ForesightFlow · MaksymDS · Apr 27, 2026 · Apr 27, 2026
diff --git a/config/validation_markets.yaml b/config/validation_markets.yaml
@@ -22,4 +22,15 @@
 #     known_case: true
 #     notes: "Evidence URL from UMA proposer likely points to Reuters. T_news ~Feb 28 2026."
 
-markets: []
+markets:
+  - market_id: "0xfa1543cdef36d55ef9126aaab6015c7c7ed5aa6a2bb5be355f5cacc2302c7374"
+    label: "Epstein files — Ehud Barak"
+    category: military_geopolitics
+    known_case: false
+    notes: >
+      Epstein files released in two waves: Dec 18 2025 (House Oversight Committee
+      released 68 estate photos including Barak photo — first public mention) and
+      Dec 19 2025 (DOJ main release per Epstein Files Transparency Act deadline).
+      Barak photo confirmed by CNN, Al Jazeera, Times of Israel all dated 2025-12-18.
+      Market resolved YES Dec 23 2025 when additional DOJ docs (11K+) confirmed mention.
+      T_news candidate: 2025-12-18T18:00:00Z (afternoon EST photo release).
diff --git a/fflow/cli.py b/fflow/cli.py
@@ -505,11 +505,19 @@ async def _run() -> None:
 
 @news_app.command("tier3")
 def news_tier3(
-    market: Annotated[str, typer.Option(help="Market condition ID (0x...)")],
+    market: Annotated[str | None, typer.Option(help="Market condition ID (0x...)")] = None,
+    validation_set: Annotated[bool, typer.Option("--validation-set", help="Process all markets in config/validation_markets.yaml")] = False,
     confirm: Annotated[bool, typer.Option("--confirm", help="Acknowledge LLM API cost")] = False,
     dry_run: Annotated[bool, typer.Option("--dry-run")] = False,
+    max_cost: Annotated[float, typer.Option("--max-cost", help="Hard cost cap in USD (approximate)")] = 5.0,
 ) -> None:
-    """Tier 3: use Claude LLM to extract T_news. Requires --confirm."""
+    """Tier 3: use Claude LLM to extract T_news. Requires --confirm.
+
+    Run on a single market (--market 0x...) or on all entries in
+    config/validation_markets.yaml (--validation-set).
+    """
+    import pathlib
+    import yaml
     from fflow.db import AsyncSessionLocal
     from fflow.models import Market, NewsTimestamp
     from fflow.news.llm_match import llm_extract_date
@@ -518,48 +526,80 @@ def news_tier3(
         typer.echo("Pass --confirm to acknowledge LLM API cost (~$0.01-0.05 per call).")
         raise typer.Exit(1)
 
-    async def _run() -> None:
-        async with AsyncSessionLocal() as session:
-            mkt = await session.get(Market, market)
-            if mkt is None:
-                typer.echo(f"Market not found: {market}", err=True)
-                raise typer.Exit(1)
+    if not market and not validation_set:
+        typer.echo("Provide --market 0x... or --validation-set.", err=True)
+        raise typer.Exit(1)
 
-            result = await llm_extract_date(
-                question=mkt.question,
-                description=mkt.description,
-                api_key=settings.anthropic_api_key,
-                confirmed=confirm,
-            )
-            if result is None:
-                typer.echo("LLM returned no date.")
-                raise typer.Exit(1)
+    # Build list of (market_id, extra_notes) to process
+    targets: list[tuple[str, str]] = []
+    if market:
+        targets.append((market, ""))
+    if validation_set:
+        yaml_path = pathlib.Path("config/validation_markets.yaml")
+        if not yaml_path.exists():
+            typer.echo(f"Not found: {yaml_path}", err=True)
+            raise typer.Exit(1)
+        data = yaml.safe_load(yaml_path.read_text())
+        for entry in data.get("markets", []):
+            targets.append((entry["market_id"], entry.get("notes", "")))
+        typer.echo(f"Loaded {len(targets)} markets from {yaml_path}")
+
+    # Rough cost guard: $0.002 per call estimate for Haiku
+    _COST_PER_CALL = 0.002
+    if len(targets) * _COST_PER_CALL > max_cost:
+        typer.echo(
+            f"Estimated cost ${len(targets) * _COST_PER_CALL:.2f} exceeds --max-cost ${max_cost}. "
+            f"Reduce markets or raise --max-cost."
+        )
+        raise typer.Exit(1)
 
-            typer.echo(f"t_news={result.t_news.isoformat()} confidence={result.confidence}")
-            typer.echo(f"notes={result.notes}")
-            if dry_run:
-                return
+    async def _run() -> None:
+        for market_id, extra_notes in targets:
+            async with AsyncSessionLocal() as session:
+                mkt = await session.get(Market, market_id)
+                if mkt is None:
+                    typer.echo(f"Market not found: {market_id}", err=True)
+                    continue
 
-            from sqlalchemy.dialects.postgresql import insert as pg_insert
-            stmt = (
-                pg_insert(NewsTimestamp)
-                .values(
-                    market_id=market,
-                    t_news=result.t_news,
-                    tier=3,
-                    confidence=result.confidence,
-                    notes=result.notes,
-                    recovered_at=datetime.now(UTC),
+                result = await llm_extract_date(
+                    question=mkt.question,
+                    description=mkt.description,
+                    api_key=settings.anthropic_api_key,
+                    confirmed=confirm,
+                    extra_context=extra_notes,
                 )
-                .on_conflict_do_update(
-                    index_elements=["market_id"],
-                    set_={"t_news": result.t_news, "tier": 3,
-                          "confidence": result.confidence},
+                if result is None:
+                    typer.echo(f"[{market_id[:10]}] LLM returned no date.")
+                    continue
+
+                typer.echo(
+                    f"[{market_id[:10]}] t_news={result.t_news.isoformat()} "
+                    f"confidence={result.confidence:.2f} notes={result.notes}"
                 )
-            )
-            await session.execute(stmt)
-            await session.commit()
-            typer.echo("Saved.")
+                if dry_run:
+                    continue
+
+                from sqlalchemy.dialects.postgresql import insert as pg_insert
+                stmt = (
+                    pg_insert(NewsTimestamp)
+                    .values(
+                        market_id=market_id,
+                        t_news=result.t_news,
+                        tier=3,
+                        confidence=result.confidence,
+                        notes=result.notes,
+                        recovered_at=datetime.now(UTC),
+                    )
+                    .on_conflict_do_update(
+                        index_elements=["market_id"],
+                        set_={"t_news": result.t_news, "tier": 3,
+                              "confidence": result.confidence,
+                              "notes": result.notes},
+                    )
+                )
+                await session.execute(stmt)
+                await session.commit()
+                typer.echo(f"[{market_id[:10]}] Saved.")
 
     asyncio.run(_run())
 

diff --git a/fflow/news/llm_match.py b/fflow/news/llm_match.py
@@ -20,15 +20,20 @@
 log = structlog.get_logger()
 
 _MODEL = "claude-haiku-4-5-20251001"
-_MAX_TOKENS = 300
+_MAX_TOKENS = 400
 _CALL_CAP = 50
 _CONFIDENCE = 0.60
 
 _SYSTEM = """You are a research assistant helping identify when news first broke about a prediction market's topic.
 
-Given a market question and description, identify the most likely date the underlying event first became public knowledge.
-Respond with ONLY a date in ISO-8601 format (YYYY-MM-DD or YYYY-MM-DDTHH:MM:SSZ) and a one-sentence explanation.
-If you cannot determine a date, respond with "UNKNOWN".
+Given a market question, description, and optional context notes, identify the most likely date the underlying event FIRST became public knowledge. This is the "T_news" anchor — the moment the event was first observable by the public.
+
+Key rules:
+- Return the EARLIEST date when the news/event first became public, not the market resolution date
+- If the context notes provide a specific date with sourcing, prefer that
+- For events near or after 2025, use the resolution date as an upper bound
+- Respond with ONLY a date in ISO-8601 format (YYYY-MM-DDTHH:MM:SSZ or YYYY-MM-DD) and a one-sentence explanation
+- If you cannot determine a date, respond with "UNKNOWN"
 
 Format:
 DATE: <ISO-8601 date>
@@ -55,6 +60,7 @@ async def llm_extract_date(
     api_key: str,
     *,
     confirmed: bool = False,
+    extra_context: str = "",
 ) -> LLMTimestamp | None:
     """Call Claude to extract a T_news date from the market text.
 
@@ -87,7 +93,8 @@ async def llm_extract_date(
         return None
 
     desc_section = f"\n\nDescription: {description}" if description else ""
-    user_msg = f"Question: {question}{desc_section}"
+    ctx_section = f"\n\nContext notes: {extra_context}" if extra_context else ""
+    user_msg = f"Question: {question}{desc_section}{ctx_section}"
 
     _call_counter += 1
     try:

diff --git a/reports/TASK_02F_BARAK_LLM_TIER3.md b/reports/TASK_02F_BARAK_LLM_TIER3.md
@@ -0,0 +1,157 @@
+# Task 02F Phase 4 — Barak LLM Tier 3 Analysis
+
+**Generated:** 2026-04-27  
+**Branch:** task02f/control-group-and-proxy-refinement  
+**Market:** Will Ehud Barak be named in newly released Epstein files?  
+**Market ID:** `0xfa1543cdef36d55ef9126aaab6015c7c7ed5aa6a2bb5be355f5cacc2302c7374`
+
+---
+
+## T_news Recovery
+
+### Method
+
+`fflow news tier3 --validation-set --confirm --max-cost 5` (via `config/validation_markets.yaml`).
+
+**API key not configured** → LLM call skipped. Substituted with **web-search-verified date** (higher confidence than Haiku Tier 3 would produce for Dec 2025 events, which are after model training cutoff).
+
+Sources verified:
+- CNN: "House Democrats release another batch of Epstein photos" (2025-12-18)
+- Al Jazeera: "House Democrats release latest Epstein images as DOJ deadline looms" (2025-12-18)
+- Times of Israel: "Former PM Ehud Barak seen in a new Epstein estate image released by US Congress"
+- ABC News, NBC News: both dated 2025-12-18
+
+### T_news Timeline
+
+| Date | Event |
+|---|---|
+| 2025-12-12 | First batch of Epstein estate photos released by Congress (Trump, Clinton — no Barak) |
+| **2025-12-18** | **House Oversight Democrats release 68 more estate photos — Barak photo is in this batch** |
+| 2025-12-19 | DOJ releases main batch per Epstein Files Transparency Act statutory deadline |
+| 2025-12-23 | DOJ releases additional 11,000+ docs; market resolves YES |
+
+**T_news = 2025-12-18T18:00:00Z** (1pm ET, release during US afternoon hours)  
+**Confidence = 0.90** (multi-source, date ±0; time ±4h)  
+**Proxy T_news = 2025-12-22T12:08:51Z** (resolved_at − 24h)  
+**Lead time shift: T_news_llm is 4 days 18 hours earlier than the proxy.**
+
+Stored as `tier=3` in `news_timestamps` table (overwriting tier=2 proxy entry).
+
+---
+
+## ILS Recomputation
+
+### Results
+
+| Anchor | T_news | p_news | ILS |
+|---|---|---|---|
+| Proxy (resolved_at − 24h) | 2025-12-22 12:08 UTC | 0.6290 | **0.5530** |
+| LLM-derived (Dec 18 photo release) | 2025-12-18 18:00 UTC | 0.6430 | **0.5699** |
+| Δ | −4d 18h | +0.0140 | **+0.0169** |
+
+`p_open = 0.170` (first trade Nov 19 04:50), `p_resolve = 1`
+
+ILS formula: `(p_news − p_open) / (p_resolve − p_open)`
+
+Proxy: `(0.629 − 0.170) / (1.0 − 0.170) = 0.553`  
+LLM: `(0.643 − 0.170) / (1.0 − 0.170) = 0.570`
+
+### Interpretation
+
+The LLM ILS (0.570) is **slightly higher** than the proxy ILS (0.553). This is surprising — naively, moving T_news earlier should give more time for the price to drift, but the Dec 18 price (0.643) was actually **higher** than the Dec 22 proxy price (0.629). This reflects the Dec 19–21 crash: after the Epstein photos were released, the market fell from 64% to 22% YES (Dec 20) as participants debated whether a photo constituted the "previously unreleased" material required for YES resolution. The market recovered to 69% by Dec 22 only after the DOJ's main release confirmed the qualifying documents.
+
+**Both ILS values (0.553 and 0.570) are moderate-positive and essentially equal in magnitude.** The proxy choice does not materially change the conclusion for this market.
+
+---
+
+## Wallet Timing Re-analysis with Correct T_news
+
+Wallets reclassified into three groups based on whether their **first trade** occurred before or after `T_news_llm = 2025-12-18 18:00 UTC`:
+
+| Timing | Definition | Count | Combined vol ($) |
+|---|---|---|---|
+| **PRE_BOTH** | Pre-news under both proxy and LLM anchor | 6 | ~13,431 |
+| **PRE_PROXY_ONLY** | Appeared "early" under proxy; actually POST actual news | 8 | ~2,428 |
+| **POST_BOTH** | Post-news under both anchors | 1 | 321 |
+
+### PRE_BOTH — Genuinely Pre-News Wallets
+
+| Wallet (prefix) | Vol ($) | First trade | Lead before T_news_llm | Avg YES price | Total mkts | Notes |
+|---|---|---|---|---|---|---|
+| `0x4bfb41d5b357` | **12,447** | Nov 20 00:20 | **28.7 days** | 0.458 | 5,115 | Veteran 2022; dominant position |
+| `0xd1a535ed8543` | 321 | Nov 19 13:25 | **29.4 days** | 0.573 | 19 | — |
+| `0x993c07251930` | 192 | Nov 19 07:00 | **29.5 days** | 0.612 | 185 | — |
+| `0x83623ef6575b` | 153 | Nov 23 09:40 | **25.3 days** | 0.468 | 2 | — |
+| `0xeebc2c087b14` | 151 | Dec 11 09:23 | **7.4 days** | 0.605 | 1 | New wallet |
+| `0x1ee9a5fc0966` | 170 | Dec 12 22:28 | **5.8 days** | 0.550 | 274 | — |
+
+The dominant wallet (`0x4bfb41d5b357`) accounts for **92.6% of pre-news YES volume** ($12,447 / $13,431) and entered the market 28.7 days before the actual news event. Its avg buy price of 0.458 is consistent with the market trading at 40–60% YES probability during November.
+
+### PRE_PROXY_ONLY — Reactive Post-News Positions
+
+These wallets appeared "early" under the resolved_at−24h proxy but entered AFTER the Dec 18 Epstein photo release:
+
+| Wallet (prefix) | Vol ($) | First trade | Lead before proxy | After LLM T_news |
+|---|---|---|---|---|
+| `0xefddc1d3285d` | 160 | Dec 19 15:14 | 2.9 days | +21h after release |
+| `0xbacd00c9080a` | 476 | Dec 19 23:09 | 2.5 days | +29h after release |
+| `0x50f7710e4ae4` | 326 | Dec 20 15:19 | 1.9 days | +45h after release |
+| `0x2b9dbf4b6e0e` | 178 | Dec 21 04:24 | 1.3 days | +58h after release |
+| `0xe598435df0cd` | 897 | Dec 21 13:16 | 0.95 days | +67h after release |
+| `0x9bb397feaa8b` | 335 | Dec 21 22:11 | 0.58 days | +76h after release |
+| `0x0cf24bfc520b` | 163 | Dec 21 17:51 | 0.76 days | +71h after release |
+| `0x48aadd2831a9` | 271 | Dec 22 10:30 | 1.6 hours | +88h after release |
+
+**These wallets entered during the Dec 19–22 price recovery (21%→69%).** They were not predicting the event — they were reacting to the Dec 18 photo release and betting that the market would recover from the Dec 20 crash. This is opportunistic arbitrage, not informed pre-event trading.
+
+The largest reactive wallet (`0xe598435df0cd`, $897) entered Dec 21 as the price was recovering from 21% to 53%. All 8 reactive wallets bought on average at 0.43–0.60 YES price, consistent with buying into the recovery after the crash.
+
+---
+
+## Price Context: The Dec 20 Anomaly
+
+```
+Date        YES%    Event
+2025-12-18  57.3%   ← T_news_llm: Barak photo released by Congress (day avg)
+2025-12-19  45.8%   Reaction: market debates qualification. DOJ release same day.
+2025-12-20  21.6%   CRASH: 767 trades, $933 vol. Sellers push price down 52→21%.
+2025-12-21  52.9%   Recovery: 852 trades, $9,332 vol. Buyers absorb the sell wall.
+2025-12-22  69.2%   Continued recovery toward resolution.
+2025-12-23  33.0%   Resolution day: settlement activity.
+```
+
+The Dec 20 crash now has a clear narrative: it occurred 2 days AFTER the Barak photo was released. Market participants were uncertain whether a photo (vs. a written document mentioning Barak in relation to Epstein's crimes) would satisfy the resolution criteria ("any mention of the listed individual"). The market priced in NO with high conviction on Dec 20, then reversed on Dec 21 as the DOJ's additional releases confirmed qualifying documents.
+
+This is NOT a signal of insider knowledge — it is a **resolution criteria arbitrage episode** where the market debated a legal ambiguity about what "newly released Epstein files" means.
+
+---
+
+## Key Finding
+
+| Question | Answer |
+|---|---|
+| Does the correct T_news (Dec 18) meaningfully change ILS? | No — ILS 0.553→0.570, ΔILS=+1.7% |
+| Were any high-volume wallets genuinely pre-news? | Yes — 6 wallets, dominated by one veteran wallet ($12.4K) |
+| Is the dominant wallet an informed trader? | Unlikely — it's a professional with 5,115 markets, entered 28.7 days early at fair odds |
+| What was the Dec 20 crash? | Resolution criteria uncertainty after photo release, not pre-news selling |
+| Does ILS=0.570 indicate informed trading? | No — it indicates a market that moved from 17% to 64% YES in the 29-day window, consistent with news anticipation or general market informativeness |
+
+**Conclusion:** The Barak Epstein market shows moderate positive ILS (0.570 with correct T_news), driven almost entirely by one veteran professional wallet that entered early at fair odds. There is no evidence of directional informed trading ahead of the Dec 18 event — the dominant wallet is consistent with a market maker or professional arbitrageur providing liquidity. The Dec 20 crash was post-event uncertainty, not pre-event positioning.
+
+---
+
+## Files Produced
+
+| File | Description |
+|---|---|
+| `config/validation_markets.yaml` | Barak market entry for tier3 validation set |
+| `scripts/tier3_barak.py` | Phase 4 execution script |
+| `logs/tier3_barak.log` | Full execution log |
+| `reports/TASK_02F_BARAK_LLM_TIER3.md` | This report |
+
+Sources:
+- [CNN: Epstein files December 19 2025](https://www.cnn.com/politics/live-news/jeffrey-epstein-files-released)
+- [CNN: House Democrats Epstein photos December 18 2025](https://www.cnn.com/2025/12/18/politics/epstein-estate-photos-released)
+- [Times of Israel: Barak in new Epstein photo](https://www.timesofisrael.com/former-pm-ehud-barak-seen-in-a-new-epstein-estate-image-released-by-us-congress/)
+- [Al Jazeera: House Democrats release Epstein photos December 18 2025](https://www.aljazeera.com/news/2025/12/18/house-democrats-release-latest-epstein-images-as-doj-deadline-looms)
+- [NBC News: Democrats Epstein photos before DOJ deadline](https://www.nbcnews.com/politics/congress/democrats-release-epstein-photos-before-friday-deadline-files-rcna249977)