the BELLS selection question is recorded as sent

distribution/QUEUE.md

@@ -11,10 +11,10 @@ message was sent during reconciliation.
 | 3 | GuardBench joint-reporter ask | **WAIT until September 9**, one week after IBM. Recheck the target and duplicate issues before sending the existing [dossier](dossiers/guardbench-2024.md) with [reporter](../contrib/guardbench_joint.py). No matching Cubits11 request found among the public issues inspected September 5. | A reply, accepted patch, or usable joint-results release. |
 | 4 | Dispatch `x-film-same-scores` | **HOLD for the cold-viewer gate below and account baseline.** Current `campaigns.yaml` requires **5–8** cold viewers, not three. Then re-derive numerals at the dispatch commit; post the existing 11-second square cut and registered copy once. Record its actual permalink/time and platform denominators; never invent a dispatch date. | Verified independent reproduction. Link clicks / impressions are a verification-intent proxy, not observed `/try/` landings. |
 | 5 | Run the frozen Same Scores comprehension trial | **NEXT HUMAN ACTION.** Five minimum, eight maximum, people unfamiliar with the project. One viewing, sound off, no explanation; collect all three answers verbatim before scoring. Rules remain in `launch-units.yaml → cold_test_gate`. Two failures on the same question hold release; revise once and use a fresh cold set, never rescore. | Real answers and pass/fail diagnostics. No trial has been fabricated or recorded as completed. |
-| 6 | BELLS source-selection question | **PREPARED, separate weekly card.** Inspect the actual author announcement and existing replies first. Respect the current PRESENT-first selection rule; do not strengthen the frozen row or publish private denominator findings through an outreach shortcut. | A checkable source confirmation or correction. |
+| 6 | BELLS source-selection question | **SENT September 16:** [BELLS issue #1](https://github.com/CentreSecuriteIA/bells_leaderboard/issues/1). Exact posted text recorded in the dossier and dispatch log; awaiting an answer about selection. Do not duplicate. | A checkable source confirmation or correction. |
 | 7 | Subsequent film dispatches | **WAIT for Same Scores' actual T+14d decision.** Then change one representation at a time, at least 14 days apart, under the existing rule. No fixed dates until the first dispatch exists. | Reproduction or joint release; attention stays diagnostic. |
 | 8 | Process an external outcome | Follow [EXTERNAL_EVENTS.md](EXTERNAL_EVENTS.md) when a real artifact arrives. The three existing tracked issues had no comments at the September 5 check. An issue opened by the author is not a qualified outcome. | Verified entry, relevant regeneration, credit, and any required correction. |
-| 9 | Apply the stop rule | **NOT TRIGGERED: 3/12 technical interactions, 0 qualified outcomes.** The third is IBM's already-sent request, recorded retrospectively. Nine interactions remain. Do not infer unobserved funnel stages. | A supported diagnosis, with exactly one changed variable when the rule fires. |
+| 9 | Apply the stop rule | **NOT TRIGGERED: 4/12 technical interactions, 0 qualified outcomes.** The third is IBM's already-sent request; the fourth is the BELLS selection question. Eight interactions remain. Do not infer unobserved funnel stages. | A supported diagnosis, with exactly one changed variable when the rule fires. |
 | 10 | Cohort B films | **HOLD.** Only the existing A–D evidence gates in the E2 brief justify further films. | The evidence that opens a gate, not another film. |
 
 ## Dates that can actually be scheduled

distribution/dispatch-log.yaml

@@ -43,3 +43,20 @@ dispatches:
       - {at: "T+7d",  rule: "same snapshot; no second film until the T+14d decision"}
       - {at: "T+14d", rule: "apply the campaign's decision rule; silence is NO_OBSERVED_RESPONSE"}
     outcome: null
+
+  - id: bells-selection-rule
+    sent_at_commit: null  # owner browser submission; checkout at send time not established
+    reproduction_commit: 4b659fb97bf50aff9793ff4a26c3410b7cf0a53c
+    adjudication:
+      - {on: substantive_response, rule: "Verify selection rule or correction against the cited artifact before awarding outcome credit."}
+      - {on: silence, rule: "Retain NO_OBSERVED_RESPONSE; do not duplicate the issue."}
+    dossier: distribution/dossiers/bells-misuse-2025.md
+    target: "CentreSecuriteIA/bells_leaderboard — issue tracker"
+    status: SENT_PENDING_RESPONSE
+    sent_by: owner
+    reported: "2026-09-16"
+    sent_at_utc: "2026-09-17T03:19:51Z"
+    permalink: https://github.com/CentreSecuriteIA/bells_leaderboard/issues/1
+    observed_comments: 0
+    message_as_sent: "Thank you for releasing per-item supervisor verdicts in `data/non_adversarial_prompts.csv`.\n\nAt commit `507566c5a4606c8e3dec0bd59a5c5fde62594951` the file has 170 rows, 82 of them with `harm_level` = `harmful`. On 9 of those 82 rows, all five specialized-supervisor columns (`lakera_guard`, `prompt_guard`, `langkit`, `nemo`, `llm_guard`) are 0. This script downloads the file at that commit, checks its SHA-256, and recomputes the count: https://github.com/Cubits11/cubits11.github.io/blob/4b659fb97bf50aff9793ff4a26c3410b7cf0a53c/scripts/reanalyze_bells_subset.py\n\nYour FAQ and the paper's Dataset Access appendix call the playground rows \"representative examples\". How were these 170 prompts selected from the full non-adversarial set? A sampling rule, a filter, manual choice, or a link to an existing description would answer it.\n\nI report 9/82 only as a count on these 170 released rows, not as an estimate for the full benchmark or for any deployed system. If I have misread the release, a correction is just as useful."
+    outcome: null

distribution/dossiers/bells-misuse-2025.md

@@ -1,6 +1,8 @@
 # Dossier — bells-misuse-2025
 
-Row: `bells-misuse-2025` · classification PRESENT (computable_via_item_release) · reconstruction PARTIALLY_IDENTIFIED · status: PREPARED, nothing sent.
+Row: `bells-misuse-2025` · classification PRESENT (computable_via_item_release) · reconstruction PARTIALLY_IDENTIFIED · status: SENT September 16, 2026 (America/New_York), RESPONSE PENDING.
+
+Issue: https://github.com/CentreSecuriteIA/bells_leaderboard/issues/1 · sent 2026-09-17T03:19:51Z · zero comments when verified.
 
 **1. What the source publishes.** Per-supervisor results for twelve systems, and a released per-item subset: 170 non-adversarial prompts (82 harmful, 50 benign, 38 borderline) with binary verdict columns for five specialized supervisors, at a pinned commit (`507566c5…`, hash-verified in CI). Nothing joint is printed; the headline population's per-item outcomes and the ~4,165 adversarial prompts have no per-item release. The subset's selection rule is unstated.
 
@@ -18,8 +20,14 @@ Row: `bells-misuse-2025` · classification PRESENT (computable_via_item_release)
 
 **8. Correction condition.** Evidence that a released column is not the labelled system's verdict, that the subset was released with a stated rule the row misreports, or that a fuller per-item release existed on or before 2026-08-27. Any of these corrects MC-002 or the row under their registered falsifiers.
 
-**The ask, verbatim:**
+**The ask, verbatim (as posted):**
 
-> Your released 170-prompt subset lets anyone recompute what the five specialized supervisors miss *together*: 9 of 82 harmful prompts, against an independence plug-in of 2.87 — and the leave-one-out unions show three of the five add no exclusive coverage on that stratum. One question so the row about your work is right: how were the 170 prompts selected? If the row is wrong, say so — corrections land dated in the public file.
+> Thank you for releasing per-item supervisor verdicts in `data/non_adversarial_prompts.csv`.
+>
+> At commit `507566c5a4606c8e3dec0bd59a5c5fde62594951` the file has 170 rows, 82 of them with `harm_level` = `harmful`. On 9 of those 82 rows, all five specialized-supervisor columns (`lakera_guard`, `prompt_guard`, `langkit`, `nemo`, `llm_guard`) are 0. This script downloads the file at that commit, checks its SHA-256, and recomputes the count: https://github.com/Cubits11/cubits11.github.io/blob/4b659fb97bf50aff9793ff4a26c3410b7cf0a53c/scripts/reanalyze_bells_subset.py
+>
+> Your FAQ and the paper's Dataset Access appendix call the playground rows "representative examples". How were these 170 prompts selected from the full non-adversarial set? A sampling rule, a filter, manual choice, or a link to an existing description would answer it.
+>
+> I report 9/82 only as a count on these 170 released rows, not as an estimate for the full benchmark or for any deployed system. If I have misread the release, a correction is just as useful.
 
 **Channel.** GitHub issues on `CentreSecuriteIA/bells_leaderboard` (the row's recorded route). One message.

distribution/outcomes.yaml

@@ -46,8 +46,16 @@ diagnostics:
   cold_comprehension_trials: []
   # An invitation or open issue is diagnostic only; it is never credited as
   # an external reproduction, correction, release, PR, or cold run.
-  technical_interactions: 3     # replies that carried a number, source, or correction
+  technical_interactions: 4     # replies that carried a number, source, or correction
   technical_interaction_log:
+    - date: "2026-09-16"
+      kind: upstream_issue
+      url: https://github.com/CentreSecuriteIA/bells_leaderboard/issues/1
+      summary: >
+        Owner asked how the 170 released BELLS rows were selected, acknowledging
+        the published representative-examples description. Posted September 16
+        in America/New_York (September 17 UTC). Open with no comments when verified;
+        diagnostic only, not an independent reproduction or source correction.
     - date: "2026-09-02"
       kind: upstream_issue
       url: https://github.com/IBM/Adversarial-Prompt-Evaluation/issues/7

distribution/traction/events.json

@@ -52,13 +52,13 @@
     "novelty": "snapshot; not asserted newly published"
   },
   {
-    "id": "70050c08430f7c32",
+    "id": "c93d03ba49aff94d",
     "kind": "metrics",
     "source": {
       "path": "distribution/outcomes.yaml",
-      "sha256": "2cefa442652376b870e95f86840e5716cc91da5ea91f9be46585d2a19ff23b66",
-      "commit": "26a9d578ffe77deecc191f02ae578e6f8ae6e2c0",
-      "url": "https://github.com/Cubits11/cubits11.github.io/blob/26a9d578ffe77deecc191f02ae578e6f8ae6e2c0/distribution/outcomes.yaml",
+      "sha256": "91f5decd619427fdb700ce5a458f73453402fc06c9b25a5c1ed039a74c64ef38",
+      "commit": "baf018bfd3997716cc8800f51861bb81edd98652",
+      "url": "https://github.com/Cubits11/cubits11.github.io/blob/baf018bfd3997716cc8800f51861bb81edd98652/distribution/outcomes.yaml",
       "binding": "committed"
     },
     "draft_ids": [],

distribution/traction/manifest.json

@@ -2,7 +2,7 @@
   "version": 1,
   "generator": "scripts/distribute.py",
   "outputs": {
-    "events.json": "5f7baafe4f8a98f8823333777efbf0899ef1375a6e18c23e66abd78a062fe0a1",
+    "events.json": "343833167a7794e4269e40c21868b52fcd3055d9ce358c593c1646cf3538e8e6",
     "drafts.json": "6916cc6433575eace3023a124bcbf93ecffe3ac5ce6637e6b4661c6ff1128419",
     "queue.json": "ca0558ccf563ec99695cfc68a39a2a5c7e19a9c33dbaf232f09689c589bf7789",
     "dashboard.json": "988051f7c9d5b82684071ab29fdb52a891ce0f93470405f402cc6beff62d04bb",

docs/RESEARCH_INDEX.md

@@ -13,7 +13,7 @@ Registry v0.4 · last owner review 2026-09-14 · one question: *Do published per
 | paired outcome releases | 0 |
 | upstream prs | 0 |
 | human cold runs | 0 |
-| technical interactions (diagnostic, not an outcome) | 3 |
+| technical interactions (diagnostic, not an outcome) | 4 |
 
 Zero is the recorded value where it is zero. The stop rule and the procedure for recording an outcome are in `distribution/EXTERNAL_EVENTS.md`.
 

docs/graph/repo-graph.json

@@ -4902,66 +4902,17 @@
    "severity": "RECORD",
    "sha256": "22e956020cdfdd43c328f02441f003c933ee4e7083d02d9d24e414e4b42ae2e6"
   },
-  {
-   "ahead": 15,
-   "behind": 0,
-   "branch": "claude/inferential-adequacy",
-   "detector": "D4",
-   "severity": "TOPOLOGY"
-  },
   {
    "ahead": 1,
    "behind": 0,
-   "branch": "claude/issue-forms-ingress",
-   "detector": "D4",
-   "severity": "TOPOLOGY"
-  },
-  {
-   "ahead": 17,
-   "behind": 0,
-   "branch": "claude/joint-cell-visible",
-   "detector": "D4",
-   "severity": "TOPOLOGY"
-  },
-  {
-   "ahead": 14,
-   "behind": 0,
-   "branch": "claude/seo-falsifiable-web",
-   "detector": "D4",
-   "severity": "TOPOLOGY"
-  },
-  {
-   "ahead": 17,
-   "behind": 0,
-   "branch": "claude/stack-merge",
-   "detector": "D4",
-   "severity": "TOPOLOGY"
-  },
-  {
-   "ahead": 15,
-   "behind": 0,
-   "branch": "origin/claude/inferential-adequacy",
+   "branch": "claude/bells-sent-recorded",
    "detector": "D4",
    "severity": "TOPOLOGY"
   },
   {
    "ahead": 1,
    "behind": 0,
-   "branch": "origin/claude/issue-forms-ingress",
-   "detector": "D4",
-   "severity": "TOPOLOGY"
-  },
-  {
-   "ahead": 17,
-   "behind": 0,
-   "branch": "origin/claude/joint-cell-visible",
-   "detector": "D4",
-   "severity": "TOPOLOGY"
-  },
-  {
-   "ahead": 14,
-   "behind": 0,
-   "branch": "origin/claude/seo-falsifiable-web",
+   "branch": "origin/claude/bells-sent-recorded",
    "detector": "D4",
    "severity": "TOPOLOGY"
   },
@@ -4973,8 +4924,8 @@
   }
  ],
  "generated_from": {
-  "branch": "claude/stack-merge",
-  "head": "4173a8c"
+  "branch": "claude/bells-sent-recorded",
+  "head": "baf018b"
  },
  "nodes": [
   {
@@ -5825,110 +5776,137 @@
    "type": "experiment"
   },
   {
-   "ahead_of_origin_main": 15,
+   "ahead_of_origin_main": 1,
    "behind_origin_main": 0,
+   "head": "baf018b",
+   "id": "claude/bells-sent-recorded",
+   "last_commit": "2026-09-17",
+   "reachable_from_origin_main": false,
+   "type": "branch"
+  },
+  {
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 5,
    "head": "6a51abb",
    "id": "claude/inferential-adequacy",
    "last_commit": "2026-09-17",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 1,
-   "behind_origin_main": 0,
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 19,
    "head": "026554c",
    "id": "claude/issue-forms-ingress",
    "last_commit": "2026-09-17",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 17,
-   "behind_origin_main": 0,
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 3,
    "head": "4173a8c",
    "id": "claude/joint-cell-visible",
    "last_commit": "2026-09-17",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 14,
-   "behind_origin_main": 0,
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 6,
    "head": "5fe44bc",
    "id": "claude/seo-falsifiable-web",
    "last_commit": "2026-09-15",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 17,
-   "behind_origin_main": 0,
-   "head": "4173a8c",
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 1,
+   "head": "cc7c7ec",
    "id": "claude/stack-merge",
    "last_commit": "2026-09-17",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
    "ahead_of_origin_main": 0,
    "behind_origin_main": 0,
-   "head": "4b659fb",
+   "head": "d3766ae",
    "id": "main",
-   "last_commit": "2026-09-07",
+   "last_commit": "2026-09-17",
    "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
    "ahead_of_origin_main": 0,
    "behind_origin_main": 0,
-   "head": "4b659fb",
+   "head": "d3766ae",
    "id": "origin",
-   "last_commit": "2026-09-07",
+   "last_commit": "2026-09-17",
    "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 15,
+   "ahead_of_origin_main": 1,
    "behind_origin_main": 0,
+   "head": "baf018b",
+   "id": "origin/claude/bells-sent-recorded",
+   "last_commit": "2026-09-17",
+   "reachable_from_origin_main": false,
+   "type": "branch"
+  },
+  {
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 5,
    "head": "6a51abb",
    "id": "origin/claude/inferential-adequacy",
    "last_commit": "2026-09-17",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 1,
-   "behind_origin_main": 0,
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 19,
    "head": "026554c",
    "id": "origin/claude/issue-forms-ingress",
    "last_commit": "2026-09-17",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 17,
-   "behind_origin_main": 0,
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 3,
    "head": "4173a8c",
    "id": "origin/claude/joint-cell-visible",
    "last_commit": "2026-09-17",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
-   "ahead_of_origin_main": 14,
-   "behind_origin_main": 0,
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 6,
    "head": "5fe44bc",
    "id": "origin/claude/seo-falsifiable-web",
    "last_commit": "2026-09-15",
-   "reachable_from_origin_main": false,
+   "reachable_from_origin_main": true,
+   "type": "branch"
+  },
+  {
+   "ahead_of_origin_main": 0,
+   "behind_origin_main": 1,
+   "head": "cc7c7ec",
+   "id": "origin/claude/stack-merge",
+   "last_commit": "2026-09-17",
+   "reachable_from_origin_main": true,
    "type": "branch"
   },
   {
    "ahead_of_origin_main": 0,
    "behind_origin_main": 0,
-   "head": "4b659fb",
+   "head": "d3766ae",
    "id": "origin/main",
-   "last_commit": "2026-09-07",
+   "last_commit": "2026-09-17",
    "reachable_from_origin_main": true,
    "type": "branch"
   },

films/flagship/source-map.json

@@ -17,7 +17,7 @@
     "films/the-falsifier-that-fired/manifest.yaml": "89814cb92b88aa87ab546798f737a9d8ff3a4a80548d8375cd67d2b3d3148acd",
     "scripts/reanalyze_bells_subset.py": "c39643ca63146931aaaa8614477453adcd5d411501d251d33d12bdb811759a88",
     "films/thirteen-worlds/manifest.yaml": "351adf9ccbd42f54c53645251378aaeb0255e5507874cfe78d59009476d79828",
-    "missing-column/reproduce/index.html": "80280667a8e99287856837e17586b12d63106c2790107395761ffb17863ff29f",
+    "missing-column/reproduce/index.html": "1dcdcb4f7c87b1e9968a2435036a3c49a253baef911eecfe050832279e49ca92",
     "films/the-correction-invitation/manifest.yaml": "9b87c56e01095ea1db9152d67c12f193afc670a0ed44a54bc664652b5680602d"
   },
   "sections": [

metrics/ledger_snapshot.json

@@ -85,7 +85,7 @@
   "outcomes": {
     "qualified_total": 0,
     "categories": 5,
-    "technical_interactions": 3,
+    "technical_interactions": 4,
     "stop_threshold": 12
   }
 }

missing-column/reproduce/index.html

@@ -217,7 +217,7 @@ <h2 id="file-h">File your run — match or mismatch</h2>
       <a class="u" href="mailto:bhavepranavwork@gmail.com">email</a>. The project-wide
       <a class="u" href="/distribution/outcomes.yaml">qualified-outcome ledger</a> currently
       records 0 qualified external outcomes;
-      it also records 3 technical interactions,
+      it also records 4 technical interactions,
       which is diagnostic rather than evidence. That public null is not a growth metric. MC-004 has no recorded independent
       reproduction — this page is the standing invitation, not evidence that one exists.</p>
   </section>