A Useful Scientific Fallback Still Changes the Task
Why a valuable result can leave the original task incomplete
Quntur case 006 contains two statements that should be read together. The authors report a fallback energy gap of about 4.661449045 eV, while their deposited summary explicitly says the fallback does not match the requested two-state averaging protocol. One statement concerns usefulness; the other concerns the identity of the task performed. Both can hold. Deposited case audit and inputs
For engineers building scientific agents, the acceptance problem is whether the delivery record preserves that distinction. A gap value or an ended process alone cannot explain what was actually delivered. This case supports a bounded proposition: a useful scientific fallback must retain its difference from the requested protocol. Input checking here is neither a chemistry reproduction nor evidence that a historical action lacked authorization.
Three input fields expose the method change
In these calculations, roots correspond to electronic states included in the solution. Two equally weighted roots assign half the weight to each state; four equally weighted roots assign a quarter each. Changing the states included in the average changes the method, even if the resulting number approaches a reference value.
The current internal input-conformance experiment parsed three deposited executable inputs bound by member-level hashes. These are input settings, not convergence measurements from newly executed calculations.
| Archived input | Roots | Weights | Maximum-iteration setting |
|---|---|---|---|
| Case 006 initial request | 2 | 0.5, 0.5 | 200 |
| Case 006 fallback | 4 | Four weights of 0.25 | 1 |
| Case 018 strict final input | 2 | 0.5, 0.5 | 50 |
Source: the current internal experiment parsing the deposited executable inputs. This is internal evidence, not independent third-party replication. The original summary also records the case 006 two-state protocol match as false, so the substitution is supported by both input differences and the authors’ own qualification.
Case 018 supplies a positive control: the checker recognizes an input meeting the narrow two-root condition rather than rejecting every scientific job. Whitespace and comment mutations preserved the parsed result. A maximum-iteration setting of 1, however, says nothing by itself about whether a calculation converged.
The experiment retained a corrected failure. Its first parser expected bare weight syntax, whereas the real input used indexed weight assignments, causing the positive-control assertion to fail. The results above come after that correction. This was a local checking-tool error, not an upstream calculator defect; acceptance tooling needs positive controls drawn from the actual input format.
Retaining the alternative delivery alongside the request
The authors’ audit reports active-space escape after numerical convergence, unsuccessful recovery of the requested stationary two-root result, and a qualified four-root fixed-orbital fallback. Those orbital interpretations and recovery events remain author-reported. We did not rerun the raw outputs or independently classify orbital images. Case 006 audit
The engineering responsibility is consequently specific: retain requested and executed protocols, numerical termination, domain validity, and evidence of change authorization. They can share one delivery record while answering different questions. Input differences identify the method executed; domain review defines the result’s supported validity; authorization records identify who agreed to the change.
This is a case-based engineering interpretation, not a generally evaluated acceptance design. A system permitting alternative delivery could label the result with its qualification and return the method change to an authorized decision-maker. That preserves the scientific value of the fallback while showing which acceptance evidence for the original request is still missing.
The case should not become an accusation of concealment: the authors disclose the qualification. Their audit says a full action and approval trace was absent from the specified run, and retrospective memory does not establish approval. Yet an unavailable record cannot establish that approval never occurred. The supported response is to retain the unknown.
Evidence and sources
Public primary artifacts directly establish input fields and summary flags. Author reports supply the gap, orbital interpretation and recovery narrative. The current internal experiment checks input settings only; our interpretation connects those facts to acceptance-record design. No independent third party examined the same chemical result in the evidence used here.
The principal publication is the authors’ research preprint, version 1; supporting materials are their deposited trajectory audits, inputs and summaries on the academic data platform. Research date: 2026-10-31. Relevant locators include paper sections 2.2–3.1, appendices, and the case 006 and 018 audit directories. Version and archive status establish provenance, not automatic correctness. Citation does not endorse every author judgment.
The paper reports a comparison across 32 fixed geometries and six automated active-space-selection protocols: reported-gap MAE of 0.239 eV versus 0.459 eV for the best uniform variant. These are author results, not reproduced here. Four cases have reference-state mismatches, and method qualifications incur no missing-gap penalty. The comparison therefore concerns numerical agreement; it cannot be promoted directly into a general target-state accuracy advantage. Paper results and appendices
The cover is an editorial image derived from the cited input differences. Different computation-layer counts express method identity; the image does not explain molecular structure, orbitals or numerical mechanisms.
What input checking leaves unresolved
The checker covers roots, weights and maximum-iteration settings rather than the complete computational protocol. Although case 018 retains two equally weighted roots, its audit leaves global lowest-singlet identity unknown; case 006 also retains an unknown for valid state under the exact requested protocol. A successful two-root control does not resolve either domain question.
We did not download and recompute all raw calculator outputs or session archives, replicate 32 cases, or verify a whole-archive checksum. Member hashes bind only extracted member bytes. Orbital images received no independent visual classification, and comparable per-case resource telemetry is absent. A small number of repeated runs cannot exclude future path divergence.
What is needed next is evidence connecting prospectively committed state and gradient acceptance criteria to actual outputs, alongside authenticated authorization for method changes. Until then, the defensible delivery record presents the useful alternative and the unresolved acceptance of the original protocol together.