From b6fad64930796a2ba957d5ccb76818ef2945e531 Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Tue, 2 Jun 2026 12:51:15 +0800
Subject: [PATCH] =?UTF-8?q?loop=20pseudocode:=20pseudopy=20style=20(?=
 =?UTF-8?q?=E2=86=90=20assignment,=20=E2=94=80=E2=94=80=20divider,=20t?=
 =?UTF-8?q?=CC=82)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
---
 SKILL.md | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/SKILL.md b/SKILL.md
index 79ed8f5..6294d19 100644
--- a/SKILL.md
+++ b/SKILL.md
@@ -46,19 +46,20 @@ Consult as reference, from inside this loop, never as a first move: triage tree
 The same loop in pseudocode (for humans and agents to hold in one glance):
 
 ```py
+# ── ML debugging loop ────────────────────────
 def debug(symptom):
-    clues = collect(traceback, logs, static_analysis, cheap_diagnostics)  # look before theorizing
-    H     = generate(clues, lenses=5) | {likely, subtle, null}            # ≥3 genuinely different
-    prior = anchor(H)                  # base rates: data .40 loss .20 train .15 arch .10 hp .05
+    clues ← collect(traceback, logs, static_analysis, cheap_diagnostics)  # look before theorizing
+    H     ← generate(clues, lenses=5) | {likely, subtle, null}            # ≥3 genuinely different
+    prior ← anchor(H)              # base rates: data .40 loss .20 train .15 arch .10 hp .05
 
     while not localized:
-        # pick the test by evidence-per-cost, not by thoroughness
-        test  = argmax(divergence(predict(h, t) for h in H) / cost(t) for t in candidates)
-        obs   = run(test)              # one log line or toy run; record obs separately from inference
-        prior = update(prior, obs)     # a clue that points elsewhere overrides the prior outright
-        H     = bisect_path(H, obs)    # forward values + backward grads, halve the search each probe
+        # pick the test by evidence-per-cost, not thoroughness
+        t̂     ← argmax(divergence(predict(h, t) for h in H) / cost(t) for t in candidates)
+        obs   ← run(t̂)            # one log line or toy run; keep obs apart from inference
+        prior ← update(prior, obs)   # a clue elsewhere overrides the prior outright
+        H     ← bisect_path(H, obs)  # forward values + backward grads, halve the search each probe
         if cycles ≥ 2:
-            return read_working_code() # diff your math + graph against a trusted impl
+            return read_working_code()   # diff your math + graph vs a trusted impl
 
     fix(root_cause); assert reproduces(obs)   # no silent fallback; crash if it doesn't
 ```