Upload folder using huggingface_hub

app/src/je_validation/console/static/app.js

@@ -793,6 +793,34 @@ async function openPastRun(runId) {
   openLiveView(runId, snap.contract);
 }
 
+// ---- deep links from the Tasks tab -----------------------------------------
+//
+// tabs.js translates #console/run/<run_id> into a "je:open-run" event (it is
+// the only hash reader). Opening the run scrolls to the live view; the pulse
+// on the panel and on the run's history row says "this is the one you
+// clicked". Unknown run ids degrade to the plain console (openPastRun
+// already swallows the failed fetch).
+
+function pulseOnce(el) {
+  if (!el) return;
+  el.classList.remove('deep-flash');
+  void el.offsetWidth;              // restart the animation on repeat clicks
+  el.classList.add('deep-flash');
+  setTimeout(() => el.classList.remove('deep-flash'), 1600);
+}
+
+async function openDeepLinkRun(runId) {
+  await loadRunHistory();           // ensure the row exists before pulsing it
+  await openPastRun(runId);
+  if (!Console.live || Console.live.runId !== runId) return;
+  pulseOnce(document.querySelector('#live .panel'));
+  const idEl = [...document.querySelectorAll('#history-list .history-id')]
+    .find((el) => el.textContent === runId);
+  if (idEl) pulseOnce(idEl.closest('.history-row'));
+}
+
+window.addEventListener('je:open-run', (e) => openDeepLinkRun(e.detail));
+
 // ---- live view: SSE client -------------------------------------------------
 //
 // A fresh EventSource cannot send Last-Event-ID (the browser only adds it on

app/src/je_validation/console/static/content/tasks.html

@@ -15,24 +15,65 @@
     <div class="task-card"><h3><span class="je-id">JE-01</span> Broken entries
       <span class="badge-live">Live</span></h3>
       <p>An overnight interface load has left a batch of entries damaged. Find
-      every entry that no longer balances or is missing required pieces.</p></div>
+      every entry that no longer balances or is missing required pieces.</p>
+      <!-- example:JE-01 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.04 ± 0.02 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.80 · no-evidence 0.00</span>
+        <a class="ex-link" href="#console/run/c5-grid-je01">View the full run →</a>
+      </div>
+      <!-- /example:JE-01 --></div>
     <div class="task-card"><h3><span class="je-id">JE-02</span> Coding validity
       <span class="badge-live">Live</span></h3>
       <p>Account, fund, and department codes must exist, be active in the
       period, and be an allowed combination. Find the entries that violate any
-      of it.</p></div>
+      of it.</p>
+      <!-- example:JE-02 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 1.00 ± 0.00 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.80 · no-evidence 0.22</span>
+        <a class="ex-link" href="#console/run/c5-grid-je02">View the full run →</a>
+      </div>
+      <!-- /example:JE-02 --></div>
     <div class="task-card"><h3><span class="je-id">JE-03</span> Period cut-off
       <span class="badge-live">Live</span></h3>
       <p>Entries must be posted to the period their dates belong to. Find the
-      ones recorded in the wrong period.</p></div>
+      ones recorded in the wrong period.</p>
+      <!-- example:JE-03 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.21 ± 0.02 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.80 · no-evidence 0.00</span>
+        <a class="ex-link" href="#console/run/c5-grid-je03">View the full run →</a>
+      </div>
+      <!-- /example:JE-03 --></div>
     <div class="task-card"><h3><span class="je-id">JE-04</span> Duplicate payments
       <span class="badge-live">Live</span></h3>
       <p>A vendor claims they were paid twice. Separate true duplicates from
-      legitimate split payments and properly re-issued checks.</p></div>
+      legitimate split payments and properly re-issued checks.</p>
+      <!-- example:JE-04 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.10 ± 0.30 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.80 · no-evidence 0.00</span>
+        <a class="ex-link" href="#console/run/c5-grid-je04">View the full run →</a>
+      </div>
+      <!-- /example:JE-04 --></div>
     <div class="task-card"><h3><span class="je-id">JE-05</span> Foreign currency
       <span class="badge-live">Live</span></h3>
       <p>Foreign-currency entries must translate at the official period rate.
-      Find wrong rates and translation errors.</p></div>
+      Find wrong rates and translation errors.</p>
+      <!-- example:JE-05 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.16 ± 0.02 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.71 · no-evidence 0.18</span>
+        <span class="ex-fail">1 of 5 seeds auto-failed (fabrication)</span>
+        <a class="ex-link" href="#console/run/c5-grid-je05">View the full run →</a>
+      </div>
+      <!-- /example:JE-05 --></div>
   </div>
 
   <div class="tier-head tier-l2"><span class="tier-tag">L2</span><h2>Analytical checks</h2>
@@ -44,23 +85,64 @@
       <span class="badge-live">Live</span></h3>
       <p>Payments above the approval threshold need a purchase order, an
       invoice, and a payment that agree. Find the mismatches and explain each
-      one.</p></div>
+      one.</p>
+      <!-- example:JE-06 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.83 ± 0.14 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.76 · no-evidence 0.00</span>
+        <a class="ex-link" href="#console/run/c5-grid-je06">View the full run →</a>
+      </div>
+      <!-- /example:JE-06 --></div>
     <div class="task-card"><h3><span class="je-id">JE-07</span> Subledger reconciliation
       <span class="badge-live">Live</span></h3>
       <p>The payables control account doesn't agree with its subledger.
-      Decompose the difference into the specific items that cause it.</p></div>
+      Decompose the difference into the specific items that cause it.</p>
+      <!-- example:JE-07 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.17 ± 0.08 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.80 · no-evidence 0.00</span>
+        <a class="ex-link" href="#console/run/c5-grid-je07">View the full run →</a>
+      </div>
+      <!-- /example:JE-07 --></div>
     <div class="task-card"><h3><span class="je-id">JE-08</span> Accrual reversals
       <span class="badge-live">Live</span></h3>
       <p>Period-end accruals should reverse in the next period. Find the ones
-      that never did, and recompute the ones reversed incorrectly.</p></div>
+      that never did, and recompute the ones reversed incorrectly.</p>
+      <!-- example:JE-08 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.68 ± 0.08 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.79 · no-evidence 0.00</span>
+        <a class="ex-link" href="#console/run/c5-grid-je08">View the full run →</a>
+      </div>
+      <!-- /example:JE-08 --></div>
     <div class="task-card"><h3><span class="je-id">JE-09</span> Interfund transfers
       <span class="badge-live">Live</span></h3>
       <p>A transfer between funds needs a matching counterparty leg. Find the
-      unmatched legs before consolidation.</p></div>
+      unmatched legs before consolidation.</p>
+      <!-- example:JE-09 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.59 ± 0.20 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.63 · no-evidence 0.31</span>
+        <a class="ex-link" href="#console/run/c5-grid-je09">View the full run →</a>
+      </div>
+      <!-- /example:JE-09 --></div>
     <div class="task-card"><h3><span class="je-id">JE-10</span> Budget integrity
       <span class="badge-live">Live</span></h3>
       <p>Obligations should liquidate as payments are made, and spending must
-      stay inside budget authority. Find where either has broken down.</p></div>
+      stay inside budget authority. Find where either has broken down.</p>
+      <!-- example:JE-10 -->
+      <div class="example-run">
+        <span class="ex-model">deepseek-v4-flash</span>
+        <span class="ex-score">final 0.91 ± 0.07 <small>(5 seeds)</small></span>
+        <span class="ex-base">baselines: flag-everything −1.80 · no-evidence 0.25</span>
+        <span class="ex-fail">1 of 5 seeds auto-failed (fabrication)</span>
+        <a class="ex-link" href="#console/run/c5-grid-je10">View the full run →</a>
+      </div>
+      <!-- /example:JE-10 --></div>
   </div>
 
   <div class="tier-head tier-l3"><span class="tier-tag">L3</span><h2>Cross-document evidence</h2>
@@ -121,5 +203,6 @@
 </section>
 
 <footer class="land-footer">
-  <span>Example runs for each task will appear here as recorded runs are published.</span>
+  <span>Recorded 2026-08-11 · open any run in the console's Past runs panel —
+  the same task and seed always reproduces the same case.</span>
 </footer>

app/src/je_validation/console/static/landing.css

@@ -189,6 +189,10 @@
 .tier-l3 .tier-tag { color: var(--l3); background: var(--l3bg); }
 .tier-l4 .tier-tag { color: var(--l4); background: var(--l4bg); }
 .task-grid { display: grid; grid-template-columns: 1fr 1fr; gap: 12px; }
+/* column + auto margin: the example run sits on the card floor, so the score
+ * lines of neighbouring cards align however long the description runs. */
+.task-card { display: flex; flex-direction: column; }
+.task-card > p { margin-bottom: auto; }
 .task-card h3 { display: flex; gap: 8px; align-items: baseline; }
 .task-card .je-id { font-family: var(--mono); font-size: 11px; color: var(--ink-3); }
 .badge-live, .badge-roadmap {
@@ -203,6 +207,55 @@
 .badge-live { color: var(--l1); background: var(--l1bg); }
 .badge-roadmap { color: var(--ink-3); background: var(--panel-3); }
 
+/* ---- example runs on task cards ---- */
+.example-run {
+  margin-top: 14px;
+  padding-top: 10px;
+  border-top: 1px dashed var(--line-2);
+  display: grid;
+  gap: 3px;
+  font-size: 12px;
+}
+.example-run .ex-model { font-family: var(--mono); font-size: 11px; color: var(--ink-3); }
+.example-run .ex-score {
+  font-weight: 700;
+  font-size: 14px;
+  color: var(--ink);
+  font-variant-numeric: tabular-nums;
+}
+.example-run .ex-score small { font-weight: 600; font-size: 11.5px; color: var(--ink-3); }
+.example-run .ex-base { font-family: var(--mono); font-size: 11.5px; color: var(--ink-2); }
+/* a flagged outcome, not an alarm: dot + danger ink, quieter than the link */
+.example-run .ex-fail {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  font-size: 11.5px;
+  font-weight: 600;
+  color: var(--l4);
+}
+.example-run .ex-fail::before {
+  content: "";
+  flex: none;
+  width: 6px;
+  height: 6px;
+  border-radius: 50%;
+  background: currentColor;
+}
+.example-run .ex-link {
+  justify-self: start;     /* hug the text so the focus ring frames the link */
+  margin-top: 4px;
+  font-weight: 700;
+  color: var(--accent);
+  text-decoration: none;
+}
+.example-run .ex-link:hover { text-decoration: underline; }
+/* under ~440px the mono line cannot fit; cap the measure so it breaks at the
+ * separator instead of inside "no-evidence". */
+@media (max-width: 440px) {
+  .example-run .ex-base { max-width: 36ch; }
+}
+
 /* ---- scoring ---- */
 .formula {
   background: var(--panel);

app/src/je_validation/console/static/styles.css

@@ -530,3 +530,26 @@ input[type="checkbox"] { accent-color: var(--accent); }
 .step-detail-row pre { white-space: pre-wrap; word-break: break-word; }
 
 .breakdown { margin-bottom: 14px; }
+
+/* ---- deep-link arrival pulse (Tasks tab -> a specific run) ---- */
+/* Outline-only on the big live panel (a full wash would tint the whole step
+ * log); the compact history row gets the wash too so it reads at a glance. */
+.deep-flash { animation: deep-flash 1.4s ease-out 1; }
+.history-row.deep-flash { animation-name: deep-flash-fill; }
+@keyframes deep-flash {
+  0%   { outline: 2px solid var(--accent); outline-offset: 2px; }
+  60%  { outline-color: var(--accent); }
+  100% { outline: 2px solid transparent; outline-offset: 2px; }
+}
+@keyframes deep-flash-fill {
+  0%   { outline: 2px solid var(--accent); outline-offset: 2px;
+         background-color: var(--accent-soft); }
+  60%  { outline-color: var(--accent); }
+  100% { outline: 2px solid transparent; outline-offset: 2px;
+         background-color: transparent; }
+}
+@media (prefers-reduced-motion: reduce) {
+  /* no motion: a still outline that app.js clears with the class */
+  .deep-flash { animation: none; outline: 2px solid var(--accent);
+                outline-offset: 2px; }
+}

app/src/je_validation/console/static/tabs.js

@@ -1,13 +1,22 @@
 /* Hash-routed tabs. app.js owns everything inside #tab-console and never
- * reads location.hash (verified), so routing is isolated here. */
+ * reads location.hash (verified), so routing is isolated here. A deep link
+ * of the form #console/run/<run_id> opens the console tab and hands the run
+ * id to app.js via a "je:open-run" event (app.js loads first, so its
+ * listener exists before the initial dispatch). */
 (() => {
   const TABS = ["home", "datasets", "tasks", "scoring", "console"];
 
   const current = () => {
-    const h = location.hash.replace("#", "");
+    const h = location.hash.replace("#", "").split("/")[0];
     return TABS.includes(h) ? h : "home";
   };
 
+  const runTarget = () => {
+    const parts = location.hash.replace("#", "").split("/");
+    return parts[0] === "console" && parts[1] === "run" && parts[2]
+      ? parts[2] : null;
+  };
+
   async function show(tab) {
     document.querySelectorAll(".tabbar a").forEach((a) =>
       a.classList.toggle("active", a.dataset.tab === tab));
@@ -27,6 +36,10 @@
       }
     }
     window.scrollTo(0, 0);
+    const run = runTarget();
+    if (tab === "console" && run) {
+      window.dispatchEvent(new CustomEvent("je:open-run", { detail: run }));
+    }
   }
 
   window.addEventListener("hashchange", () => show(current()));

app/src/je_validation/envir/briefs.py

@@ -59,6 +59,28 @@ TASK_RULES: dict[str, str] = {
         "have actually seen in tool results) earns no outcome credit."),
 }
 
+# Per-task reporting requirements (c5): how findings must be REPRESENTED in
+# the submission — flag targets and evidence citations. A versioned output
+# contract, kept separate from TASK_RULES on purpose (sol 2026-08-11): a
+# seed-invariant procedural rule is still not world ontology, and admitting
+# it there would widen the world-rules exception. Same leak test applies:
+# each sentence must be true and identical for every seed and every defect
+# placement — id formats and fixed row conventions only, never per-instance
+# facts. Scored by the JE-10 outcome verifier (flag target, via seeded
+# defect ids) and the JE-10 process gate (citation factor), both c5.
+REPORTING_RULES: dict[str, str] = {
+    "JE-10": (
+        "Reporting requirements: every budgeted account has one BUDGET "
+        "STATUS SUMMARY entry, id 'JE10-BUDGET-<account>'. Raise a "
+        "budget_exceeded exception by flagging that summary entry - not "
+        "the individual payments posted to the account. If you submit "
+        "issue type wrong_document, cite in evidence_ids the obligation "
+        "register's obligating_doc_id field value (the OD-... id you "
+        "compared with the flagged entry); the master:obligation:... "
+        "record key may also be cited as provenance, but it does not "
+        "replace that field-value citation."),
+}
+
 PROMPT_VERSIONS: tuple[str, ...] = ("minimal", "standard", "detailed")
 
 # standard-level additions. The stated rule must match the SCORED rule
@@ -103,6 +125,8 @@ def _minimal(task: dict) -> str:
     parts = [task["brief"]]
     if (rule := TASK_RULES.get(task["id"])) is not None:
         parts.append(rule)
+    if (rep := REPORTING_RULES.get(task["id"])) is not None:
+        parts.append(rep)
     extra = CLAIM_INSTRUCTIONS.get(task["id"])
     if extra is not None:
         parts.append(f"{CLAIM_FORMAT.lstrip()}\n\n{extra}")

app/src/je_validation/scoring/gates.py

@@ -32,6 +32,12 @@ def _is_flag(s: StepRecord) -> bool:
     return s.tool == "disposition" and s.args.get("verdict") in FLAGGING
 
 
+def _dget(d, key):
+    """Disposition field access: production passes state.Disposition objects,
+    tests pass plain dicts (same duality as verifiers._get)."""
+    return d[key] if isinstance(d, dict) else getattr(d, key)
+
+
 def _token_in(token: str, expr: str) -> bool:
     """Whole-numeric-token match: plain substring would let entry amount 100
     "match" expression "1000-900"."""
@@ -309,8 +315,25 @@ def je09_next_period(steps: list[StepRecord], dispositions: dict,
 
 def je10_account_aggregate(steps: list[StepRecord], dispositions: dict,
                            instance: Instance) -> GateResult:
+    """Two per-flag factors, multiplied (standing gate algebra). Budget
+    factor (c4), order-sensitive per spec: a budget_exceeded flag needs an
+    account-level aggregate EARLIER in the episode. Citation factor (c5),
+    order-free over episode-final dispositions: a wrong_document flag is
+    supported only if its disposition cites the obligation register's
+    obligating_doc_id field value for the flagged entry - the reporting
+    requirement REPORTING_RULES['JE-10'] states. Agent-visible data only:
+    the entry header names its obligation_id and the register row carries
+    the correct doc id; a wrong_document flag on an entry with no
+    obligation has no correct doc to cite and is unsupported. Entries are
+    identified by the dispositions dict key - the authoritative id
+    everywhere in scoring."""
     account_of = {e.entry_id: str(e.header.get("account", ""))
                   for e in instance.entries}
+    register = instance.master_data.get("obligation", {})
+    correct_doc_of = {
+        e.entry_id: (register.get(e.header.get("obligation_id")) or {}
+                     ).get("obligating_doc_id")
+        for e in instance.entries}
     grouped = False
     filtered_accounts: set[str] = set()
     b_flagged = b_supported = 0
@@ -328,8 +351,20 @@ def je10_account_aggregate(steps: list[StepRecord], dispositions: dict,
                 b_supported += (grouped or account_of.get(
                     s.args.get("entry_id")) in filtered_accounts)
     b_rate = b_supported / b_flagged if b_flagged else 1.0
-    return GateResult(b_rate, {"budget_flags": b_flagged,
-                               "with_account_aggregate": b_supported})
+    wd_flagged = wd_supported = 0
+    for eid, d in dispositions.items():
+        if (_dget(d, "verdict") in FLAGGING
+                and _dget(d, "issue_type") == "wrong_document"):
+            wd_flagged += 1
+            correct = correct_doc_of.get(eid)
+            wd_supported += (correct is not None
+                             and correct in set(_dget(d, "evidence_ids") or ()))
+    wd_rate = wd_supported / wd_flagged if wd_flagged else 1.0
+    return GateResult(b_rate * wd_rate,
+                      {"budget_flags": b_flagged,
+                       "with_account_aggregate": b_supported,
+                       "wrong_document_flags": wd_flagged,
+                       "with_doc_citation": wd_supported})
 
 
 def je05_rate_recompute(steps: list[StepRecord], dispositions: dict,

app/src/je_validation/scoring/verifiers/je10.py

@@ -1,11 +1,15 @@
-"""JE-10: encumbrance liquidation. F1 x issue accuracy with exact-overage and
-obligating-document-citation conditions folded into issue correctness."""
+"""JE-10: encumbrance liquidation. F1 x issue accuracy with the exact-overage
+condition folded into issue correctness (spec: stated overage must equal
+budget - (actual + encumbrance), exact). The obligating-document citation is
+a REPORTING requirement priced by the process gate (c5,
+gates.je10_account_aggregate) - the spec's outcome line is "exception set and
+type matched against seed", so citation selection never alters outcome."""
 from __future__ import annotations
 
 from collections.abc import Iterable
 
 from je_validation.scoring.claims import claims_of
-from je_validation.scoring.verifiers import OutcomeVerdict, _get, f1_stats, flagged_ids, issue_of
+from je_validation.scoring.verifiers import OutcomeVerdict, f1_stats, flagged_ids, issue_of
 
 
 def verify(dispositions: dict, entry_ids: Iterable[str], heldout) -> OutcomeVerdict:
@@ -24,10 +28,6 @@ def verify(dispositions: dict, entry_ids: Iterable[str], heldout) -> OutcomeVerd
             # spec sign: budget - (actual + encumbrance), negative when over
             claim = claims_of(dispositions, eid).get("available_cents")
             ok = claim == heldout.claims[eid]["available_cents"]
-        if ok and dtype == "wrong_document":
-            cited = set(_get(dispositions[eid], "evidence_ids"))
-            ok = any(x in cited for x in heldout.evidence_map[eid]
-                     if not x.startswith("master:"))
         issue_correct += ok
     fn = len(set(seeded) & ids) - tp
     precision, recall, f1 = f1_stats(tp, fp, fn)

data/runs/console/{run-a5453c3d/baseline_no_evidence/seed3 → c5-grid-je01/deepseek_deepseek-v4-flash/seed1}/score.json

@@ -1,16 +1,16 @@
 {
   "outcome": 0.0,
-  "process": 0.0,
+  "process": 0.5,
   "rationale": null,
-  "combined": 0.0,
+  "combined": 0.05555555555555556,
   "fp_rate": 0.0,
   "fp_penalty": 0.0,
-  "final": 0.0,
+  "final": 0.05555555555555556,
   "fabrications": [],
   "auto_fail": false,
   "hard_fail": false,
   "judge_pass": null,
   "passed": null,
   "judge": "absent",
-  "cost_usd": 0.0
+  "cost_usd": 0.22998800869240005
 }

data/runs/console/{run-a5453c3d/baseline_no_evidence/seed5 → c5-grid-je01/deepseek_deepseek-v4-flash/seed2}/score.json

@@ -1,16 +1,16 @@
 {
   "outcome": 0.0,
-  "process": 0.0,
+  "process": 0.495,
   "rationale": null,
-  "combined": 0.0,
+  "combined": 0.055,
   "fp_rate": 0.0,
   "fp_penalty": 0.0,
-  "final": 0.0,
+  "final": 0.055,
   "fabrications": [],
   "auto_fail": false,
   "hard_fail": false,
   "judge_pass": null,
   "passed": null,
   "judge": "absent",
-  "cost_usd": 0.0
+  "cost_usd": 0.45799382661999993
 }

data/runs/console/c5-grid-je01/deepseek_deepseek-v4-flash/seed3/score.json

@@ -0,0 +1,16 @@
+{
+  "outcome": 0.0,
+  "process": 0.3333333333333333,
+  "rationale": null,
+  "combined": 0.037037037037037035,
+  "fp_rate": 0.0,
+  "fp_penalty": 0.0,
+  "final": 0.037037037037037035,
+  "fabrications": [],
+  "auto_fail": false,
+  "hard_fail": false,
+  "judge_pass": null,
+  "passed": null,
+  "judge": "absent",
+  "cost_usd": 0.39998567448000016
+}

data/runs/console/c5-grid-je01/deepseek_deepseek-v4-flash/seed4/score.json

@@ -0,0 +1,16 @@
+{
+  "outcome": 0.0,
+  "process": 0.16666666666666666,
+  "rationale": null,
+  "combined": 0.018518518518518517,
+  "fp_rate": 0.0,
+  "fp_penalty": 0.0,
+  "final": 0.018518518518518517,
+  "fabrications": [],
+  "auto_fail": false,
+  "hard_fail": false,
+  "judge_pass": null,
+  "passed": null,
+  "judge": "absent",
+  "cost_usd": 0.36706333776000016
+}

data/runs/console/c5-grid-je01/deepseek_deepseek-v4-flash/seed5/score.json

@@ -0,0 +1,16 @@
+{
+  "outcome": 0.0,
+  "process": 0.14285714285714285,
+  "rationale": null,
+  "combined": 0.015873015873015872,
+  "fp_rate": 0.0,
+  "fp_penalty": 0.0,
+  "final": 0.015873015873015872,
+  "fabrications": [],
+  "auto_fail": false,
+  "hard_fail": false,
+  "judge_pass": null,
+  "passed": null,
+  "judge": "absent",
+  "cost_usd": 0.38351098015999985
+}

data/runs/console/{run-a5453c3d → c5-grid-je01}/results.json

@@ -8,130 +8,29 @@
         "std": 0.0
       },
       "process": {
-        "mean": 0.6870666666666667,
-        "std": 0.16309751684191887
+        "mean": 0.32757142857142857,
+        "std": 0.17161501105130808
       },
       "rationale": "pending judge calibration",
       "combined": {
-        "mean": 0.12124705882352942,
-        "std": 0.028781914736809213
+        "mean": 0.03639682539682539,
+        "std": 0.019068334561256454
       },
       "final": {
-        "mean": 0.12124705882352942,
-        "std": 0.028781914736809213
+        "mean": 0.03639682539682539,
+        "std": 0.019068334561256454
       },
       "fail_rate": 0.0
     },
     "reports": [
-      {
-        "outcome": 0.0,
-        "process": 0.6666666666666666,
-        "rationale": "pending judge calibration",
-        "combined": 0.11764705882352941,
-        "fp_rate": 0.02,
-        "fp_penalty": 0.0,
-        "final": 0.11764705882352941,
-        "fabrications": [],
-        "auto_fail": false,
-        "hard_fail": false,
-        "judge_pass": null,
-        "passed": null,
-        "judge": "absent"
-      },
       {
         "outcome": 0.0,
         "process": 0.5,
         "rationale": "pending judge calibration",
-        "combined": 0.08823529411764706,
-        "fp_rate": 0.02,
-        "fp_penalty": 0.0,
-        "final": 0.08823529411764706,
-        "fabrications": [],
-        "auto_fail": false,
-        "hard_fail": false,
-        "judge_pass": null,
-        "passed": null,
-        "judge": "absent"
-      },
-      {
-        "outcome": 0.0,
-        "process": 0.9500000000000001,
-        "rationale": "pending judge calibration",
-        "combined": 0.16764705882352943,
-        "fp_rate": 0.02,
-        "fp_penalty": 0.0,
-        "final": 0.16764705882352943,
-        "fabrications": [],
-        "auto_fail": false,
-        "hard_fail": false,
-        "judge_pass": null,
-        "passed": null,
-        "judge": "absent"
-      },
-      {
-        "outcome": 0.0,
-        "process": 0.6466666666666667,
-        "rationale": "pending judge calibration",
-        "combined": 0.11411764705882353,
-        "fp_rate": 0.02,
-        "fp_penalty": 0.0,
-        "final": 0.11411764705882353,
-        "fabrications": [],
-        "auto_fail": false,
-        "hard_fail": false,
-        "judge_pass": null,
-        "passed": null,
-        "judge": "absent"
-      },
-      {
-        "outcome": 0.0,
-        "process": 0.672,
-        "rationale": "pending judge calibration",
-        "combined": 0.11858823529411765,
-        "fp_rate": 0.06,
-        "fp_penalty": 0.0,
-        "final": 0.11858823529411765,
-        "fabrications": [],
-        "auto_fail": false,
-        "hard_fail": false,
-        "judge_pass": null,
-        "passed": null,
-        "judge": "absent"
-      }
-    ]
-  },
-  "openai/gpt-5.6-sol": {
-    "aggregate": {
-      "n": 5,
-      "variance_measurable": true,
-      "outcome": {
-        "mean": 0.0,
-        "std": 0.0
-      },
-      "process": {
-        "mean": 0.8142857142857143,
-        "std": 0.18454300975376878
-      },
-      "rationale": "pending judge calibration",
-      "combined": {
-        "mean": 0.14369747899159663,
-        "std": 0.0325664134859592
-      },
-      "final": {
-        "mean": 0.14369747899159663,
-        "std": 0.0325664134859592
-      },
-      "fail_rate": 0.0
-    },
-    "reports": [
-      {
-        "outcome": 0.0,
-        "process": 0.75,
-        "rationale": "pending judge calibration",
-        "combined": 0.1323529411764706,
-        "fp_rate": 0.08,
+        "combined": 0.05555555555555556,
+        "fp_rate": 0.0,
         "fp_penalty": 0.0,
-        "final": 0.1323529411764706,
+        "final": 0.05555555555555556,
         "fabrications": [],
         "auto_fail": false,
         "hard_fail": false,
@@ -141,12 +40,12 @@
       },
       {
         "outcome": 0.0,
-        "process": 0.5714285714285714,
+        "process": 0.495,
         "rationale": "pending judge calibration",
-        "combined": 0.10084033613445377,
-        "fp_rate": 0.04,
+        "combined": 0.055,
+        "fp_rate": 0.0,
         "fp_penalty": 0.0,
-        "final": 0.10084033613445377,
+        "final": 0.055,
         "fabrications": [],
         "auto_fail": false,
         "hard_fail": false,
@@ -156,12 +55,12 @@
       },
       {
         "outcome": 0.0,
-        "process": 1.0,
+        "process": 0.3333333333333333,
         "rationale": "pending judge calibration",
-        "combined": 0.17647058823529413,
-        "fp_rate": 0.02,
+        "combined": 0.037037037037037035,
+        "fp_rate": 0.0,
         "fp_penalty": 0.0,
-        "final": 0.17647058823529413,
+        "final": 0.037037037037037035,
         "fabrications": [],
         "auto_fail": false,
         "hard_fail": false,
@@ -171,12 +70,12 @@
       },
       {
         "outcome": 0.0,
-        "process": 0.75,
+        "process": 0.16666666666666666,
         "rationale": "pending judge calibration",
-        "combined": 0.1323529411764706,
-        "fp_rate": 0.08,
+        "combined": 0.018518518518518517,
+        "fp_rate": 0.0,
         "fp_penalty": 0.0,
-        "final": 0.1323529411764706,
+        "final": 0.018518518518518517,
         "fabrications": [],
         "auto_fail": false,
         "hard_fail": false,
@@ -186,12 +85,12 @@
       },
       {
         "outcome": 0.0,
-        "process": 1.0,
+        "process": 0.14285714285714285,
         "rationale": "pending judge calibration",
-        "combined": 0.17647058823529413,
+        "combined": 0.015873015873015872,
         "fp_rate": 0.0,
         "fp_penalty": 0.0,
-        "final": 0.17647058823529413,
+        "final": 0.015873015873015872,
         "fabrications": [],
         "auto_fail": false,
         "hard_fail": false,

data/runs/console/{run-a5453c3d → c5-grid-je01}/run.json

@@ -1,18 +1,17 @@
 {
-  "run_id": "run-a5453c3d",
+  "run_id": "c5-grid-je01",
   "status": "finished",
   "request": {
-    "run_id": "run-a5453c3d",
-    "task_id": "JE-05",
+    "run_id": "c5-grid-je01",
+    "task_id": "JE-01",
     "models": [
       "deepseek/deepseek-v4-flash",
-      "openai/gpt-5.6-sol",
       "baseline:flag_everything",
       "baseline:no_evidence"
     ],
     "seed_count": 5,
     "knobs": {
-      "population": 60,
+      "population": 400,
       "defect_density": 28,
       "distractor_ratio": 2,
       "doc_noise": 10
@@ -20,56 +19,49 @@
     "tools_enabled": [
       "query_ledger",
       "get_entry",
-      "get_master_data",
       "recompute",
       "disposition",
       "submit"
     ],
-    "step_budget": 270,
-    "token_ceiling": 229500000,
+    "step_budget": 1800,
+    "token_ceiling": 1530000000,
     "pricing": {
       "deepseek/deepseek-v4-flash": {
-        "prompt_price": 8.82e-08,
-        "completion_price": 1.764e-07
-      },
-      "openai/gpt-5.6-sol": {
-        "prompt_price": 5e-06,
-        "completion_price": 3e-05
+        "prompt_price": 1.4e-07,
+        "completion_price": 2.8e-07
       }
     },
     "prompt_version": "standard"
   },
   "contract": {
-    "task_id": "JE-05",
+    "task_id": "JE-01",
     "tier": "L1",
     "seed": 1,
     "generator": {
       "knobs": {
-        "population": 60,
+        "population": 400,
         "defect_density": 28,
         "distractor_ratio": 2,
         "doc_noise": 10
       },
       "sources": [
-        "DS-A",
-        "DS-J"
+        "DS-A"
       ]
     },
     "tools_enabled": [
       "query_ledger",
       "get_entry",
-      "get_master_data",
       "recompute",
       "disposition",
       "submit"
     ],
     "tools_withheld": [],
-    "step_budget": 270,
+    "step_budget": 1800,
     "scoring": {
       "weights": {
-        "outcome": 0.7,
-        "process": 0.15,
-        "rationale": 0.15
+        "outcome": 0.8,
+        "process": 0.1,
+        "rationale": 0.1
       },
       "judge": "global_rubric_v1",
       "judges_per_episode": 1
@@ -77,12 +69,11 @@
     "sweep": {
       "models": [
         "deepseek/deepseek-v4-flash",
-        "openai/gpt-5.6-sol",
         "baseline:flag_everything",
         "baseline:no_evidence"
       ],
       "seeds": 5,
-      "episodes": 20
+      "episodes": 15
     },
     "prompt_version": "standard"
   },
@@ -92,11 +83,6 @@
     "deepseek_deepseek-v4-flash/seed3": "finished",
     "deepseek_deepseek-v4-flash/seed4": "finished",
     "deepseek_deepseek-v4-flash/seed5": "finished",
-    "openai_gpt-5.6-sol/seed1": "finished",
-    "openai_gpt-5.6-sol/seed2": "finished",
-    "openai_gpt-5.6-sol/seed3": "finished",
-    "openai_gpt-5.6-sol/seed4": "finished",
-    "openai_gpt-5.6-sol/seed5": "finished",
     "baseline_flag_everything/seed1": "finished",
     "baseline_flag_everything/seed2": "finished",
     "baseline_flag_everything/seed3": "finished",
@@ -108,5 +94,5 @@
     "baseline_no_evidence/seed4": "finished",
     "baseline_no_evidence/seed5": "finished"
   },
-  "cost_usd": 12.381930227651985
+  "cost_usd": 1.838541827712399
 }

data/runs/console/c5-grid-je01/seed1/bundle/heldout/defect_ledger.json

@@ -0,0 +1 @@
+{"claims":{},"defects":[{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0000219389","revealing_evidence_id":"2024-0000219389:lines"},{"correct_verdict":"flag","defect_type":"dup_line_seq","entry_id":"2024-0018941205","revealing_evidence_id":"2024-0018941205:lines"},{"correct_verdict":"flag","defect_type":"orphan_leg","entry_id":"2024-0018944498","revealing_evidence_id":"2024-0018944498:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018945399","revealing_evidence_id":"2024-0018945399:lines"},{"correct_verdict":"flag","defect_type":"orphan_leg","entry_id":"2024-0018950877","revealing_evidence_id":"2024-0018950877:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018952717","revealing_evidence_id":"2024-0018952717:lines"},{"correct_verdict":"flag","defect_type":"orphan_leg","entry_id":"2024-0018953134","revealing_evidence_id":"2024-0018953134:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018955782","revealing_evidence_id":"2024-0018955782:lines"},{"correct_verdict":"flag","defect_type":"zero_amount_line","entry_id":"2024-0018958794","revealing_evidence_id":"2024-0018958794:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018966176","revealing_evidence_id":"2024-0018966176:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018967117","revealing_evidence_id":"2024-0018967117:lines"},{"correct_verdict":"flag","defect_type":"zero_amount_line","entry_id":"2024-0018972058","revealing_evidence_id":"2024-0018972058:lines"}],"evidence_map":{"2024-0000219389":["2024-0000219389:lines"],"2024-0018941205":["2024-0018941205:lines"],"2024-0018944498":["2024-0018944498:lines"],"2024-0018945399":["2024-0018945399:lines"],"2024-0018950877":["2024-0018950877:lines"],"2024-0018952717":["2024-0018952717:lines"],"2024-0018953134":["2024-0018953134:lines"],"2024-0018955782":["2024-0018955782:lines"],"2024-0018958794":["2024-0018958794:lines"],"2024-0018966176":["2024-0018966176:lines"],"2024-0018967117":["2024-0018967117:lines"],"2024-0018972058":["2024-0018972058:lines"]},"required_docs":[]}

data/runs/console/c5-grid-je01/seed1/bundle/instance/meta.json

@@ -0,0 +1 @@
+{"seed":1,"task_id":"JE-01"}

data/runs/console/c5-grid-je01/seed1/bundle/manifest.json

@@ -0,0 +1 @@
+{"generator_version":"4c.1","knobs":{"n_entries":400},"seed":1,"snapshot_id":"2026-08-04","spec_sha256":"90d55c077a0acea9a57b95ecd63a70df0316fc271d873a38afa195a766f23978","task_id":"JE-01"}

data/runs/console/c5-grid-je01/seed2/bundle/heldout/defect_ledger.json

@@ -0,0 +1 @@
+{"claims":{},"defects":[{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018942198","revealing_evidence_id":"2024-0018942198:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018944317","revealing_evidence_id":"2024-0018944317:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018946773","revealing_evidence_id":"2024-0018946773:lines"},{"correct_verdict":"flag","defect_type":"orphan_leg","entry_id":"2024-0018950131","revealing_evidence_id":"2024-0018950131:lines"},{"correct_verdict":"flag","defect_type":"orphan_leg","entry_id":"2024-0018951225","revealing_evidence_id":"2024-0018951225:lines"},{"correct_verdict":"flag","defect_type":"zero_amount_line","entry_id":"2024-0018952858","revealing_evidence_id":"2024-0018952858:lines"},{"correct_verdict":"flag","defect_type":"orphan_leg","entry_id":"2024-0018960305","revealing_evidence_id":"2024-0018960305:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018965646","revealing_evidence_id":"2024-0018965646:lines"},{"correct_verdict":"flag","defect_type":"dup_line_seq","entry_id":"2024-0018965792","revealing_evidence_id":"2024-0018965792:lines"},{"correct_verdict":"flag","defect_type":"zero_amount_line","entry_id":"2024-0018969623","revealing_evidence_id":"2024-0018969623:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018969733","revealing_evidence_id":"2024-0018969733:lines"},{"correct_verdict":"flag","defect_type":"unbalanced","entry_id":"2024-0018969865","revealing_evidence_id":"2024-0018969865:lines"}],"evidence_map":{"2024-0018942198":["2024-0018942198:lines"],"2024-0018944317":["2024-0018944317:lines"],"2024-0018946773":["2024-0018946773:lines"],"2024-0018950131":["2024-0018950131:lines"],"2024-0018951225":["2024-0018951225:lines"],"2024-0018952858":["2024-0018952858:lines"],"2024-0018960305":["2024-0018960305:lines"],"2024-0018965646":["2024-0018965646:lines"],"2024-0018965792":["2024-0018965792:lines"],"2024-0018969623":["2024-0018969623:lines"],"2024-0018969733":["2024-0018969733:lines"],"2024-0018969865":["2024-0018969865:lines"]},"required_docs":[]}