add Cold Start tab: SSL pretraining lifts new-merchant fraud catch ~5x

new encoder/src/demo/cold_start.py renders the result tab: three stat
cards, the two-panel figure (fraud caught vs label history; fraud
caught vs approval rate at 1% labels), a 7-column operating-point
table read from docs/figures/recall_at_approval.json, the "how it
works" paragraph for the SSL stage, and a synthetic-data scope block.

wired as a new tab in both app.py (standalone 4-tab) and
copilot_app_unified.py (7-tab unified). figure is base64-embedded so
the tab is self-contained html. *.png added to LFS tracking.

result on synthetic data (held-out test, n=20k, 3.7% base rate, 1%
labels = stand-in for a newly onboarded entity, 95% approval rate):
fraud caught 7.5% -> 36.0% (~4.8x), false-decline rate 4.9% -> 3.8%.
roc-auc 0.575 -> 0.814; pr-auc 0.048 -> 0.210. whole-book bonus at
full labels and 1% decline budget: +16 pts.

mechanism: stage-1 self-supervised next-feature prediction through the
frozen lfm2.5 base anchors the encoder's value tables before the
supervised fine-tune. same connector-pretraining stage lfm2.5-audio
and lfm2.5-vl already use.

.gitattributes

@@ -1,3 +1,4 @@
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.npy filter=lfs diff=lfs merge=lfs -text
 *.npz filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -22,11 +22,16 @@ per-surface LoRA adapts the attention layers.
 
 ## What this demo shows
 
-Six tabs over one shared LFM2.5-350M backbone:
+Seven tabs over one shared LFM2.5-350M backbone:
 
 - **Multi-Head Demo** — the original 4-task-head encoder demo. Live
   inference on curated customer archetypes: fraud probability,
   next-merchant prediction, amount-bucket prediction, MCC prediction.
+- **Cold Start** — a brief self-supervised pretraining stage ("SSL")
+  catches ~5× more fraud at a fixed approval rate on a newly onboarded
+  merchant with almost no labels, while declining fewer good customers.
+  The same connector-pretraining stage LFM2.5-Audio and LFM2.5-VL use.
+  Synthetic-data demonstration with operating-point numbers + figure.
 - **Why Liquid** — architectural pitch for the encoder-on-frozen-backbone
   recipe.
 - **Integration** — build-it-yourself guide.

docs/figures/recall_at_approval.json

@@ -0,0 +1,211 @@
+{
+  "base_rate": 0.03665,
+  "n_test": 20000,
+  "split": "test",
+  "runs": {
+    "ssl_1pct": {
+      "roc_auc": 0.8140344300750755,
+      "pr_auc": 0.2104627331942034,
+      "operating_points": {
+        "appr99": {
+          "approval_rate": 0.99,
+          "decline_rate": 0.01145,
+          "recall": 0.12551159618008187,
+          "false_decline_rate": 0.007110603622774692,
+          "precision": 0.4017467248908297
+        },
+        "appr98": {
+          "approval_rate": 0.98,
+          "decline_rate": 0.02005,
+          "recall": 0.19508867667121418,
+          "false_decline_rate": 0.01339077178595526,
+          "precision": 0.3566084788029925
+        },
+        "appr97": {
+          "approval_rate": 0.97,
+          "decline_rate": 0.03005,
+          "recall": 0.2660300136425648,
+          "false_decline_rate": 0.021072299787200915,
+          "precision": 0.324459234608985
+        },
+        "appr95": {
+          "approval_rate": 0.95,
+          "decline_rate": 0.05005,
+          "recall": 0.3601637107776262,
+          "false_decline_rate": 0.03825193335755437,
+          "precision": 0.26373626373626374
+        }
+      }
+    },
+    "nossl_1pct": {
+      "roc_auc": 0.5746099668824207,
+      "pr_auc": 0.04786141744299679,
+      "operating_points": {
+        "appr99": {
+          "approval_rate": 0.99,
+          "decline_rate": 0.01,
+          "recall": 0.021828103683492497,
+          "false_decline_rate": 0.009550007785332434,
+          "precision": 0.08
+        },
+        "appr98": {
+          "approval_rate": 0.98,
+          "decline_rate": 0.02005,
+          "recall": 0.03956343792633015,
+          "false_decline_rate": 0.0193076244355634,
+          "precision": 0.07231920199501247
+        },
+        "appr97": {
+          "approval_rate": 0.97,
+          "decline_rate": 0.03,
+          "recall": 0.05320600272851296,
+          "false_decline_rate": 0.029117143302018995,
+          "precision": 0.065
+        },
+        "appr95": {
+          "approval_rate": 0.95,
+          "decline_rate": 0.0502,
+          "recall": 0.07503410641200546,
+          "false_decline_rate": 0.04925520319717652,
+          "precision": 0.054780876494023904
+        }
+      }
+    },
+    "ssl_10pct": {
+      "roc_auc": 0.9393715555037555,
+      "pr_auc": 0.8570152888812671,
+      "operating_points": {
+        "appr99": {
+          "approval_rate": 0.99,
+          "decline_rate": 0.02215,
+          "recall": 0.6016371077762619,
+          "false_decline_rate": 0.00010380443244926559,
+          "precision": 0.9954853273137697
+        },
+        "appr98": {
+          "approval_rate": 0.98,
+          "decline_rate": 0.02215,
+          "recall": 0.6016371077762619,
+          "false_decline_rate": 0.00010380443244926559,
+          "precision": 0.9954853273137697
+        },
+        "appr97": {
+          "approval_rate": 0.97,
+          "decline_rate": 0.03,
+          "recall": 0.7803547066848567,
+          "false_decline_rate": 0.0014532620542897182,
+          "precision": 0.9533333333333334
+        },
+        "appr95": {
+          "approval_rate": 0.95,
+          "decline_rate": 0.05015,
+          "recall": 0.8472032742155525,
+          "false_decline_rate": 0.019826646597809725,
+          "precision": 0.6191425722831505
+        }
+      }
+    },
+    "nossl_10pct": {
+      "roc_auc": 0.9259714016664364,
+      "pr_auc": 0.7596305449815982,
+      "operating_points": {
+        "appr99": {
+          "approval_rate": 0.99,
+          "decline_rate": 0.0139,
+          "recall": 0.37789904502046384,
+          "false_decline_rate": 5.1902216224632795e-05,
+          "precision": 0.9964028776978417
+        },
+        "appr98": {
+          "approval_rate": 0.98,
+          "decline_rate": 0.02,
+          "recall": 0.5266030013642565,
+          "false_decline_rate": 0.0007266310271448591,
+          "precision": 0.965
+        },
+        "appr97": {
+          "approval_rate": 0.97,
+          "decline_rate": 0.03,
+          "recall": 0.6657571623465212,
+          "false_decline_rate": 0.005813048217158873,
+          "precision": 0.8133333333333334
+        },
+        "appr95": {
+          "approval_rate": 0.95,
+          "decline_rate": 0.05,
+          "recall": 0.7612551159618008,
+          "false_decline_rate": 0.022940779571287696,
+          "precision": 0.558
+        }
+      }
+    },
+    "ssl_100pct": {
+      "roc_auc": 0.9561845809915676,
+      "pr_auc": 0.8820119418197224,
+      "operating_points": {
+        "appr99": {
+          "approval_rate": 0.99,
+          "decline_rate": 0.01595,
+          "recall": 0.43519781718963163,
+          "false_decline_rate": 0.0,
+          "precision": 1.0
+        },
+        "appr98": {
+          "approval_rate": 0.98,
+          "decline_rate": 0.02015,
+          "recall": 0.5497953615279673,
+          "false_decline_rate": 0.0,
+          "precision": 1.0
+        },
+        "appr97": {
+          "approval_rate": 0.97,
+          "decline_rate": 0.03005,
+          "recall": 0.7939972714870396,
+          "false_decline_rate": 0.000986142108268023,
+          "precision": 0.9683860232945092
+        },
+        "appr95": {
+          "approval_rate": 0.95,
+          "decline_rate": 0.05,
+          "recall": 0.8663028649386084,
+          "false_decline_rate": 0.01894430892199097,
+          "precision": 0.635
+        }
+      }
+    },
+    "nossl_100pct": {
+      "roc_auc": 0.9452378866918683,
+      "pr_auc": 0.7709143923552446,
+      "operating_points": {
+        "appr99": {
+          "approval_rate": 0.99,
+          "decline_rate": 0.01005,
+          "recall": 0.2742155525238745,
+          "false_decline_rate": 0.0,
+          "precision": 1.0
+        },
+        "appr98": {
+          "approval_rate": 0.98,
+          "decline_rate": 0.02,
+          "recall": 0.5252387448840382,
+          "false_decline_rate": 0.0007785332433694919,
+          "precision": 0.9625
+        },
+        "appr97": {
+          "approval_rate": 0.97,
+          "decline_rate": 0.03035,
+          "recall": 0.669849931787176,
+          "false_decline_rate": 0.006020657082057404,
+          "precision": 0.8088962108731467
+        },
+        "appr95": {
+          "approval_rate": 0.95,
+          "decline_rate": 0.05005,
+          "recall": 0.7639836289222374,
+          "false_decline_rate": 0.02288887735506306,
+          "precision": 0.5594405594405595
+        }
+      }
+    }
+  }
+}

encoder/src/demo/app.py

@@ -36,6 +36,7 @@ from src.demo.decode import TransactionDecoder
 from src.demo.merchant_catalog import DemoMerchantCatalog
 from src.demo.profile_inference import format_profile_html, infer_profile
 
+from encoder.src.demo.cold_start import render_cold_start
 from encoder.src.demo.inference import EncoderDemoModel
 from encoder.src.demo.render import (
     format_amount_predictions,
@@ -380,6 +381,11 @@ def _build_integration_tab_contents() -> None:
     gr.HTML(render_encoder_integration())
 
 
+def _build_cold_start_tab_contents() -> None:
+    """Render the Cold Start tab content (HTML self-supervised-pretraining result)."""
+    gr.HTML(render_cold_start())
+
+
 def create_app(
     model: EncoderDemoModel,
     data: DemoData,
@@ -396,6 +402,8 @@ def create_app(
         with gr.Tabs():
             with gr.Tab("Demo"):
                 _build_demo_tab_contents(model, data, decoder, merchant_catalog, app)
+            with gr.Tab("Cold Start"):
+                _build_cold_start_tab_contents()
             with gr.Tab("Why Liquid"):
                 _build_why_liquid_tab_contents()
             with gr.Tab("Integration"):

encoder/src/demo/cold_start.py

@@ -0,0 +1,346 @@
+"""Cold-start (self-supervised pretraining) result tab for the encoder demo.
+
+A third piece of demo-specific content alongside `render_why_encoder` and
+`render_encoder_integration` in render.py. It presents one result: on a newly
+onboarded merchant with almost no fraud labels, adding a brief self-supervised
+pretraining stage ("SSL") catches far more fraud at the same approval rate
+while declining fewer good customers. The lift concentrates where label history
+is scarce (new merchants, new markets) and shrinks as history accumulates.
+
+Design choices, and why:
+  - Numbers are read at render time from the recall-at-approval JSON produced by
+    `encoder/scripts/recall_at_approval.py`. That artifact is the single source
+    of truth; nothing here is hardcoded, so the tab can never drift from the
+    measured numbers.
+  - The figure is base64-embedded as a data URI so the tab is self-contained
+    HTML, matching the other content tabs (no Gradio static-file-serving config).
+  - Design tokens are imported from render.py so there is one design system.
+
+Scope of claim: every number is synthetic-data, held-out test split. The tab
+labels it a demonstration throughout. It is never a production or customer
+guarantee; the pilot produces the number on the customer's portfolio.
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import logging
+from pathlib import Path
+from typing import Any
+
+# Single source of truth for the design system lives in render.py. Importing the
+# tokens (rather than redefining them) keeps this tab visually identical to the
+# Why Liquid / Integration tabs.
+from encoder.src.demo.render import (
+    _BG_CARD,
+    _BG_CARD_ALT,
+    _BORDER,
+    _BORDER_SUBTLE,
+    _CONTAINER_WIDTH,
+    _FONT_MONO,
+    _RADIUS_CARD,
+    _TEXT,
+    _TEXT_DIM,
+    _TEXT_MUTED,
+)
+
+log = logging.getLogger(__name__)
+
+# Repo-root-relative defaults (robust to the process cwd). This file lives at
+# encoder/src/demo/, so parents[3] is the repository root.
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+# Neutral figure (no customer name in caption) for the generic Gradio demo.
+# The dLocal-specific figure (docs/figures/dlocal_cold_start.png) is used in
+# the dLocal deck only; pass figure_path explicitly to switch.
+_FIG_DEFAULT = _REPO_ROOT / "docs" / "figures" / "cold_start_demo.png"
+_JSON_DEFAULT = _REPO_ROOT / "docs" / "figures" / "recall_at_approval.json"
+
+# Teal = "with SSL" in the figure; gray = "without". Reuse the same mapping for
+# the table so the tab's color language matches the embedded chart.
+_SSL = "#0F766E"          # teal-700, with SSL
+_SSL_BG = "rgba(15,118,110,0.06)"
+_NO = "#6B7280"           # gray-500, without SSL
+
+
+# ---------------------------------------------------------------------------
+# Small formatters and JSON accessors
+# ---------------------------------------------------------------------------
+
+
+def _img_data_uri(path: Path) -> str:
+    """Base64-encode a PNG as an inline data URI. Returns '' if the file is absent.
+
+    The figure is cosmetic; a missing file degrades to the table-only view with
+    a logged warning rather than crashing the app.
+    """
+    try:
+        b64 = base64.b64encode(path.read_bytes()).decode("ascii")
+    except OSError as exc:
+        log.warning("cold-start figure not found at %s (%s); rendering without it", path, exc)
+        return ""
+    return f"data:image/png;base64,{b64}"
+
+
+def _op(runs: Any, tag: str, appr: int, field: str) -> float:  # noqa: ANN401 (json is Any)
+    """One operating-point field (recall / false_decline_rate / precision) from the JSON."""
+    return float(runs[tag]["operating_points"][f"appr{appr}"][field])
+
+
+def _pct(x: float) -> str:
+    """Recall / percentage with one decimal, e.g. 0.3601 -> '36.0%'."""
+    return f"{x * 100:.1f}%"
+
+
+def _fdr(x: float) -> str:
+    """False-decline rate; small values keep two decimals so they don't read as zero."""
+    p = x * 100
+    if p >= 1.0:
+        return f"{p:.1f}%"
+    if p > 0.0:
+        return f"{p:.2f}%"
+    return "0.0%"
+
+
+def _lift(ssl: float, base: float) -> str:
+    """Multiplicative lift as 'N.N×'.
+
+    Computed from the displayed (1-decimal-percent) recall values, not the raw
+    fractions, so the lift column equals SSL% / no-SSL% when a reader divides the
+    two columns by eye (and matches the talk track). Guarded against divide-by-zero.
+    """
+    s = round(ssl * 100, 1)
+    b = round(base * 100, 1)
+    return f"{s / max(b, 1e-9):.1f}×"
+
+
+# ---------------------------------------------------------------------------
+# HTML fragments
+# ---------------------------------------------------------------------------
+
+
+def _stat_card(big: str, big_color: str, label: str, detail: str) -> str:
+    """One headline stat: a large colored number, a label, and a supporting line."""
+    return f"""
+    <div style="padding: 14px 16px; background: {_BG_CARD_ALT}; border: 1px solid {_BORDER};
+                border-radius: {_RADIUS_CARD};">
+      <div style="font-size: 28px; font-weight: 700; color: {big_color};
+                  letter-spacing: -0.02em; line-height: 1;">{big}</div>
+      <div style="font-size: 12px; font-weight: 600; color: {_TEXT}; margin: 8px 0 4px 0;">
+        {label}</div>
+      <div style="font-size: 11px; color: {_TEXT_DIM}; line-height: 1.45;">{detail}</div>
+    </div>"""
+
+
+def _table_header(cols: list[str]) -> str:
+    ths = ""
+    for i, c in enumerate(cols):
+        align = "left" if i == 0 else "right"
+        ths += (
+            f'<th style="padding: 6px 10px; font-size: 10px; color: {_TEXT_DIM};'
+            f' text-transform: uppercase; letter-spacing: 0.05em; text-align: {align};'
+            f' font-weight: 600;">{c}</th>'
+        )
+    return f"<tr style='border-bottom: 1px solid {_BORDER};'>{ths}</tr>"
+
+
+def _table_row(cells: list[str], highlight: bool = False) -> str:
+    bg = f"background: {_SSL_BG};" if highlight else ""
+    tds = ""
+    for i, c in enumerate(cells):
+        align = "left" if i == 0 else "right"
+        weight = "600" if highlight and i == 0 else "400"
+        tds += (
+            f'<td style="padding: 6px 10px; font-family: {_FONT_MONO}; font-size: 11px;'
+            f' color: {_TEXT}; text-align: {align}; font-weight: {weight};">{c}</td>'
+        )
+    return f"<tr style='border-bottom: 1px solid {_BORDER_SUBTLE}; {bg}'>{tds}</tr>"
+
+
+# ---------------------------------------------------------------------------
+# Main render
+# ---------------------------------------------------------------------------
+
+
+def render_cold_start(
+    json_path: Path | None = None,
+    figure_path: Path | None = None,
+) -> str:
+    """Render the Cold Start (SSL result) tab content as an HTML string.
+
+    Reads every number from the recall-at-approval JSON (single source of truth)
+    and base64-embeds the result figure so the tab is self-contained.
+
+    Args:
+        json_path: recall_at_approval.json. Defaults to docs/figures/ in the repo.
+        figure_path: dlocal_cold_start.png. Defaults to docs/figures/ in the repo.
+
+    Returns:
+        HTML string suitable for a ``gr.HTML`` block.
+    """
+    json_path = json_path or _JSON_DEFAULT
+    figure_path = figure_path or _FIG_DEFAULT
+
+    raw: Any = json.loads(json_path.read_text())
+    runs: Any = raw["runs"]
+    n_test = int(raw["n_test"])
+    base_rate = float(raw["base_rate"])
+
+    # Headline (cold start = 1% labels, 95% approval).
+    cs_ssl = _op(runs, "ssl_1pct", 95, "recall")
+    cs_no = _op(runs, "nossl_1pct", 95, "recall")
+    cs_fdr_ssl = _op(runs, "ssl_1pct", 95, "false_decline_rate")
+    cs_fdr_no = _op(runs, "nossl_1pct", 95, "false_decline_rate")
+    # Whole-book bonus at full labels, tight 1% decline budget.
+    wb_ssl = _op(runs, "ssl_100pct", 99, "recall")
+    wb_no = _op(runs, "nossl_100pct", 99, "recall")
+    wb_delta_pts = (wb_ssl - wb_no) * 100
+
+    stats = "".join([
+        _stat_card(
+            _lift(cs_ssl, cs_no), _SSL,
+            "more fraud caught on a new merchant",
+            f"At a 95% approval rate with only 1% of labels: {_pct(cs_ssl)} of fraud "
+            f"caught vs {_pct(cs_no)} without pretraining.",
+        ),
+        _stat_card(
+            f"&minus;{(cs_fdr_no - cs_fdr_ssl) * 100:.1f} pts", _SSL,
+            "fewer good customers declined",
+            f"False-decline rate {_fdr(cs_fdr_ssl)} vs {_fdr(cs_fdr_no)} at the same "
+            f"approval rate. More fraud caught and fewer false declines, together.",
+        ),
+        _stat_card(
+            f"+{wb_delta_pts:.0f} pts", _TEXT,
+            "whole-book bonus at full labels",
+            f"Even with full label history, {_pct(wb_ssl)} vs {_pct(wb_no)} fraud caught "
+            "at a tight 1% decline budget. Not only a new-entity trick.",
+        ),
+    ])
+
+    # Business-units table. (label, frac_tag, approval, highlight-as-cold-start)
+    specs: list[tuple[str, str, int, bool]] = [
+        ("1% &mdash; new entity", "1pct", 99, True),
+        ("1% &mdash; new entity", "1pct", 95, True),
+        ("10%", "10pct", 99, False),
+        ("100% &mdash; full", "100pct", 99, False),
+        ("100% &mdash; full", "100pct", 95, False),
+    ]
+    rows = ""
+    for label, frac, appr, hl in specs:
+        ssl_r = _op(runs, f"ssl_{frac}", appr, "recall")
+        no_r = _op(runs, f"nossl_{frac}", appr, "recall")
+        ssl_f = _op(runs, f"ssl_{frac}", appr, "false_decline_rate")
+        no_f = _op(runs, f"nossl_{frac}", appr, "false_decline_rate")
+        rows += _table_row(
+            [label, f"{appr}%", _pct(ssl_r), _pct(no_r), _lift(ssl_r, no_r),
+             _fdr(ssl_f), _fdr(no_f)],
+            highlight=hl,
+        )
+
+    # Ranking quality, for the evaluator who wants AUC alongside operating points.
+    roc_1 = (float(runs["nossl_1pct"]["roc_auc"]), float(runs["ssl_1pct"]["roc_auc"]))
+    roc_100 = (float(runs["nossl_100pct"]["roc_auc"]), float(runs["ssl_100pct"]["roc_auc"]))
+    pr_1 = (float(runs["nossl_1pct"]["pr_auc"]), float(runs["ssl_1pct"]["pr_auc"]))
+    pr_100 = (float(runs["nossl_100pct"]["pr_auc"]), float(runs["ssl_100pct"]["pr_auc"]))
+
+    fig_uri = _img_data_uri(figure_path)
+    if fig_uri:
+        figure_block = f"""
+      <div style="padding: 12px; background: {_BG_CARD}; border: 1px solid {_BORDER};
+                  border-radius: {_RADIUS_CARD}; margin-bottom: 16px; text-align: center;">
+        <img src="{fig_uri}" alt="Cold-start fraud-catch lift from self-supervised pretraining"
+             style="max-width: 100%; height: auto; border-radius: 8px;" />
+      </div>"""
+    else:
+        figure_block = ""
+
+    return f"""
+    <div style="max-width: {_CONTAINER_WIDTH}; margin: 0 auto; padding: 16px;
+                font-family: -apple-system, BlinkMacSystemFont, Segoe UI, Roboto, sans-serif;">
+
+      <!-- Lead -->
+      <h2 style="margin: 0 0 4px 0; color: {_TEXT}; font-size: 22px; font-weight: 700;
+                 letter-spacing: -0.02em;">
+        Cold Start: the Fraud-Catch Lift Where You Have No History Yet
+      </h2>
+      <p style="color: {_TEXT_DIM}; font-size: 13px; margin: 0 0 20px 0; line-height: 1.55;">
+        A newly onboarded merchant or market has almost no fraud labels, so a model trained
+        only on labels has nothing to stand on. Adding one brief self-supervised pretraining
+        stage &mdash; before any labels &mdash; anchors the model on transaction <i>behavior</i>
+        learned across the rest of the book. The result: it catches far more fraud the day you
+        enter a new market, at the same approval rate, while turning away fewer good customers.
+      </p>
+
+      <!-- Headline stats -->
+      <div style="display: grid; grid-template-columns: 1fr 1fr 1fr; gap: 12px; margin-bottom: 16px;">
+        {stats}
+      </div>
+
+      <!-- Figure -->
+      {figure_block}
+
+      <!-- Operating-point table -->
+      <div style="padding: 16px 20px; background: {_BG_CARD}; border: 1px solid {_BORDER};
+                  border-radius: {_RADIUS_CARD}; margin-bottom: 16px;">
+        <h3 style="color: {_TEXT}; margin: 0 0 4px 0; font-size: 15px; font-weight: 600;">
+          Fraud Caught at a Fixed Approval Rate
+        </h3>
+        <p style="color: {_TEXT_MUTED}; font-size: 12px; line-height: 1.55; margin: 0 0 10px 0;">
+          What a fraud team actually runs: hold the approval rate fixed (decline the riskiest
+          tail), then measure how much fraud you catch and how many good customers you decline.
+          Label history stands in for entity age &mdash; <b>1% = a new entity</b>, 100% = a
+          mature one. The lift is largest where history is scarce and shrinks as it accumulates.
+        </p>
+        <table style="width: 100%; border-collapse: collapse; margin-bottom: 8px;">
+          {_table_header(["Label history", "Approval", "Fraud caught (SSL)",
+                          "Fraud caught (no SSL)", "Lift", "False decl. (SSL)", "(no SSL)"])}
+          {rows}
+        </table>
+        <p style="font-size: 11px; color: {_TEXT_DIM}; margin: 0; line-height: 1.5;">
+          Ranking quality (test split): ROC-AUC {roc_1[0]:.3f} &rarr; <b>{roc_1[1]:.3f}</b> at 1%
+          labels, {roc_100[0]:.3f} &rarr; <b>{roc_100[1]:.3f}</b> at full labels.
+          PR-AUC {pr_1[0]:.3f} &rarr; <b>{pr_1[1]:.3f}</b> and
+          {pr_100[0]:.3f} &rarr; <b>{pr_100[1]:.3f}</b> &mdash; PR-AUC is the honest metric at a
+          {base_rate:.1%} fraud base rate.
+        </p>
+      </div>
+
+      <!-- How it works -->
+      <div style="padding: 16px 20px; background: {_BG_CARD}; border: 1px solid {_BORDER};
+                  border-radius: {_RADIUS_CARD}; margin-bottom: 16px;">
+        <h3 style="color: {_TEXT}; margin: 0 0 4px 0; font-size: 15px; font-weight: 600;">
+          How It Works: One Added Pretraining Stage
+        </h3>
+        <p style="color: {_TEXT_MUTED}; font-size: 13px; line-height: 1.6; margin: 0;">
+          The encoder originally trained in a single supervised pass with randomly-initialized
+          per-feature embeddings. We add a stage-1 self-supervised step &mdash; next-feature
+          prediction over the transaction corpus, run through the frozen LFM2.5 base &mdash; that
+          anchors the encoder&rsquo;s value tables <i>before</i> the supervised fine-tune. It is
+          the same connector-pretraining stage LFM2.5-Audio and LFM2.5-VL already use, which the
+          transaction encoder had skipped. The base stays frozen and the shared-base,
+          many-surfaces story is unchanged; this adds an unsupervised warm-up, not a new model.
+        </p>
+      </div>
+
+      <!-- Scope / caveat -->
+      <div style="padding: 14px 18px; background: {_BG_CARD_ALT}; border: 1px solid {_BORDER};
+                  border-radius: {_RADIUS_CARD}; margin-bottom: 16px;">
+        <div style="font-family: {_FONT_MONO}; font-size: 10px; color: {_TEXT_DIM};
+                    font-weight: 600; text-transform: uppercase; letter-spacing: 0.05em;
+                    margin-bottom: 6px;">Scope of claim</div>
+        <p style="color: {_TEXT_MUTED}; font-size: 12px; line-height: 1.6; margin: 0;">
+          Synthetic data, held-out test split (n={n_test:,}, fraud base rate {base_rate:.1%}).
+          Label-scarcity (1% &asymp; 1,700 labeled sequences) stands in for a newly onboarded
+          entity before label history accumulates. <b>The absolute numbers don&rsquo;t
+          transfer; the shape and the mechanism do.</b> The pilot produces the real number on
+          the customer&rsquo;s portfolio &mdash; hold out a recently onboarded market, score fraud
+          on it cold against the current baseline, at the customer&rsquo;s approval rate.
+        </p>
+      </div>
+
+      <div style="font-family: {_FONT_MONO}; font-size: 10px; color: {_TEXT_DIM}; text-align: center;">
+        Same architecture, same frozen base, same fine-tune &mdash; trained with vs without the
+        self-supervised stage. One variable changed.
+      </div>
+    </div>
+    """

encoder/src/demo/copilot_app_unified.py

@@ -42,6 +42,7 @@ from src.demo.decode import TransactionDecoder
 from src.demo.merchant_catalog import DemoMerchantCatalog
 
 from encoder.src.demo.app import (
+    _build_cold_start_tab_contents,
     _build_demo_tab_contents,
     _build_integration_tab_contents,
     _build_why_liquid_tab_contents,
@@ -94,6 +95,8 @@ def build_unified_ui(
                     multihead_merchant_catalog,
                     app,
                 )
+            with gr.Tab("Cold Start"):
+                _build_cold_start_tab_contents()
             with gr.Tab("Why Liquid"):
                 _build_why_liquid_tab_contents()
             with gr.Tab("Integration"):
@@ -116,12 +119,12 @@ def _load_multihead(
     device: str,
 ) -> tuple[EncoderDemoModel, DemoData, TransactionDecoder, DemoMerchantCatalog]:
     """Load the original multi-head V3 model + curated test data."""
-    print(f"[multihead] schema + data ...")
+    print("[multihead] schema + data ...")
     schema_cfg = load_schema(schema)
     data = DemoData(data_dir, schema_cfg)
     merchant_catalog = DemoMerchantCatalog(schema_cfg)
     decoder = TransactionDecoder(schema_cfg, merchant_catalog)
-    print(f"[multihead] loading EncoderDemoModel ...")
+    print("[multihead] loading EncoderDemoModel ...")
     model = EncoderDemoModel(
         model_config_path=config,
         schema_path=schema,
@@ -148,7 +151,7 @@ def _load_copilots(
     device: torch.device,
 ) -> tuple[CopilotModel, CollectionsCopilotModel, FraudPatternCopilotModel]:
     """Load the three Co-Pilot surfaces. Each has its own backbone copy."""
-    print(f"[copilot 1/3] loading Dispute ...")
+    print("[copilot 1/3] loading Dispute ...")
     dispute_model = CopilotModel.from_paths(
         checkpoint_path=dispute_checkpoint,
         model_config_path=dispute_config,
@@ -157,7 +160,7 @@ def _load_copilots(
         cast_path=dispute_cast,
         device=device,
     )
-    print(f"[copilot 2/3] loading Collections ...")
+    print("[copilot 2/3] loading Collections ...")
     collections_model = CollectionsCopilotModel.from_paths(
         checkpoint_path=collections_checkpoint,
         model_config_path=collections_config,
@@ -166,7 +169,7 @@ def _load_copilots(
         cast_path=collections_cast,
         device=device,
     )
-    print(f"[copilot 3/3] loading Fraud ...")
+    print("[copilot 3/3] loading Fraud ...")
     fraud_model = FraudPatternCopilotModel.from_paths(
         checkpoint_path=fraud_checkpoint,
         model_config_path=fraud_config,