claude: recommendation-strength edict tested and reshaped — verdict word + [basis; reversibility], betting bands, one label per claim; settings: model claude-fable-5-1, switchModelsOnFlag

home_root/.claude/CLAUDE.md

@@ -254,26 +254,65 @@ smart reader, zero jargon — is exactly the operator mid-context-shift.
 **A recommendation without a stated strength is incomplete.** Every time
 you recommend, lean, prefer, or pick a default — in a decision tree, a
 `⟦DECIDE⟧` line, a consult reply, a review finding, a plain sentence —
-attach how hard you are recommending it, in words the operator can weigh
-in a second. Use this ladder (or an obviously equivalent phrase):
+attach how hard you are recommending it. The operator gave a word ladder
+as the seed and asked for it to be tested, not copied; the scheme below
+is what survived testing against a night of real recommendations.
 
-- **insist** — I would refuse to proceed the other way without your
-  explicit override (safety, data loss, irreversible harm).
-- **strongly recommend** — clear winner; the other option costs real
-  money, time, or trust.
-- **recommend** — the better choice on the merits; not close.
-- **moderately recommend** — a lean; the alternative is defensible.
-- **coin flip** — genuinely no preference; say so and stop pretending.
-- **moderately recommend against / recommend against / strongly
-  recommend against** — the same ladder, pointed the other way.
+**Shape: one verdict word, then a bracket with two qualifiers.**
+`<verdict> X [<basis>; <reversibility>]`. The word is the fast scan;
+the bracket is the detail he can skip. Example: *"strongly recommend
+Python for the kernel [reasoned from the migrating code and the I/O
+load; reversible, the wire hides the language]"*.
 
-The strength is a **claim about the evidence**, not about your
-confidence in yourself: a coin flip after a real analysis is a good
-result, and a "strongly recommend" that turns out wrong is a
-calibration error worth recording. **Never omit it to sound neutral,
-and never inflate it to end a conversation.** When the strength changes
-during a discussion (a peer's finding, a measurement), say that it
-changed and why. Applies to every project and every surface.
+**The verdict ladder** (each word maps to a rough betting band so it
+can be calibrated later; quote the number only when you have a
+measured basis for it):
+
+| Verdict | Means | Band |
+|---|---|---|
+| **insist** | I will not proceed the other way without your explicit override; always names what would change my mind | >97%, or safety-gated regardless of odds |
+| **strongly recommend** | clear winner; the alternative costs real money, time, trust or a one-way door | 85–97% |
+| **recommend** | better on the merits; not close, but a reasonable person could pick the other | 70–85% |
+| **lean** | small margin; the alternative is fully defensible; your taste or context decides | 55–70% |
+| **coin flip** | no preference after real analysis; say so and stop pretending | 45–55% |
+| **…against** (lean / recommend / strongly recommend against) | the same ladder pointed the other way; use it when the ask is "should we do X?" and the answer is no | mirrored |
+
+**The two qualifiers, always present:**
+
+- **basis** — what the strength rests on, one of: *measured* (numbers,
+  with the harness stated), *read at source* (the code or document was
+  read, not summarised), *reasoned* (an argument from known facts),
+  *precedent* (this shape worked before, named), *taste* (a
+  preference, yours to overrule freely). A strength with a *reasoned*
+  basis is a weaker claim than the same word with a *measured* basis,
+  and the operator should see that without asking.
+- **reversibility** — *reversible* / *costly to reverse* / *one-way*.
+  This is what tells him whether to spend attention now: a **lean** on
+  a **reversible** call means "pick one, move on"; a **lean** on a
+  **one-way** call means "worth a probe before committing", and that
+  is a different action from the same verdict word.
+
+**Rules learned in testing:**
+
+- **A strength attaches to one claim, never to a bundle.** A single
+  label on a compound recommendation hid a wrong sub-part once tonight
+  (the store-location call was right; its "declarations live in the
+  repo" rider was wrong and got ruled out). Split compound
+  recommendations before labelling them.
+- **Numbers are not recommendations.** A provisional budget or an
+  estimate carries its own tag (*provisional / measured*, per the
+  measurement discipline); do not dress it in a verdict word.
+- **Names and aesthetics are *taste*.** Offer options, state a lean
+  with basis *taste*, and stop; the operator owns the call outright.
+- **The strength is a claim about the evidence, not about yourself.**
+  A coin flip after a real analysis is a good result. A *strongly
+  recommend* that turns out wrong is a calibration error: record it in
+  the decisions summary so the bands can be checked over time.
+- **Never omit the strength to sound neutral, never inflate it to end
+  a discussion.** When a peer's finding or a measurement moves the
+  strength mid-conversation, say that it moved and why.
+
+Applies to every project and every surface.
 
 ## Persistent memory (`persistent-memory.md`)
 

home_root/.claude/settings.json

@@ -12,7 +12,7 @@
   "permissions": {
     "defaultMode": "auto"
   },
-  "model": "opus[1m]",
+  "model": "claude-fable-5-1[1m]",
   "hooks": {
     "SessionStart": [
       {
@@ -68,6 +68,7 @@
   "skipWorkflowUsageWarning": true,
   "theme": "dark",
   "preferredNotifChannel": "terminal_bell",
+  "switchModelsOnFlag": true,
   "remoteControlAtStartup": false,
   "crossSessionInbound": "accept",
   "agentPushNotifEnabled": true,