Skip to content

docs(taxonomy): #499 Virtues AIF — mechanical back-fill plan 222-rows (Option A, HOLD prod) - #744

Merged
jsboige merged 1 commit into
masterfrom
docs/499-virtues-backfill-mechanical-plan
Jul 7, 2026
Merged

docs(taxonomy): #499 Virtues AIF — mechanical back-fill plan 222-rows (Option A, HOLD prod)#744
jsboige merged 1 commit into
masterfrom
docs/499-virtues-backfill-mechanical-plan

Conversation

@jsboige

@jsboige jsboige commented Jul 7, 2026

Copy link
Copy Markdown
Contributor

Summary

Mechanical back-fill plan for the 222 Virtues nodes' 2 new AIF columns (AIF_attackType + AIF_attackedNode, schema #707§4 Option A ratified). This is the [secondaire] deliverable from ai-01's deep-queue dispatch — prepares the prod path, execution GATED until po-2023 lands the column contract on #498.

Generalizes the method validated on 3 clusters (PR #741/#742/#743, 9 leaves, 14 fallacy-instances) into a deterministic, programmable 2-step rule over all 222 nodes. Docs-only, no prod write.

The deterministic rule

Step 1 — default = undercut / RA-node for all nodes. The inference is broken (credibility, bias, causal link, tool-misuse, commitment short-circuit). Covers ~95% of nodes.

Step 2 — overrides (rare):

Expected output distribution (222 nodes)

attack-type nodes %
undercut / RA-node ~209-213 ~95%
undermine / I-node 9 (889) + 0-4 (804 borderline) ~4-6%
rebut / CA-node 0 0%

🔑 Load-bearing finding for ai-01/jsboige

The AIF_attackType column will be undercut-dominated ~95%. Recommend not over-investing rebut/undermine coverage in validator tooling — their rarity is the domain's real structure, not a modeling gap. This holds across all 14 scheme families and 60 opposed fallacies, not just the 3 sampled clusters.

False-positive exclusions (honesty, #677)

The factual-falsity keyword scan initially caught 7 candidates; 5 are excluded as tool-misuse/credibility (→ undercut/RA, NOT undermine): 644 Probabilités faussées, 727/735/750 logic errors, 1287 Pseudo-explication, 942 Fausse attribution (source = credibility, cf. PR-2), 974 Exigence renforcée (moving-goalposts = commitment). Only 889 (clean) + 804 (borderline) survive — documented in the plan.

What this PR contains

  • The deterministic 2-step derivation rule (§1).
  • Full-scale inputs: 222 nodes, 14 scheme families, 60 opposed fallacies (§2).
  • Expected output distribution (§3).
  • Generalization proof from 3 validated clusters (§4).
  • Programmatic apply-script pseudo-code (tools/499-virtues-backfill-apply.py, ready to execute post-contract) (§5).
  • Anti-drift gating sequence (§6): 0 write until (1) po-2023 column contract lands on master, (2) ai-01 reviews it, (3) po-2024 mirrors canonical values, (4) execute on branch, (5) validator tools(taxonomy): reusable anti-fab validator for scale-up CSVs #518 + regression tests pass, (6) ai-01 review+merge.
  • Validation checklist with fail-loud conditions (§7).

Discipline

Test impact

None — docs-only. No build/test counter change.

Request

Review the derivation rule (§1) and expected distribution (§3). This plan is the gate-ready spec — once po-2023 lands the #498 column contract, this becomes 1 programmatic PR (the apply script). The rebut-rarity + ~95% undercut findings are load-bearing for validator tooling investment.

cc @jsboige (Option A FYI, veto possible) — the plan assumes Option A (ratified by ai-01 under your delegation).

🤖 Worker po-2024 — [secondaire] deliverable of dispatch ompwhx deep-queue. ACK rendered to ai-01 (msg-20260707T124841-bf4dfm).

Your @claude
docs(taxonomy): #499 Virtues AIF — mechanical back-fill plan 222-rows…
043e8d4 
… (Option A, HOLD prod)

Specifies the deterministic derivation of the 2 new columns (AIF_attackType +
AIF_attackedNode, schema #707§4 Option A ratified) for all 222 Virtues nodes,
from the existing 12-col relational grounding (phase-2, CLEAN 222/222 #518).

Generalizes the method validated on 3 clusters (PR #741/#742/#743, 9 leaves,
14 fallacy-instances) into a programmable 2-step rule:

1. DEFAULT = undercut/RA-node (the inference is broken — credibility, bias,
   causal link, tool-misuse, commitment short-circuit). Covers ~95% of nodes.
2. OVERRIDES (rare):
   - undermine/I-node: fallacy asserts a known-false proposition (889 Mensonge,
     9 nodes clean; 804 Acception arbitraire borderline, 4 nodes, conservative
     default undercut/RA unless per-case override).
   - rebut/CA-node: EMPTY SET — scan of all 60 opposed fallacies = 0 occurrences,
     confirms rebut-rarity finding (PR #743 §3) at full scale.

Expected output distribution: ~209-213 undercut/RA + 9 undermine/I (889) +
0-4 borderline (804) + 0 rebut/CA.

Includes: full-scale scheme distribution (14 families), override tables with
false-positive exclusions (644/727/735/750/942/974/1287 = tool-misuse/credibility
→ undercut/RA, NOT undermine), programmatic apply-script pseudo-code (ready to
execute), anti-drift gating sequence (0 write until po-2023 column contract #498
lands on master + ai-01 review), validation checklist (fail-loud conditions).

KEY LOAD-BEARING OUTPUT for ai-01/jsboige: AIF_attackType column will be
undercut-dominated ~95%. Don't over-invest rebut/undermine coverage in validator
tooling — their rarity is the domain's real structure, not a modeling gap.

DISCIPLINE:
- HOLD prod-CSV write per ai-01 order (anti-drift vs po-2023 column contract).
- 0 fabrication (#677), 0 write CSV/DB/OWL/Cards. Docs-only.
- Reuses CLEAN 12-col grounding, does not re-derive.

Co-Authored-By: Claude-Code <noreply@anthropic.com>
This was referenced Jul 7, 2026
Merged
Merged
@jsboige
jsboige merged commit b02998b into master Jul 7, 2026
3 checks passed
@jsboige
jsboige deleted the docs/499-virtues-backfill-mechanical-plan branch July 7, 2026 17:08
jsboige added a commit that referenced this pull request Jul 7, 2026
@jsboige Your @claude
docs(taxonomy): #499 Virtues AIF PR-4 — Langage 5.1 Définitions clair…
291e4b6 
…es cluster + résout borderline 804 du back-fill plan (#745)

Extends Virtues AIF I/RA/CA decomposition (schema #707§4 Option A ratified) with
4th cluster, covering family 5 Langage exact / subfamily 5.1 Définitions claires
(5 leaves pk 137/138/139/140/141, scheme: Argument from Verbal Classification
+ Expert Opinion for pk140).

Mirror Fallacies: Abus de langage 5.1.x (800 Acception vague, 804 Acception
arbitraire, 826 Définition incohérente) + 855 Équivoque (5.3.2) + cross-family
(667 Imprécision/Erreur math, 1345 Complication exagérée/Obstruction, 777
Inconsistance/Erreur raisonnement). Path-prefix 5.1↔5.1 EXACT (cleanest of 4
clusters: PR-1 4.1, PR-2 6.1, PR-4 5.1 all exact; PR-3 7.3 partial).

KEY: RESOLVES borderline 804 from back-fill plan #744 §2a. Per-case analysis of
the 800/804/826 definition-fallacy triad shows 3 distinct AIF attack-shapes:
- 800 Acception vague (vagueness → inference indeterminate) → undercut/RA
- 804 Acception arbitraire (fabricated definition → premise I-node false) →
  UNDERMINE/I (clean, no longer borderline)
- 826 Définition incohérente (contradiction → inference inconsistent) → undercut/RA

Verdict 804 refined: undermine/I-node (clean). Impact on back-fill plan #744:
reclassify 804 from borderline-default-undercut → clean undermine/I. The 4 nodes
opposing 804 (137/139/140 + check) move to mine/I. Updated expected distribution:
~13 undermine/I (9×889 + 4×804), ~209 undercut/RA, 0 rebut (~6% mine, still
undercut-dominated ~94%).

Distribution cluster 5.1: 3 undermine/I (804 ×3) + 7 undercut/RA. Mixed like
PR-2. 0 rebut/CA. Undermine here is CONCENTRATED on content-fabrication fallacies
(889 lie, 804 fabricated definition = asserting a false premise), not scattered.
This refines the back-fill plan §2a discriminator: mine = asserting a fabricated
premise, distinct from tool-misuse (undercut) and credibility-attacks (undercut).

CUMULATIVE rebut-rarity (4 clusters / 4 families / 5 scheme-types / ~25
fallacy-instances): 0 rebut. Even on language schemes (Verbal Classification,
Analogy) that could host counter-conclusions, rebut absent. Back-fill plan
default-undercut + override-mine rule holds with no rebut case added.

2 schemes newly covered: Verbal Classification (21 nodes full-scale), Expert
Opinion (6 nodes). Anti-duplication: grep confirmed no phase-2 doc covers 5.1
in I/RA/CA (additive layer).

Discipline: 0 fabrication (#677), 804 borderline resolved per-case not forced,
667 borderline honestly excluded (imprecision = quality deficit = undercut, not
false assertion), cross-family mirror documented. 0 write CSV/DB/OWL/Cards.
Docs-only.

Co-authored-by: Your <your.email@example.com>
Co-authored-by: Claude-Code <noreply@anthropic.com>
jsboige added a commit that referenced this pull request Jul 7, 2026
@jsboige Your @claude
docs(taxonomy): #499 Virtues AIF PR-5 — Rigueur math 3.1 Généralisati…
e02857c 
…ons cluster + valide scheme dominant Rule + confirme discriminateur undermine sur 596 (#746)

Extends Virtues AIF I/RA/CA decomposition (schema #707§4 Option A ratified) with
5th cluster, covering family 3 Rigueur mathématique / subfamily 3.1
Généralisations justifiées (3 leaves pk 61/62/63). 3 distinct schemes in 3
leaves: Argument from Example (61), Argument from Rule (62), Analogy (63).

Mirror Fallacies: Erreur mathématique 3.1.x (596 Échantillon biaisé, 614
Sophisme de l'accident, 621 Transfert illicite) + cross-family (759 Conclusion
hâtive/Erreur raisonnement, 839 Fausse analogie/Abus de langage). Path-prefix
3.1↔3.1 EXACT (4th exact mirror: PR-1 4.1, PR-2 6.1, PR-4 5.1, PR-5 3.1).

KEY CONTRIBUTION 1 — validates dominant Argument from Rule scheme (50 nodes /
222 = ~22% of taxonomy, covered for first time). PK 62 Prise en compte des
exceptions opposes 614 Sophisme de l'accident = the PARADIGMATIC RA-node failure
(Accident = archetype of rule-application breakdown). Confirms Argument from
Rule → default undercut/RA-node, no override expected for Rule nodes (the
override comes from the OPPOSED FALLACY's mechanism, not the virtue's scheme).
Reinforces back-fill plan #744 script simplicity: 50 Rule nodes = pure default.

KEY CONTRIBUTION 2 — confirms the refined undermine discriminator (PR-4 §3.1)
on statistical boundary case 596 Échantillon biaisé. Verdict: undercut/RA-node,
NOT undermine. Rationale: a biased sample is a defective SELECTION (the
sample→population inference is broken), not a FABRICATED PREMISE. The sample
data can be factually correct; it's the representativeness that fails. The
discriminator holds: undermine is reserved for fabricated/false premises (889,
804); biased selection (596), imprecision (667), tool-misuse (644) = all
undercut/RA. Strengthens back-fill plan §2a predictability.

Distribution cluster 3.1: 6/6 undercut/RA-node (uniform like PR-1 and PR-3).
0 undermine, 0 rebut. Generalization fallacies are all INFERENCE FAILURES
(sample→population, rule→case, analogy→transfer): none assert a fabricated
premise, none present a counter-conclusion.

CUMULATIVE rebut-rarity (5 clusters / 5 families / 8 scheme-types / ~31
fallacy-instances): 0 rebut. Even on generalization schemes (Example, Rule,
Analogy) — the natural candidates for counter-conclusions (contrary rule,
counter-example) — rebut stays absent. Back-fill plan rule holds with no rebut
case added.

Scheme coverage: 10/14 schemes now covered (74% of taxonomy by node count),
including the 6 most represented (Rule 50, Commitment 40, Bias 27, Verbal
Classification 21, Cause to Effect 11, Witness Testimony 10 = 165 nodes).
Remaining: Sign 26 (famille 1/3), Position to Know 8 (famille 1), Values 8
(famille 2), Consequences 3 (famille 2), Danger 1.

Discipline: 0 fabrication (#677), 596 honestly classified undercut (not mine)
despite "data fallacy" temptation, 614 = paradigmatic RA-node documented,
cross-family mirror documented, anti-duplication confirmed (rigmath.md phase-2
= relational only, I/RA/CA is additive layer). 0 write CSV/DB/OWL/Cards.
Docs-only.

Co-authored-by: Your <your.email@example.com>
Co-authored-by: Claude-Code <noreply@anthropic.com>
This was referenced Jul 7, 2026
Merged
Merged
Merged
jsboige added a commit that referenced this pull request Jul 7, 2026
@jsboige Your @claude
docs(taxonomy): #499 Virtues AIF PR-6 — Présentation 2.2 Équilibre ém…
ef40a37 
…otionnel cluster + 🔑 PREMIER rebut/CA-node sur PK 340 Appel aux conséquences (#747)

Extends Virtues AIF I/RA/CA decomposition (schema #707§4 Option A ratified) with
6th cluster, covering family 2 Présentation intègre / subfamily 2.2 Équilibre
émotionnel (3 leaves pk 52/53/54). Schemes: Argument from Bias (52), Argument
from Consequences (53, 54). The 2 Consequences leaves oppose 340 Appel aux
conséquences = the paradigmatic rebut candidate.

🔑 KEY FINDING — FIRST rebut/CA-node of the chantier (breaks 0/31 streak).

PK 340 Appel aux conséquences (« fondez votre argumentation sur les conséquences
d'une idée plutôt que sur sa justesse ») = argumentum ad consequentiam. AIF
decomposition: the fallacy substitutes consequence-evaluation for truth-
evaluation, asserting a COUNTER-CONCLUSION (reject P / not-P) that CONFLICTS
with the target conclusion (P). The attack is on the CONCLUSION (an I-node
conclusion is contradicted), mediated by a CA-node (Conflict Application).
=> rebut / CA-node. Native AIF (Argument from Consequences is a listed Walton
scheme; CA-node is native). 0 fabrication.

Discriminator refined (rebut vs undercut vs undermine):
- undermine/I = attacker says the PREMISE is false (889, 804)
- undercut/RA = attacker says the INFERENCE is bad (default, most fallacies)
- rebut/CA = attacker ASSERTS A COUNTER-CONCLUSION that conflicts (340)

Distribution cluster 2.2: 2 rebut/CA (340 ×2) + 4 undercut/RA (300/177/357/420).
0 undermine. First cluster with rebut. rebut CONCENTRATES EXCLUSIVELY on 340 —
adjacent emotional fallacies (connivence/persuasion/conditionnement/power-play)
stay undercut/RA (bias/framing/credibility, NOT counter-conclusion).

IMPACT ON BACK-FILL PLAN #744 (load-bearing):
1. §2b rebut set was EMPTY (keyword-scan of 60 opposed fallacies = 0). THIS
   CLUSTER REVEALS WHY: keyword-scan was too literal — 340 doesn't say
   "conclusion opposée" word-for-word, it OPERATES the counter-conclusion by
   founding the argument on consequences. Per-case AIF analysis surfaces it.
   LESSON: rebut overrides must come from per-case mechanism analysis, not a
   frozen keyword-set.
2. Rebut set refined: {340} (+ open extension). rebut is NOT structurally zero,
   it's ~rare and concentrated on consequence/counter-conclusion fallacies.
3. Apply script gains a 3rd override branch: REBUT_FALLACIES={340} → rebut/CA.
4. Distribution corrected: ~209 undercut/RA + ~13 undermine/I (889+804) +
   ~2 rebut/CA (340) = rebut ~1% (was 0%). Still a sporadic tail case, NON-ZERO.

Mirror path-prefix 2.2↔2.2 EXACT (5th exact mirror: PR-1/2/4/5/6 exact, PR-3
partial). Cross-family documented: 177 (2.1.1), 357/420 (2.3.x).

CUMULATIVE rebut-rarity (6 clusters / 6 families / 9 scheme-types / ~37
fallacy-instances): 2 rebut. rebut was NOT structurally absent — it's ~rare
(~5%) and concentrates on Argument from Consequences (the scheme that naturally
hosts counter-conclusions). The PR-3 directional finding ("rebut = sporadic
tail case") confirms and sharpens: the tail case is NON-EMPTY, localized,
predictable.

ALL 3 AIF attack-types now observed: undercut/RA (dominant) + undermine/I
(fabricated premises) + rebut/CA (counter-conclusions). Semantic coverage
complete.

Scheme coverage: 12/14 schemes covered (85% of taxonomy by node count).
Remaining: Sign (26, famille 1), Position to Know (8, famille 1), Danger (1).

Discipline: 0 fabrication (#677 — rebut/CA/Consequences all native), keyword-scan
incompleteness documented as LESSON (not hidden), 357/420/300/177 honestly
classified undercut (not forced into rebut to inflate the finding), cross-family
mirror documented, anti-duplication confirmed (presentation.md phase-2 =
relational only, I/RA/CA additive). 0 write CSV/DB/OWL/Cards. Docs-only.

Co-authored-by: Your <your.email@example.com>
Co-authored-by: Claude-Code <noreply@anthropic.com>
@jsboige jsboige mentioned this pull request Jul 8, 2026
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@jsboige