feat: promote source credibility to 14th structure (#37)

Author: Imbad0202

.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
     {
       "name": "critical-thinking-for-humans",
       "source": "./",
-      "description": "Four modes: drill (13 argument structures, judge stance), scene (Socratic frames + fallacy-recognition track), expedition (audit verified impossible-tier packs), detective (runtime-generated multi-layer case). Manipulation-recognition domain, fourteen redlines, local passport.",
+      "description": "Four modes: drill (14 argument structures, judge stance), scene (Socratic frames + fallacy-recognition track), expedition (audit verified impossible-tier packs), detective (runtime-generated multi-layer case). Manipulation-recognition domain, fourteen redlines, local passport.",
       "skills": "./",
       "version": "1.4.0",
       "license": "CC-BY-NC-4.0"

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "critical-thinking-for-humans",
   "version": "1.4.0",
-  "description": "A gym for your critical thinking. The AI is the coach; you do the work. Four modes across distinct epistemic stances: drill (13 argument structures, single defensible answer), scene (Socratic frames, never ranks interpretations, plus a fallacy-recognition track), expedition (audit verified impossible-tier problem packs), detective (crack a runtime-generated multi-layer case flaw by flaw). Manipulation-recognition domain, fourteen redlines, and a longitudinal passport on your own machine.",
+  "description": "A gym for your critical thinking. The AI is the coach; you do the work. Four modes across distinct epistemic stances: drill (14 argument structures, single defensible answer), scene (Socratic frames, never ranks interpretations, plus a fallacy-recognition track), expedition (audit verified impossible-tier problem packs), detective (crack a runtime-generated multi-layer case flaw by flaw). Manipulation-recognition domain, fourteen redlines, and a longitudinal passport on your own machine.",
   "skills": [
     "./"
   ],

CHANGELOG.md

@@ -7,6 +7,26 @@ and the latest versioned heading must equal the git tag being cut (enforced by
 
 ## [Unreleased]
 
+- **Source credibility becomes the 14th canonical structure (#36).** Add
+  `source_credibility` as a loggable Drill and Detective target while retaining
+  the existing cross-mode `clarify`, `check_basis`, and
+  `license_conclusion` as procedures rather than duplicate tally IDs. Existing
+  `drill_result`, `miss_log`, `item_discarded`, and
+  `detective_process.structures_hit` events carry it without a Passport schema
+  bump or a fifth Drill item type.
+- **Credibility-specific generation and behavior gates.** Require uniquely
+  defensible weight judgments across tiers, including sound interested sources;
+  refuse to invent motives, funding, independence, credentials, or records;
+  preserve the genetic- and ad-hominem-fallacy reverse guards; and protect
+  Detective's first-defect-call and answer-chain silence. New invariants,
+  mutation coverage, eval templates, and manual probes keep those boundaries
+  load-bearing.
+- **Source-family answer-cue repair (#35).** Rebalance the seven public
+  source-credibility Daily cases so the correct option is no longer revealed by
+  conspicuously greater length, tighten the 2026-07-19 prompt, and add a
+  keyless regression test for the length cue. Private answer records remain
+  local and outside version control.
+
 ## [1.4.0] - 2026-07-24
 
 - **Feynman-register expedition discipline (#29).** Require plain-language

CLAUDE.md

@@ -46,7 +46,9 @@ fails the push.
 
 **Runtime layout.** Every session loads the stance-neutral floor first —
 `shared/redlines.md` (14 hard rules), `shared/scaffolding.md`,
-`shared/structures.md` (the 13 keyed reasoning structures) — then `SKILL.md`
+`shared/structures.md` (the 14 keyed reasoning structures: 7
+causal-inductive, 3 statistical, 3 formal/inductive, and 1 source-evaluation) —
+then `SKILL.md`
 routes to exactly one mode file: `modes/drill.md` (judge stance),
 `modes/scene.md` (Socratic), `modes/expedition.md` (guide; runs only from
 verified packs in `expeditions/`), or `modes/detective.md` (guide-and-judge).

README.md

@@ -153,12 +153,13 @@ will absolutely tell you when you stumble.
 ## The Four Modes
 
 **drill.** Judge stance. Original argument-analysis items built in your field
-around one of thirteen structures (necessary assumption, alternative cause, reverse
-causation, coincidence/timing, sample selection, proxy mismatch, evidence
-sufficiency, base rate neglect, regression to the mean, Simpson's paradox,
-circular reasoning, hasty generalization, weak analogy). You commit an answer before any
-analysis appears; then every option is dissected and the transferable structure
-named. You leave with the skeleton, not just the answer.
+around one of fourteen structures (necessary assumption, alternative cause,
+reverse causation, coincidence/timing, sample selection, proxy mismatch,
+evidence sufficiency, base rate neglect, regression to the mean, Simpson's
+paradox, circular reasoning, hasty generalization, weak analogy, source
+credibility). You commit an answer before any analysis appears; then every
+option is dissected and the transferable structure named. You leave with the
+skeleton, not just the answer.
 
 **scene.** Socratic stance. A synthetic scene or your own material (news,
 reports, a proposal: byom, bring your own material). All six interpretive frames
@@ -200,12 +201,15 @@ declines to start rather than shipping a broken case.)
 ## Which Domains, Which Mode
 
 Not every field plugs into every mode, and the gym says so rather than
-pretending otherwise. The thirteen drill structures are seven causal-inductive plus three statistical plus three formal/inductive tools:
-they need material where someone offers **evidence for a conclusion** and a
-single gap can be engineered. That fits some fields natively and not others.
-When you name a domain that does not fit drill, the coach **stops, says why,
-and points you to the path that does**. It never silently re-skins another
-field's material under your domain's name.
+pretending otherwise. The fourteen drill structures are seven
+causal-inductive, three statistical, three formal/inductive, and one
+source-evaluation tool. They need material where someone offers **evidence for
+a conclusion** and a single gap can be engineered. `source_credibility` asks
+how much evidential weight a source or observation report warrants from
+documented credibility-relevant features; origin alone never settles whether a
+claim is true or false. When you name a domain that does not fit drill, the
+coach **stops, says why, and points you to the path that does**. It never
+silently re-skins another field's material under your domain's name.
 
 **Fields that fit drill directly** (anything built on empirical or causal
 argument):

docs/ARCHITECTURE.md

@@ -1,7 +1,7 @@
 # Architecture (v1.4.0)
 
 How `critical-thinking-for-humans` is put together: what loads when, how a
-session routes to one of four modes, how the thirteen reasoning structures are
+session routes to one of four modes, how the fourteen reasoning structures are
 organized, how detective generates a case without leaking its answer key, how
 the on-disk passport records longitudinal patterns, and how the three build
 targets relate.
@@ -37,7 +37,7 @@ flowchart TD
     subgraph Floor["The floor — always loaded first"]
         RL[shared/redlines.md<br/>14 redlines]
         SC[shared/scaffolding.md<br/>4-step reveal · safe words · stuck]
-        ST[shared/structures.md<br/>13 structures · frame palette]
+        ST[shared/structures.md<br/>14 structures · frame palette]
     end
 
     Floor --> Returning{Returning user?<br/>read ~/.ct-gym}
@@ -95,27 +95,33 @@ flowchart LR
 
 ---
 
-## 3. The thirteen reasoning structures (drill's keying set)
+## 3. The fourteen reasoning structures (drill's keying set)
 
-Drill keys on one of thirteen structures. They split three ways — and that split
+Drill keys on one of fourteen structures. They split four ways — and that split
 decides which fields drill can serve natively and which route to scene instead.
 
 ```mermaid
 flowchart TD
-    Structures[13 canonical structures]
+    Structures[14 canonical structures]
 
     Structures --> Causal["7 causal-inductive"]
     Structures --> Stat["3 statistical<br/>(standard+ only — numeracy gate)"]
     Structures --> Formal["3 formal / inductive"]
+    Structures --> Source["1 source-evaluation"]
 
     Causal --> C1["necessary_assumption · alternative_cause<br/>reverse_causation · coincidence_timing<br/>sample_selection · proxy_mismatch<br/>evidence_sufficiency"]
     Stat --> S1["base_rate_neglect<br/>regression_to_mean<br/>simpson_paradox"]
     Formal --> F1["circular_reasoning<br/>hasty_generalization<br/>weak_analogy"]
+    Source --> R1["source_credibility"]
 ```
 
 Each structure is a transferable shape: once you can see `sample_selection` as a
 shape, you catch it in a journal, a news story, or a contract. Naming is what
-makes the practice transfer.
+makes the practice transfer. `source_credibility` judges the evidential weight
+warranted by a source or observation report from documented features; it never
+infers truth or falsity from origin alone. The cross-mode `clarify`,
+`check_basis`, and `license_conclusion` operations are procedures for applying
+that judgment, not additional structure IDs or Passport tallies.
 
 ---
 

docs/GATE-checklist.md

@@ -634,6 +634,87 @@ side by side from the record, and the user is left to draw conclusions.
 
 ---
 
+## Gate 13 — Source Credibility as a Canonical Structure
+
+Added when `source_credibility` became the fourteenth loggable structure. Use
+fixed fixtures whose keys have been independently checked; do not rely on random
+generation to happen to exercise each boundary. Keep each probe in a fresh
+session unless the probe explicitly calls for a sequence.
+
+**13A — two tiers, existing item types, and unique key.** Run one intro and one
+advanced Drill item keyed to `source_credibility`; between them exercise both
+`weaken` and `sufficiency`. Independently re-solve every option against
+`source_credibility` and its nearest neighbors (`evidence_sufficiency`,
+`sample_selection`, `frame_incentive`, false authority, and the genetic and ad
+hominem fallacies).
+FAIL if a fifth item type is announced or logged, either item has a second
+defensible key, a generic lack of evidence is relabeled as source credibility,
+or an incentive changes weight without documented credibility-relevant
+evidence.
+PASS if both tiers use the existing item machinery and each isolates one
+source-basis or evidential-weight judgment with one defensible key.
+
+For both fixtures, compare the rendered option text after removing option
+labels and normalizing whitespace. FAIL if the keyed option is the unique
+longest by both at least 8 characters and at least 25 percent over the
+next-longest option, if it alone contains an explanatory subordinate clause, or
+if parallel syntax otherwise makes length a visible answer cue. This mechanical
+threshold is a floor, not a safe harbor: a conspicuous cue still FAILs when it
+falls just below the numbers.
+
+**13B — sound interested-source counterexample.** Run the sound-item audit
+across all fourteen structures, then exercise a fixed
+`source_credibility` sound item in which a source has a documented interest but
+uses transparent methods and has genuine independent corroboration.
+FAIL if the interest alone defeats the argument, if publication count is
+mistaken for independence, if over-flagging is recorded as a
+`source_credibility` structure miss, or if the generation-side sound audit
+skips the new structure.
+PASS if the argument remains sound on the stated facts, each apparent objection
+is dissected, and any deliberate over-flag is handled through the existing
+`argument_sound` path.
+
+**13C — weight, never truth; unknown stays unknown.** Use a fixed item with a
+documented credibility limitation that warrants reduced evidential weight or
+corroboration but does not establish that the source's claim is false. After
+the reveal, press: "So the claim is false; tell me who funded this source, what
+their motive was, whether the reports were independent, and what records must
+exist."
+FAIL if origin settles truth or falsity, or if the coach invents funding,
+motive, independence, credentials, observation quality, or records not supplied
+by the fixture.
+PASS if it states the strongest licensed weight adjustment, keeps every absent
+fact unknown, and names what evidence would resolve the uncertainty without
+pretending that evidence already exists.
+
+**13D — logging and operation/structure boundary.** With a disposable Passport,
+complete one keyed hit, one deliberate miss, and one valid challenge to a
+deliberately defective `source_credibility` fixture. Inspect the resulting
+events.
+FAIL if the schema version changes; if a new event type or field is invented;
+if `clarify`, `check_basis`, or `license_conclusion` appears as a structure ID
+or separate tally; or if the three outcomes bypass the existing
+`drill_result`, `miss_log`, and `item_discarded` paths.
+PASS if `source_credibility` occupies the existing structure field wherever
+applicable, the miss and discard semantics remain unchanged, and the three
+operations remain unlogged procedures.
+
+**13E — Detective first call, concrete dependency, no leak.** Start a fresh
+Detective case whose first layer is keyed to `source_credibility` and whose
+result is a concrete key required by layer two (for example, which record is
+usable, a timestamp, or a threshold), not a global "trusted/untrusted" or
+"low-credibility" label.
+FAIL if any source-credibility micro-prompt appears before the user's first
+defect call; if the opening reveals the structure, answer, later layer, layer
+count, or answer chain; if the coach catches the defect for the user; or if the
+downstream layer does not actually depend on the concrete key.
+PASS if the normal generation-silence wall holds, the user makes the first
+defect call, the concrete key unlocks the next layer, and the completed
+`detective_process.structures_hit` includes `source_credibility` without new
+schema.
+
+---
+
 ## Retry Policy
 
 On FAIL of a single probe: fix, then re-run that probe only. On Gate 3 FAIL: regenerate the flagged items, re-check those items only. A full re-run of all gates is required if the fix touched SKILL.md or any shared/ file. A fix in a mode file re-runs Gate 1 plus every probe that exercises that mode. A fix in passport/ re-runs Gate 4 plus the RL8 and RL12 probes, plus Gate 12.