Skip to content

feat(taxonomy): #498 AIF — serialize P1 tranche-1d+1e+1f prod write (16 SUFFIX-ONLY rows, attack 121->137) - #783

Merged
jsboige merged 1 commit into
masterfrom
feat/498-p1def-write
Jul 11, 2026
Merged

feat(taxonomy): #498 AIF — serialize P1 tranche-1d+1e+1f prod write (16 SUFFIX-ONLY rows, attack 121->137)#783
jsboige merged 1 commit into
masterfrom
feat/498-p1def-write

Conversation

@jsboige

@jsboige jsboige commented Jul 11, 2026

Copy link
Copy Markdown
Contributor

#498 AIF — serialize P1 tranche-1d+1e+1f prod write (16 rows, attack 121→137)

Worker po-2024 · Base cbf9b7c8 (baseline 121, post-tranche-1c) · Statut prod write (GO ai-01 msg-ftqy51).

Batched serialization of the 3 reviewed+GO'd propositions (#778 1d, #780 1e, #781 1f). ai-01 concur'd on every MED verdict including the same-token pk33/pk808 divergence.

16 disjoint PKs cell-filled (empty → value), 2 AIF columns only

tranche famille PKs distribution
1d Erreur mathématique 595/614/632/677 4 undercut/RA
1e Insuffisance 2/33/34/43/71 5 undercut/RA
1f Abus de langage (MIXTE) 800/808/814/833/839(undercut)/856/876 6 undermine/I + 1 undercut/RA

Batch : 10 undercut / 6 undermine. attack layer 121 → 137 (fully-modeled 46 → 62).

Byte-exact verified 2×

  1. Self-proof (each apply script) : 0 mismatch, 1409×104, CRLF+BOM preserved.
    • 1d: 121→125 (+60 B) · 1e: 125→130 (+75 B) · 1f: 130→137 (+105 B)
  2. INDEPENDENT (tmp/_verify_p1def_batch.py, backup pre-batch state-121 vs prod state-137) :
    • exactly 16 PKs changed in AIF_attackType/AIF_attackedNode only = 32 cell changes
    • 0 stray change anywhere else (byte-preservation of all 1408 rows × 104 cols)
    • 0 value mismatch — every cell matches the expected MAP
    • pre-state of all 16 PKs was empty (fill, not overwrite)

Full attack-layer distribution now (137)

82 undercut / 50 undermine / 5 rebut

Gate / safety

Handoff ai-01

  • Review + merge → attack 137 EN PROD (35/36 SUFFIX-ONLY serialized ; reste Tricherie 8 = tranche-1g, prochaine proposition gated).

🤖 Worker po-2024 — batched write 1d+1e+1f (16 rows, 121→137).

…16 SUFFIX-ONLY rows, attack 121->137)

Batched prod write of 3 reviewed+GO'd propositions (master cbf9b7c, baseline 121
post-tranche-1c). ai-01 reviewed the 3 propositions (#778/#780/#781) and
concur'd on every verdict including the same-token pk33/pk808 divergence
(msg-ftqy51 [GO WRITE]).

16 disjoint PKs, 2 AIF columns cell-filled (empty -> value):
  - tranche-1d (Erreur mathematique, 4): 595/614/632/677 -> all undercut/RA-node
  - tranche-1e (Insuffisance, 5): 2/33/34/43/71 -> all undercut/RA-node
  - tranche-1f (Abus de langage, 7, MIXTE): 800/808/814/833/856/876 ->
    undermine/I-node + 839 undercut/RA-node

Distribution this batch: 10 undercut / 6 undermine. attack layer 121 -> 137
(fully-modeled 46 -> 62). Full attack distribution now: 82 undercut / 50
undermine / 5 rebut.

Byte-exact verified 2x:
  - self-proof in each apply script (0 mismatch, 1409x104, CRLF+BOM)
  - INDEPENDENT backup-vs-prod (tmp/Fallacies-backup-pre-p1def-batch.csv vs prod):
    exactly 16 PKs changed in ATI/ANI only = 32 cell changes, 0 stray, 0 value
    mismatch, pre-state of all 16 was empty (fill not overwrite).

Only prod CSV in diff (16 ins/16 del). 0 token fabricated (TYPE rows already
skos-vetted). Secret-scan: no credential in diff.

Co-authored-by: Claude-Code <noreply@anthropic.com>
@jsboige
jsboige merged commit 4161b67 into master Jul 11, 2026
3 checks passed
@jsboige
jsboige deleted the feat/498-p1def-write branch July 11, 2026 19:02
jsboige added a commit that referenced this pull request Jul 11, 2026
…FIX-ONLY, famille Tricherie, MIXTE gated, LAST FAMILY) (#784)

P1 reconciliation back-fill, tranche-1g = 8 SUFFIX-ONLY rows
(888/973/1020/1023/1066/1087/1148/1175), the whole "Tricherie" family
(Arranger les faits + Changement de cap + Pensee biaisee). LAST SUFFIX-ONLY
family -> closes the 36-row back-fill
(5 1b + 7 1c + 4 1d + 5 1e + 7 1f + 8 1g = 36).

HETEROGENEOUS (MIXTE): 3 undermine / 5 undercut. Three defect natures:
  - relational manipulation (deceptive arrangement, goalpost-moving, biased
    reasoning) -> premise-level (undermine/I): 888, 973, 1023 (precedents
    PositionToKnow pk70, ConflictingGoals pk356, Bias pk70).
  - sunk-cost / effort-justification -> inference rule undercut via Waste
    scheme (undercut/RA): 1020, 1148 (precedent Waste_Inference pk432).
  - cognitive induction biases (mere exposure, anecdotal, attribution) ->
    weak induction (undercut/RA): 1066, 1087, 1175 (1066/1175 Example_Inference
    novel; 1087 desc-twin of tranche-1e pk34).

6/8 tokens have a typed precedent; 2/8 novel. attackType = NEW judgment;
NODE deterministic (#707§4a); 0 token fabricated.

3 MED flags (888, 1066, 1175):
  - pk888 (Presentation trompeuse): SAME-TOKEN DIVERGENCE with tranche-1e pk2
    (LackOfPTKReliability_Scheme). pk2->undercut (induction); pk888->undermine
    (deceptive arrangement = biased premise, reinforced by PositionToKnow pk70).
    2nd such divergence (after 1f pk808/pk33) — reinforces desc-driven principle.
  - pk1066 (Effet de simple exposition): Example_Inference novel; undercut
    (familiarity->acceptable defective rule); alt undermine.
  - pk1175 (Biais d'attribution): Example_Inference novel; undercut (attribution
    over-systematic); alt undermine.

5 HIGH: 973 (ConflictingGoals precedent), 1020 (Waste precedent sunk cost),
1023 (Bias+Presumptive precedent), 1087 (desc twin pk34), 1148 (Waste precedent).

0 prod write (gated dry-run). Baseline dynamic (write-driven, disjoint PKs):
121->129 (master cbf9b7c) / 137->145 (if 1d+1e+1f #783 merged). Byte-exact:
0 mismatch, 1409x104, CRLF+BOM, +120 B.

- docs/taxonomy/498-reconciliation-p1g.md (proposition, MIXTE, last family)
- docs/taxonomy/498-reconciliation-p1g-annotations.csv (8 rows, 12 cols)
- tools/498-p1g-apply.py (gated apply, dry-run default, mirrors p1f)

Co-authored-by: Your <your.email@example.com>
Co-authored-by: Claude-Code <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant