turn (`lcflip`, `v2`, `v5b_onpolicy`,
+`v4_offpolicy_concat`) are untouched. Reason: a probe read window that starts at the last user turn no longer
+spans the whole passage (these rows held 44 percent of the window tokens of the corpus).
+`provenance/longcontext_single_turn_split.json` lists every rewritten row with the sha256 of its original
+`messages_json`, the variant pools and the builder's sha256.
+
 ## Provenance
 
 `provenance/BUILD.json` records the parent revision and parquet sha256s, the seed (20260929),
 the family cap (8,000), the per-family before/after counts and the builder's sha256;
 `provenance/new_row_screen.json` lists every screened new row with its judge and Fable outcome;
-`provenance/covariate_trim.json` holds the per-family covariate trim logs. The builder is
+`provenance/covariate_trim.json` holds the per-family covariate trim logs;
+`provenance/longcontext_single_turn_split.json` records the 2026-10-02 longcontext re-rendering. The builder is
 `sandbox/oa_backdoor/scripts/dataset_generation/build_v2_slotin.py` in the Deception Pod monorepo.
 
 ## Columns and loading

data/deceptive_deceptive-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3b1c3c0e50c24632c87e7c1443a0280c508464428f4b931bd655318d5d57e8da
-size 173018309
+oid sha256:44e90fe947b16bf9729b831cdf394c849f901bc0c5fca6d95477ef2966f8661e
+size 173034028

data/deceptive_honest-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b80b3616ca2ad9de176c1d094c4b4324ecd5ebe8a2583fe647210ff8dcc258c2
-size 168555904
+oid sha256:4330b8ec9859428d7be1f7a6065d03c0b111c2e492fed2e5252743d7671df68e
+size 168611316

data/honest_deceptive-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8f27d8ef9cc7f2ba4fa446c457a0c737b06b59698dfc853fb090c09bda52acc6
-size 216786956
+oid sha256:257531ac79ac9db1416dda1ff115848017aa136dce928cd184027d8324dcf912
+size 216661429

data/honest_honest-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:555563638e48bb8c30ccffee4d7074e549538e4293aa1a592123dfb5310522ed
-size 194030526
+oid sha256:a805f8cc5e1a0f8cc1ae6ba090ee6c79e2596cf58c3cbdf8725d8d5f8dc0e8ae
+size 194052652

provenance/build_longcontext_single_turn_split.py

@@ -0,0 +1,68 @@
+"""Build the FIBS-v1 revision that re-renders the 1,085 single-turn longcontext rows as four messages.
+
+Parent: AlignmentResearch/fibs-v1 @ 1596f18267bc45e14f9a08d0f9949015ce0b1d2b (local byte-identical copy under SRC).
+Only `messages_json` changes, only on rows of family longcontext whose message list has exactly one user turn and whose
+user turn ends in a horizontal-rule separator followed by the question. Every other row and column is copied unchanged.
+Variant choice is keyed on sha256(original messages_json), so the build is reproducible from the parent alone.
+"""
+import hashlib, json, re, sys, os, datetime
+import pyarrow as pa, pyarrow.parquet as pq, pyarrow.compute as pc
+pa.set_cpu_count(1); pa.set_io_thread_count(1)
+SRC = "/home/dev/persistent/SCRATCH/rm_scaling_whitebox/q39_fibs_v1/source"
+OUT = sys.argv[1]
+SEP = re.compile(r"\n[ \t]*(?:-{3,}|\*{3,})[ \t]*\n?")
+FILLERS = ["Acknowledged.", "Understood.", "Got it.", "Noted.", "Okay, I've read it.", "I've read the passage.",
+           "Read and noted.", "Understood, I have the text.", "Okay.", "Got it, go ahead.", "Noted. Ready for your question.",
+           "I have it.", "Acknowledged, ready when you are.", "Understood. What would you like to know?", "Read it.",
+           "Okay, I've gone through it.", "Noted, thanks.", "All read.", "Received and read.", "Understood, ask away."]
+REQUESTS = ["Reply with a brief acknowledgement only; I'll ask my question next.",
+            "Just confirm you've read this, nothing else. My question follows.",
+            "Please acknowledge in one short line and wait for my question.",
+            "No response needed beyond a short acknowledgement. Question coming next.",
+            "Acknowledge briefly that you have read the above; I will ask about it in a moment."]
+PREAMBLES = ["Now, about what you just read:", "Here is my question about the passage:", "Based on the text above,",
+             "Referring to what I showed you:", "My question:", "About the passage you just read:",
+             "Okay. From the text above,", "Thanks. Now,", "Right, here's the question:", ""]
+SINGLE_TURN_SOURCES = {"longcontext_v3_bookqa", "longcontext_lc3flip", "longcontext_lc5flip", "longcontext_v5a_onpolicy",
+                       "longcontext_v5c3_intrace", "longcontext_v5c5_novels"}
+def pick(seq, key, salt): return seq[int(hashlib.sha256(f"{salt}:{key}".encode()).hexdigest(), 16) % len(seq)]
+def rewrite(mj, source):
+    msgs = json.loads(mj)
+    users = [m for m in msgs if m["role"] == "user"]
+    if source not in SINGLE_TURN_SOURCES or len(users) != 1: return None, "not single-turn"
+    u = users[0]["content"]; hits = list(SEP.finditer(u))
+    if not hits: return None, "no separator"
+    m = hits[-1]; passage, question = u[:m.start()].rstrip(), u[m.end():].strip()
+    if not question or len(question) > 800: return None, "question length"
+    key = hashlib.sha256(mj.encode()).hexdigest()
+    final_user = question if "\n" in question else (pick(PREAMBLES, key, "pre") + " " + question).strip()
+    new = [m for m in msgs if m["role"] == "system"] + [
+        {"role": "user", "content": passage + "\n\n" + pick(REQUESTS, key, "req")},
+        {"role": "assistant", "content": pick(FILLERS, key, "fill")},
+        {"role": "user", "content": final_user}]
+    assert [m["role"] for m in msgs if m["role"] not in ("system", "user")] == [], "unexpected roles in single-turn row"
+    return json.dumps(new, ensure_ascii=False), key
+os.makedirs(f"{OUT}/data", exist_ok=True); os.makedirs(f"{OUT}/provenance", exist_ok=True)
+record = {"parent_repo": "AlignmentResearch/fibs-v1", "parent_revision": "1596f18267bc45e14f9a08d0f9949015ce0b1d2b", "parent_parquet_sha256": {},
+          "rule": "family longcontext, one user turn, split at the last horizontal-rule line; messages_json only", "sources": sorted(SINGLE_TURN_SOURCES),
+          "fillers": FILLERS, "requests": REQUESTS, "preambles": PREAMBLES, "variant_key": "sha256(original messages_json), salted 'req'/'fill'/'pre'",
+          "rows": [], "skipped": [], "builder_sha256": hashlib.sha256(open(__file__, "rb").read()).hexdigest(), "created_at": datetime.datetime.now(datetime.timezone.utc).isoformat()}
+for quad in ("deceptive_deceptive", "deceptive_honest", "honest_deceptive", "honest_honest"):
+    fn = f"data/{quad}-00000-of-00001.parquet"; record["parent_parquet_sha256"][quad] = hashlib.sha256(open(f"{SRC}/{fn}", "rb").read()).hexdigest()
+    t = pq.read_table(f"{SRC}/{fn}")
+    mj_col = t.column("messages_json").to_pylist(); fam = pc.equal(t.column("deception_type"), "longcontext")
+    src_col = t.column("source").to_pylist(); n_changed = 0; new_col = list(mj_col)
+    # family lives in `deception_type`? check quickly against source prefix; the census used `family` from source
+    for i, (mj, src) in enumerate(zip(mj_col, src_col)):
+        if not str(src).startswith("longcontext"): continue
+        new, info = rewrite(mj, src)
+        if new is None: record["skipped"].append({"quadrant": quad, "row": i, "source": src, "why": info}); continue
+        new_col[i] = new; n_changed += 1; record["rows"].append({"quadrant": quad, "row": i, "source": src, "orig_messages_sha256": info})
+    idx = t.schema.get_field_index("messages_json")
+    t2 = t.set_column(idx, pa.field("messages_json", pa.string()), pa.array(new_col, type=pa.string()))
+    t2 = t2.replace_schema_metadata(t.schema.metadata)
+    pq.write_table(t2, f"{OUT}/{fn}", compression="snappy", row_group_size=1000)
+    print(f"{quad}: rows {t.num_rows}, rewritten {n_changed}")
+record["n_rewritten"] = len(record["rows"]); record["n_skipped"] = len(record["skipped"])
+json.dump(record, open(f"{OUT}/provenance/longcontext_single_turn_split.json", "w"), indent=1, ensure_ascii=False)
+print("total rewritten", record["n_rewritten"], "skipped", record["n_skipped"], {s["why"] for s in record["skipped"]})