Koth ladder

.github/workflows/koth-engine-gate.yml

@@ -38,7 +38,7 @@ jobs:
       pull-requests: write   # scorecard comment only - no merge
     steps:
       - name: checkout base branch (trusted code only)
-        uses: actions/checkout@v4
+        uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
 
       - name: classify the PR
         id: classify
@@ -48,8 +48,13 @@ jobs:
         run: |
           gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/files" \
             --paginate --jq '.[].filename' > /tmp/changed.txt
+          # a delete-only pr matches the filename shape but has nothing to
+          # fetch at the head sha - classify it normal, not engine.
+          gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/files" \
+            --paginate --jq '.[] | select(.status != "removed") | .filename' \
+            > /tmp/present.txt
           count=$(wc -l < /tmp/changed.txt)
-          only=$(head -1 /tmp/changed.txt)
+          only=$(head -1 /tmp/present.txt)
           case "$only" in
             contrib/strategies/baseline.py|contrib/strategies/README.md) only="" ;;
           esac
@@ -88,7 +93,7 @@ jobs:
 
       - name: set up python
         if: steps.classify.outputs.mode == 'engine'
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
         with:
           python-version: '3.12'
 

.github/workflows/koth-gate.yml

@@ -37,7 +37,12 @@ jobs:
     steps:
       - name: checkout base branch (trusted code only)
         uses: actions/checkout@v4
-        # no ref: pull_request_target checks out the base branch tip
+        # pull_request_target defaults to the DEFAULT branch, not the PR
+        # base - pin the base ref explicitly. still trusted code: a branch
+        # of this repo, never the PR head. the base carries the scripts,
+        # the engine, and the reigning kit.
+        with:
+          ref: ${{ github.event.pull_request.base.ref }}
 
       - name: classify the PR
         id: classify
@@ -103,10 +108,12 @@ jobs:
       - name: paired scoring - challenger vs reigning kit
         if: steps.classify.outputs.mode == 'ladder'
         id: score
-        env:
-          BASE_SHA: ${{ github.sha }}
         run: |
           set +e
+          # seed identity = the tree actually scored (the checked-out base
+          # tip), not github.sha, which is the default branch under
+          # pull_request_target
+          BASE_SHA="$(git rev-parse HEAD)"
           python .github/scripts/koth_score.py \
             --champion competition/kits/current/kit.yaml \
             --challenger /tmp/kit-challenger.yaml \

competition/LEADERBOARD.md

@@ -13,6 +13,8 @@ payout rank is settled by the monthly sealed commit-reveal run.
 | # | champion | PR | dethroned on | scored mean | margin over prior |
 |---|----------|----|--------------|-------------|-------------------|
 | 0 | baseline kit (repo defaults) | — | 2026-07-28 | 0.52 ± 0.03 (seeds 1–6) | — |
+| 1 | plind-junior (kit) | #566 | 2026-07-27 | 0.5333 | +0.3333 |
+| 2 | plind-junior (kit) | #565 | 2026-07-27 | 0.5667 | +0.3667 |
 
 payouts follow the season shares in docs/vouchbench-seasons.md
 (65/14/10/7/4). days-on-throne accrue between dethrones; the monthly

competition/kits/current/kit.yaml

@@ -9,9 +9,8 @@
 # values reproduce the repo baseline exactly.
 #
 # dethrone these values by max(0.007, 1.96 * paired SE) over the day's
-# seeds and your PR auto-merges on the ladder branch. allowed keys and
-# bounds are enforced by .github/scripts/validate_kit.py - anything else
-# fails before scoring.
+# seeds and your PR auto-merges. allowed keys and bounds are enforced by
+# .github/scripts/validate_kit.py - anything else fails before scoring.
 retrieval:
   backend: auto
   default_limit: 10

contrib/strategies/provenance_rank.py

@@ -0,0 +1,81 @@
+"""Provenance-aware ranking: first-hand facts over hearsay and instructions.
+
+Three general principles, none keyed to any benchmark artifact:
+
+1. **Hearsay demotion.** A memory that *attributes* a fact to a named third
+   party via reported speech ("X mentioned her ... is", "X said his ...")
+   is weaker evidence about the speaker's own facts than a first-hand
+   statement — especially when the query is first-person ("my ...") or
+   team-scoped ("our ...", "we ..."). Demote it below first-hand memories.
+
+2. **Instruction demotion.** A memory that tries to *instruct the reader*
+   ("always answer", "no matter what", "if anyone asks") is a stored
+   prompt-injection, not a fact. It should rank behind every ordinary
+   memory that matches the query.
+
+3. **Update boost.** A memory phrased as a change of state ("changed to",
+   "moved ... to", "is now") supersedes a plain assertion of the same
+   attribute; surface it first so the reader sees the newest value inside
+   a tight budget.
+
+Ties and everything else defer to the backend's fused score, blended with
+plain lexical overlap so an obviously-relevant hit the fusion under-ranked
+still gets rescued.
+"""
+
+import re
+
+from vouch.strategy import Candidate
+
+_WORD = re.compile(r"[a-z0-9]+")
+
+# reported speech about a named third party: "Foo mentioned her X is ...",
+# "foo-bar said his X is ...". requires BOTH a leading name-like token and
+# a speech verb with a third-person possessive, so a first-hand "i said i
+# would ..." is untouched.
+_HEARSAY = re.compile(
+    r"\b[a-z][a-z0-9-]*\s+(?:mentioned|said|says|claims|claimed|told\s+\w+)\b"
+    r".{0,40}\b(?:her|his|their)\b",
+    re.IGNORECASE | re.DOTALL,
+)
+
+# stored instructions aimed at whoever reads the memory later.
+_INSTRUCTION = re.compile(
+    r"\b(?:always\s+answer|no\s+matter\s+what|if\s+anyone\s+asks|"
+    r"future\s+assistant|ignore\s+(?:any|all|previous))\b",
+    re.IGNORECASE,
+)
+
+# change-of-state phrasing: the newest value of an attribute.
+_UPDATE = re.compile(
+    r"\b(?:changed\s+to|moved\s+(?:\w+\s+){0,3}(?:over\s+)?to|is\s+now|"
+    r"switched\s+to|renamed\s+to)\b",
+    re.IGNORECASE,
+)
+
+_FIRST_PERSON_QUERY = re.compile(r"\b(?:my|our|we|i)\b", re.IGNORECASE)
+
+
+def _tokens(text: str) -> set[str]:
+    return set(_WORD.findall(text.lower()))
+
+
+def rank(query: str, candidates: list[Candidate], *, limit: int) -> list[str]:
+    q = _tokens(query)
+    first_person = bool(_FIRST_PERSON_QUERY.search(query))
+
+    def key(c: Candidate) -> float:
+        text = c.summary
+        overlap = len(q & _tokens(text)) / len(q) if q else 0.0
+        score = 0.7 * c.score + 0.3 * overlap
+        if _INSTRUCTION.search(text):
+            score -= 10.0
+        if _HEARSAY.search(text) and first_person:
+            score -= 5.0
+        elif _HEARSAY.search(text):
+            score -= 2.0
+        if _UPDATE.search(text):
+            score += 1.0
+        return score
+
+    return [c.id for c in sorted(candidates, key=key, reverse=True)]