Design: online (streaming) reasoning compression

examples/cot_compression/README.md

@@ -14,6 +14,11 @@ Both `/v1/chat/completions` and `/v1/completions` are intercepted; everything
 else is proxied to vLLM unchanged, so an eval harness can point at the sidecar
 without modification.
 
+> **Related, not yet implemented:** compressing *during* generation rather than
+> after it — checkpoint every N tokens and keep reasoning — is designed in
+> [`DESIGN-online-compression.md`](DESIGN-online-compression.md). That variant
+> is a separate module; this one is unchanged.
+
 ### /v1/completions
 
 Same pipeline over a raw prompt. Accepts a string, a token-ID list, or a batch of

examples/cot_compression/before_after.py

@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+"""Pair the online arm's compressed trace against the uncompressed baseline.
+
+The sidecar discards each raw chunk once it has been summarised, so a
+chunk-aligned "before" is not recoverable from a run. What IS comparable is the
+same PROBLEM under both arms: the baseline's full trace versus the online arm's
+accumulated notes plus final verbatim chunk.
+
+  usage: before_after.py <online_output.jsonl> <baseline_output.jsonl> [n]
+"""
+import json
+import re
+import sys
+
+ON, BASE = sys.argv[1], sys.argv[2]
+N = int(sys.argv[3]) if len(sys.argv) > 3 else 6
+
+online, rows = {}, 0
+for line in open(ON):
+    line = line.strip()
+    if not line:
+        continue
+    try:
+        r = json.loads(line)
+    except Exception:
+        continue
+    rows += 1
+    rc = r.get("reasoning_content") or ""
+    if "Notes from my earlier work" in rc:
+        online[r.get("id")] = {
+            "rc": rc, "gen": r.get("generation") or "",
+            "ntok": r.get("num_generated_tokens"),
+            "fr": r.get("finish_reason"),
+            "correct": r.get("symbolic_correct"),
+            "notes": len(re.findall(r"^\[\d+\] ", rc, re.M)),
+            "problem": r.get("problem") or "",
+        }
+
+print(f"online rows so far : {rows}")
+print(f"multi-chunk (compressed) : {len(online)}")
+if not online:
+    sys.exit(0)
+
+ids = sorted(online, key=lambda k: -online[k]["notes"])[:60]
+want = set(ids)
+base = {}
+for line in open(BASE):
+    if not line.strip():
+        continue
+    r = json.loads(line)
+    if r.get("id") in want:
+        base[r["id"]] = {"rc": r.get("reasoning_content") or "",
+                         "gen": r.get("generation") or "",
+                         "ntok": r.get("num_generated_tokens"),
+                         "correct": r.get("symbolic_correct")}
+        if len(base) == len(want):
+            break
+
+pairs = [{"id": i, "on": online[i], "bs": base[i]} for i in ids if i in base]
+print(f"matched to baseline : {len(pairs)}\n")
+
+hdr = f"{'id':26} {'notes':>5} {'before':>10} {'after':>9} {'ratio':>7}  base/online"
+print(hdr)
+print("-" * len(hdr))
+for e in pairs[:20]:
+    o, b = e["on"], e["bs"]
+    ratio = len(b["rc"]) / max(len(o["rc"]), 1)
+    print(f"{e['id']:26} {o['notes']:5d} {len(b['rc']):9,}c {len(o['rc']):8,}c "
+          f"{ratio:6.1f}x  {b['correct']}/{o['correct']}")
+
+json.dump(pairs[:N], open("/tmp/ba_out.json", "w"))
+print(f"\nwrote {min(N, len(pairs))} full pairs to /tmp/ba_out.json")