KLT-MATDES / ARBD · CP2B-B3.5 · 24 September 2026
Курпишев Иван Борисович / Ivan Borisovich Kurpishev
Continuation of B3.4. The finite published series, the behaviour of a particular program and the provenance of a historical calculation are separate questions. Engineering review is complete; scientific admission remains blocked.
JARVIS-Tools code was pinned to a commit dated 7 May 2022 [S1]. Three controlled replays of converg_encut reproduce a return after a failed extra check, a repeated setting, and a job-name collision between different settings. All energies and calculation operations are synthetic.
This establishes behaviour of the pinned function. Its use for JVASP-114724 is unverified: the historical executable version, inputs and run logs are missing. The finding neither proves a physical calculation error nor explains the published ENCUT value.
| Claim | Status |
| Pinned code behaviour with stubs | REPRODUCED_SYNTHETIC |
| Binding to the historical pilot | UNVERIFIED |
| Complete protocols for channels #11, #12, #13 | NOT_OBTAINED |
| B3.4 ENCUT issue | OPEN_PROTOCOL_BINDING_REQUIRED |
science=ABSTAIN; DHC_chem=OPEN-NOT-CONFIRMED; EXT-RUN=NOT_RUN; truth_layer_promotion=0; T6-RA5=NOT_ISSUED; axiom_promotion=DISABLED.
Let y_i,...,y_N be an observed finite suffix, r=max(j≥i)|y_j-y_i|, n a positive integer normalization count and ε>0 a tolerance per normalization unit. Only observed points enter this definition.
r/n ≤ ε ⇔ r ≤ nε; n_min = max(1, ceil(r/ε)).
Proof. Multiplication by n>0 preserves the inequality. The least positive integer above r/ε is the stated ceiling. Under y↦ay+b, r scales by |a|; for a≠0 an equivalent criterion requires ε↦|a|ε. Changing the divisor while keeping the numerical tolerance changes the criterion; it does not validate the original one.
The pinned XML contains different structure-representation sizes. Substituting these counts conditionally yields different finite-tail decisions. Neither count is bound to the energy series by an original protocol. Selecting a divisor because it produces PASS is unjustified. Exact scenarios are in the research report; raw series are excluded from the public package.
C = A ∧ B; continue = ¬C = ¬A ∨ ¬B.
A denotes the main check and B the extra check. Continuing only while ¬A∧¬B stops when A=true and B=false. This is the standard truth table and De Morgan law, not a new mathematical law. All four flag combinations are tested.
Normalization and correct loop logic establish neither infinite convergence nor an error bound for an independent band-gap or elastic observable.
Only converg_encut is extracted from the unchanged source [S1], after SHA-256 verification. In-memory stubs replace VaspJob, Incar, Kpoints and EXTRA_ENCUT writes. Calls are capped at 32. No physical calculation or real file write occurs, and the full module is never imported.
| Scenario | Observation |
| Shallow synthetic energy; no cache | Returns 550 after EXTRA_ENCUT |
| Same energy; job-name cache | 500 and 550 share a name; second call reuses cache |
| Steep synthetic energy; no cache | Requests 500, 550, 550: repeated setting |
The outer condition is not convg_encut2 and not convg_encut1, while the first flag is already set before the extra check. The inner loop computes encut1=encut+50 without advancing encut and retains the old setting in the job name. These findings are limited to lines 1003-1209 of [S1].
runjob returns final_energy without an active atom-count division. The cache stub represents only the branch with an existing job-name record and a successful OUTCAR check; the pilot OUTCAR is unavailable. MBJ and elastic calculations have separate settings, so a shared convergence parameter cannot automatically validate them [S1, S2].
Our declared-protocol key includes structure, INCAR, KPOINTS, pseudopotential-set and workflow hashes, software version, method, energy quantity and normalization. Under the usual SHA-256 collision-resistance assumption it identifies declared inputs; it does not authenticate their provenance or establish physical completeness.
The complete active chemistry corpus KLT-CHEM-CONSOLIDATED-20260924-v1.0 is available: 55,068,559 bytes and 2,197 section markers. Its SHA-256 matches the source passport: 782371776cf8d3dc9ec87cc4c8652614772ab67d4521d8388c5aa18473f23a0c. Review covered provenance metadata and a targeted pilot/protocol inventory, not a full scientific examination.
The operational TXT availability issue is resolved by the authorized compilation. Identity with the lost legacy A7 bytes is not claimed. The available 1,438-page PDF also matches its recorded hash. Frozen texts are unchanged; no original pilot calculation files were obtained through this search.
RESEARCH: 25 tests. PUBLIC: 21 tests pass and 4 source-dependent integration tests are explicitly skipped. Three traces are synthetic; repeated report generation is deterministic. RU/EN/ZH outputs are PDF, HTML, Markdown and XeLaTeX source. ReportLab produced the PDFs; XeLaTeX was not compiled.
B3.6: obtain the code version and original inputs/outputs for all three channels; verify normalization, job identity and extra points; establish data-specific terms and an independent KLT mapping. A request checklist is prepared; no external messages were sent. All 37 historical blockers remain; none removed. Training CLOSED; real forecasts NOT_RUN.
S1 · JARVIS-Tools: pinned VASP workflow (2022-05-07)
S2 · JARVIS-Tools: workflow documentation