AI in your research workflow

Finish a small piece of your research with something you can use: a reading queue, a clearer supported paragraph, or a checked table or plotting contribution. Keep enough of the route back that another author can see what the assistant proposed, what you decided, and what still needs work.

For the optional example, open a text editor beside the conversation view of an institutionally approved AI assistant. You will make a short screening record from six supplied synthetic training records, then check a real journal's policy and decide how this particular use would be reported. You do not need a literature-search application, an API, or a new installation. A colleague can make proposals from the same prompt, or you can work through the reference result manually and record that no assistant was used.

This is Level 1 User. The research question and scientific judgement stay yours. If your own permitted reading, writing, or code task is more useful, use it instead; the screening set is a practice route, not a compulsory research project.

Learning objectives

• Complete one limited research step and distinguish an assistant's proposal from your accepted contribution.
• Preserve source identity, uncertainty and human inclusion decisions when screening a fixed set of records.
• Verify references and factual claims in original sources, and code outputs against known inputs.
• Classify the actual use under applicable journal and institutional rules rather than assuming that all editing or search is equivalent.
• Keep a compact use log and make a specific, policy-grounded contribution and disclosure decision.

Core concepts

A useful result needs a route back

Imagine that a submission form asks what generative AI did. Your notes say only “used AI for editing,” but the assistant also ranked records and drafted plotting code. You cannot reconstruct which records were skipped or which generated calculation survived. Better disclosure wording cannot repair a lost research trail.

Keep that trail while working:

Your question + permitted inputs + applicable rules
                         |
                         v
               Bounded AI proposal
                         |
                         v
       Human source / meaning / calculation check
                  /                    \
        accepted contribution      rejected or unresolved
                  \                    /
                   v                  v
             Useful result + decision record
                         |
                         v
        Task-specific policy and disclosure decision

Permission comes before sending material. The detailed disclosure decision comes from what actually happened, not from the intended prompt alone. If the tool goes beyond the agreed task, stop and reassess that use.

Screening is a research aid, not a literature finding

For reading, an assistant can label and organise a fixed, permitted record set. That can save you from repeatedly scanning the same abstracts. It cannot establish that the search was complete, that an article is reliable, or that a proposed conclusion follows from the literature.

Keep three decisions separate:

1. Retrieval: where did these records come from? For a real search, save the database or issuer, query, filters, search date, export and source count. Six supplied records mean “these six records,” not “the literature.”
2. Screening: which records need closer reading against your stated criteria? The assistant suggests; you decide. Inspect proposed exclusions too. A false negative can hide the very source you needed.
3. Evidence use: after reading the original document, what exact claim can it support? A relevant abstract is not a quality appraisal or a licence to cite an unread result.

For real sources, match title, authors, date and identifier to an authoritative bibliographic record or issuing organisation, then open the original source. Check the relevant passage, document version and any visible correction or retraction notice. A DOI that resolves is an identity lead, not proof that the text supports your sentence. If access is missing, retain the record as unresolved; do not ask the assistant to reconstruct the paper.

Stable record IDs let you connect an export row, a proposed label, a quoted passage and a human decision. Preserve duplicate rows in the raw export and map them to the same work where you can verify that identity. Similar titles alone do not establish a duplicate, and a preprint and later article may need explicit version handling.

Language and code remain substantive research responsibilities

For drafting, a language pass may improve clarity while preserving your meaning. Stop when it adds an interpretation, causal claim, limitation, comparison or citation you did not establish. “More fluent” does not mean “more accurate.” Accept only wording you understand and can support, and retain proper attribution for quoted or paraphrased material.

For coding, generated code is an untested draft. Read it before running it in an approved, isolated working folder. Check its inputs, operations and output against small known cases; retain the accepted code with its inputs, output and environment. A plot that looks plausible can still drop observations, swap labels or calculate the wrong quantity. One successful run supports that run, not a general reproducibility claim.

Use T04-L01: Ask Your Files and Check the Evidence for exact-source document work, T08-L01: Make One Useful Code Change for an understood and reversible code contribution, or T10-L01: Question a Table Without Losing the Rows for row and total checks. All three are published routes; choose the one your question needs. T15-L02: Reproducibility with AI in the loop develops versions, parameters, hashes, reruns and reported differences.

Start with the applicable rules, not a universal AI label

Before using real material, open the target journal or venue's current author instructions and relevant AI policy in your browser. Also check institutional, collaboration, funder, ethics and data requirements. Save the page title, URL, access date and relevant wording. Agree permitted uses with co-authors before touching shared work. If rules conflict, apply the stricter compatible requirement; if the requirements cannot both be met, ask the venue and responsible institutional contact before proceeding.

Do not paste unpublished manuscripts, peer-review submissions, participant or patient data, confidential company research, credentials or licensed full text into an unapproved service. Permission depends on the account, contract, material and purpose, not just a familiar product name. Public accessibility does not automatically permit third-party upload or reuse. Use the synthetic material here if permission is unclear; T12-L01: What you must never paste covers that boundary.

Policy examples checked on 11 September 2026:

• Scientific Reports' AI policy (opens in a new tab), like the Nature Portfolio framework (opens in a new tab), distinguishes assistive uses from evaluative or interpretive uses requiring oversight, verification and disclosure. It prohibits replacement of accountable scholarly judgement. The worked policy decision below applies this to screening; the journal does not explicitly name our exercise.
• ICMJE's author guidance (opens in a new tab) says authors should describe AI use in the cover letter and submitted work in the appropriate section if applicable. AI tools cannot fulfil authorship responsibilities. Humans must review output, ensure attribution and avoid treating generated material as a primary source. Check whether and how your venue applies this guidance.
• Elsevier's journal policy (opens in a new tab) distinguishes basic grammar, spelling and punctuation checks from substantive changes to sentence structure or organisation. It requires a separate declaration for reportable manuscript-preparation uses and detailed Methods reporting for AI used in the research process. These locations do not automatically apply to another publisher.

The name of a task is not its risk classification. Sorting records by year differs from judging their relevance; correcting a typo differs from writing the interpretation. Disclosure also does not make a prohibited use acceptable.

An AI tool cannot approve the final manuscript or answer for its integrity, so it is not a scholarly author. Human authors must meet the venue's authorship criteria and accept responsibility. Do not delegate peer-review critique or judgement to an assistant. Reviewer permissions are a separate policy question, and another person's submitted manuscript remains confidential.

Keep the log smaller than the regret

Immediately after each use, note the date; displayed tool, model and provider where available; purpose; input description and data class; saved prompt/output location; checks; and retained result. If a model version is not exposed, record “not exposed,” not a guessed version. If reconstructing past work, mark missing details “unknown” and ask collaborators or inspect saved records.

A use log supports an honest account; it does not guarantee another person can recreate a response. Services, routing and model behaviour can change. Keep the original proposal and the accepted version separately so that a later reader can see your contribution.

Runnable exercise

Choose the step and prepare your workspace

You need a text editor or word processor, a browser for official policies, and an approved assistant if you choose to use one. T01-L02: Trust but verify is useful preparation. No terminal commands are needed for the screening route.

In your file manager, create a new folder named research-step-practice. In your editor, create training-sources.txt for the supplied inputs and research-step.md for your question, proposed output, human decisions and policy note. Keep the input file unchanged. Save any actual conversation as assistant-output.txt in the same folder using your tool's available copy or export action. The manuscript does not create these files for you.

For your own task, substitute a small approved source set, author-written paragraph, or table/code question. Record why the input is permitted and define an observable result: “a reading queue for these records,” “a clearer paragraph with the same claims,” or “a chart preserving these observations.” If AI use is not allowed, complete the step manually and record that decision.

1. Fix one question and the screening boundary

The optional design slice is literature triage with citations: retain source identity beside every proposal and human decision. Here, citations mean references to supplied training passages. They are not scholarly references, and no installed screening application is implied.

Copy this task into research-step.md:

TRAINING ONLY — no literature search or experiment has been performed.

Question:
Which supplied records warrant full-document reading for a comparison of
automated single-cell tracking with manually annotated reference tracks
in time-lapse plant-root images?

Candidate criteria:
- Plant-root time-lapse images.
- Individual cells linked across frames (tracking, not just segmentation).
- A reported comparison with manually annotated reference tracks.

Decision rules:
- Possibly relevant: supplied text supports the criteria or plausibly fits
  but needs a full-document check.
- Exclude from this comparison: supplied text explicitly establishes a
  mismatch. Give the quoted reason.
- Unresolved: information needed for a decision is missing.
- Duplicate: verified same source key in this training set; retain its row
  and point to the representative record.

Only a human makes final decisions. Missing information is not exclusion.
Scope: the six supplied records, not a comprehensive search.
Output: a short reading queue plus reasons for every other record.

“Segmentation” means identifying objects in an image; “tracking” links objects across successive images. A source can contain the right organism and image type yet answer the wrong methodological question. Define that distinction before judging the records.

2. Supply six clearly synthetic records

Copy this entire block into training-sources.txt. Every title, source key and passage below is invented for instruction. There are no real papers, authors, journals, DOIs or scientific results in this set. Do not add realistic-looking bibliographic details.

SYNTHETIC TRAINING-SOURCE SET v1
Created for this exercise; not retrieved from a scholarly database.
Local record locators: training-sources.txt, record ID, passage label.
Source keys are fixture identities, not DOIs or publication identifiers.

Record S01
Source key: TRAIN-A
Title: Training note A — root cell trajectories
Abstract A1: We link individual cells across time-lapse plant-root images
and compare the resulting trajectories with manually annotated reference
tracks. This training abstract supplies no performance results.
Full-document excerpt F1: The comparison unit is a cell trajectory across
frames. The reference tracks were annotated by people. Error definitions,
sample size and numerical results are not supplied in this fixture.

Record S02
Source key: TRAIN-B
Title: Training note B — animal cell trajectories
Abstract A1: We compare automated and manually annotated tracks in
time-lapse images of animal cells. Plant-root images are not studied.
Full-document excerpt: not supplied.

Record S03
Source key: TRAIN-C
Title: Training note C — root image annotations
Abstract A1: We evaluate automated plant-root image analysis against
manual annotations using frames from a time-lapse sequence.
Full-document excerpt F1: The evaluation uses one selected frame from each
sequence. We compare cell boundaries in that frame. Cells are not linked
between frames, and no reference trajectories are evaluated.

Record S04
Source key: TRAIN-D
Title: Training note D — incomplete tracking record
Abstract A1: We track individual cells in time-lapse plant-root images.
The comparison and reference-annotation details are missing from this
supplied abstract fragment.
Full-document excerpt: unavailable in this training set.

Record S05
Source key: TRAIN-A
Title: Training note A — root cell trajectories
Abstract A1: We link individual cells across time-lapse plant-root images
and compare the resulting trajectories with manually annotated reference
tracks. This training abstract supplies no performance results.
Full-document excerpt F1: Same source and version as S01; see S01 F1.

Record S06
Source key: TRAIN-E
Title: Training note E — overview of root imaging
Abstract A1: This overview discusses segmentation and tracking in plant-root
imaging. It reports no primary comparison with manually annotated reference
tracks and no new benchmark measurements.
Full-document excerpt: not supplied.

For a real source set, retain real source URLs and bibliographic identities instead of these fixture keys. Save your actual query and export if you searched. Do not copy the training provenance into real research records.

3. Ask for a bounded proposal in the conversation view

Before sending, confirm that the assistant account is approved for these inputs. Start a new conversation for the task. Paste the task, followed by this prompt and the record IDs, source keys, titles and Abstract A1 passages only. Leave the full-document excerpts in your input file for your subsequent reading step.

Make a screening proposal for the fixed training set below using my criteria.
These are synthetic records, not real scholarly evidence. Do not browse.

Return one row for every supplied record ID with:
record ID; source key; proposed status; exact supporting quote from its
supplied abstract; missing information; next human action.

Use possibly relevant, exclude from this comparison, unresolved, or
duplicate of [record ID]. Missing detail does not prove a mismatch.
Do not invent citations, methods, numerical results or full-document access.
Do not make final inclusion decisions or scientific conclusions.
Show duplicate mappings without deleting input rows. Preserve all IDs.
End with a proposed reading queue and the limits of this fixed set.

[Paste the task and six abstract records here.]

Save the actual response separately. Its wording and even its categories may differ from the reference below. If it invents a source, drops an ID or asserts it searched the literature, reject that part and return to the fixed inputs.

4. Turn the proposal into a useful human reading queue

In your editor, compare every row, including proposed exclusions, with the corresponding abstract in training-sources.txt. Then read the supplied full-document excerpts for possible candidates. A missing excerpt remains a limitation. The miniature excerpts make the distinction teachable; a real project requires the original document and enough context to judge the claim.

The following is an authored reference result, not a record of an assistant run or real research execution:

The useful handoff is short:

TRAINING reading handoff — reference result, not a real literature finding

Read next: S01 / TRAIN-A.
Why: supplied passages describe the requested tracking/reference comparison.
Need next: full methods, error definitions, sample size and results.

Resolve access/details: S04 / TRAIN-D.
Why: tracking fits; comparison information is missing, not disproved.

Excluded from this comparison: S02, S03, S06, for the recorded reasons.
Duplicate row retained: S05 -> S01.

Reconciliation: 6 input rows = 1 reading candidate + 1 unresolved
                  + 3 excluded + 1 duplicate row.
Distinct fixture sources: 5. Candidate is not final evidence inclusion.
No performance claim, complete-search claim or real citation is supported.

You now know what to read, what to obtain, and why the other records are not in this comparison. That is a useful screening result even though it cannot answer which tracking method is better.

5. Improve one failure without changing the research question

A plausible bad proposal is: “S04: exclude; no manual reference comparison.” The source actually says the detail is missing. Excluding it would turn uncertainty into an unsupported negative finding.

If this occurs, keep the original proposal and append a correction in research-step.md. Use a follow-up such as:

Revisit S04 only. Quote the words that establish a mismatch with the
comparison criterion. If no supplied passage establishes that mismatch,
change the proposal to unresolved and state the missing information.
Do not invent or retrieve a full document. Leave other record IDs intact.

Whether or not the assistant corrects itself, the human decision is unresolved because the needed evidence is absent. Check that the revised handoff still accounts for all six rows. An assistant's apology is not the check.

If your run already handled S04 correctly, use S03 as the improvement: move its abstract-level possibility to a human exclusion using F1, preserving both stages and their reasons. You have added information through reading rather than silently rewriting the screening history. Do not invent a model mistake to complete the exercise.

6. Check an actual target policy and decide what this use contributed

Permission should already be settled for the material you used. Now check the reporting decision against the actual output and retained contribution.

In your browser, open your intended journal or report venue's current policy. For practice, use Scientific Reports: Artificial Intelligence (AI) (opens in a new tab). This is a real policy used for a hypothetical target, not a claim that the synthetic exercise is a manuscript suitable for submission there. Its editorial policies page (opens in a new tab) states that the journal follows Nature Portfolio policies unless otherwise noted and links to this journal-specific AI page.

The author read these pages on 11 September 2026. Make your own access-date entry when you open them; do not borrow that date as evidence of your check.

For the reference result in this book, the decision is: this is an authored synthetic training demonstration; no actual research screening or scholarly submission occurred. There is no real study disclosure to submit. Its AI manuscript authorship is recorded under “How this book was made.”

For your actual exercise, log whether you used an assistant or worked manually. A classroom submission may have its own reporting rules. For a real research workflow using AI relevance proposals, the worked policy mapping supports a disclosure decision, with human accountability and a confirmed reporting location. If the institution, account approval or submission placement is still unresolved, mark the record not ready for submission and name the contact who can resolve it.

In research-step.md, fill this contribution note from observed work:

Question and useful result:
Input provenance and scope:
Date; displayed assistant/tool; model/provider if exposed:
AI contribution (or no AI used):
Saved prompt/output location (or not applicable):
My source, meaning or calculation checks:
What I accepted:
What I rejected or changed, and why:
Unresolved limitation and next action:
Target venue; policy title; URL; my access date; relevant wording:
Institution/collaboration rule and any unresolved decision:
Disclosure decision and task-specific reason:
Required location, or unresolved location and responsible contact:
Human author/co-author confirmation status:

For a real fixed-set screening contribution, a statement scaffold might be:

We used [actual tool/service; displayed model if available] to propose
relevance labels for [actual number and provenance of supplied records]
against [actual criteria]. [Responsible human authors] checked the proposals,
including proposed exclusions, against [materials actually inspected] and
made the final screening decisions. We retained [actual contribution].
[State actual unresolved access or scope limits where relevant.]
The human authors take responsibility for the reported work.

Replace every bracket from records, adapt it to the venue's requested form, and place it only where the current instructions require. Do not say that authors read full papers if they saw only abstracts. Do not claim that the assistant performed no search unless your saved session supports that statement. A transparent record can still describe an incomplete research step.

Alternative useful steps: the same responsibility in writing and code

The screening exercise is enough. These alternatives preserve the original reading, drafting and coding pathways without requiring three assignments.

Language contribution. In your editor, start with this synthetic author-written paragraph:

In this training table, control values were 10 and 12 and treatment values
were 14 and 16. The treatment values were higher in these four observations.
This small synthetic example does not establish a causal effect or support
generalisation to a population.

In the approved assistant's conversation, ask:

Improve readability without adding or removing claims, numbers, uncertainty
or citations. List every meaning-changing proposal separately. Do not add
a causal claim or a claim of statistical significance.

An acceptable authored reference revision is: “The synthetic control values were 10 and 12; the treatment values were 14 and 16. Treatment values were higher in these four observations, but this example establishes neither a causal effect nor a population-level conclusion.” Compare every sentence with the input before accepting it. Reject “the treatment increased the outcome” because that asserts causation. Record the retained wording and actual policy decision: a structural rewrite is not necessarily an exempt spelling check.

Code contribution. In your new practice folder, create observations.csv with the original four synthetic observations:

condition,replicate,value
control,1,10
control,2,12
treatment,1,14
treatment,2,16

If you already have an approved analysis environment, ask for a small plotting-script draft for that environment. Specify all four individual observations, labelled conditions and replicate identities, no invented uncertainty bars, no added measurements, and no silent averaging. Save the draft separately from any original research code. Read the code before running it and reject unrelated file access, uploads or installations.

Check the displayed data against the CSV: control is 10 and 12; treatment is 14 and 16; four rows went in and four observations should be represented. If an optional summary is requested separately, the arithmetic means are 11 and 15, but two mean points do not satisfy the requested four-observation plot. A difference of 4 between those means is arithmetic on synthetic numbers, not an experimental conclusion. Repair a dropped observation or unwanted aggregation and recheck the affected output. If you cannot read or safely run the draft, keep it unaccepted and use T08 or an experienced collaborator before it enters the research.

Keep the final code, input, output and environment details together, and describe only checks you performed. No script or plotting execution is supplied or claimed by this book. Your contribution note should distinguish a code draft from accepted, executed code; the venue may require methodological detail as well as a declaration.

Company research setting. The same method works for a public evidence report. Jonas can use records from issuing organisations, retaining titles, dates and versions, then check each promising and rejected source. A superseded standard is not automatically the current requirement; keep its version and decide whether the report concerns current practice or history. A source describing reported experiences does not support rewriting the report to say a programme “improves productivity.” For a plotting exercise, Jonas can use the same synthetic numbers with comparison and pilot labels. These are optional fictional framings, not claims of an actual search, company policy or successful run. The reporting location comes from his real venue and company rules.

Recover when the research trail breaks

Exit check

Deliver one inspectable research-step record, research-step.md or an equivalent document. Include the useful result and contribution note, with references to the preserved inputs and actual assistant output where applicable. It must let another author trace:

• The question, input provenance and bounded scope.
• The accepted result: a reading queue with source-linked human decisions, a supported paragraph revision, or a checked table/code contribution.
• What AI actually proposed, what the human accepted or rejected, and one meaningful correction or improvement.
• The checks actually performed and the uncertainty that remains.
• The target policy title, URL, access date and relevant wording, followed by a task-specific disclosure decision and required location or a named unresolved placement question.
• Who remains accountable and whether co-author agreement or another decision is pending.

For the synthetic screening route, the reference reconciliation is six rows, five distinct fixture sources, S01 for further reading, S04 unresolved, S02/S03/S06 excluded from this comparison, and S05 mapped to S01. Neither S01's candidacy nor that count establishes a scientific conclusion. A different result needs a reason supported by the supplied passages, not an invented source.

It fails if a source vanishes without explanation, a missing detail becomes a confident exclusion, an unperformed check is claimed, a generated citation becomes evidence without inspection, or the note hides the assistant's real contribution. A clearly named unresolved policy question can pass the learning exercise as an unfinished submission decision; it cannot pass as ready to submit. A manual route can pass with a useful checked result and an honest “no AI used” decision.

Rule to remember: You are the author. It is a tool you must be able to describe.

Source notes

• Baseline: T15-L01: AI in your research workflow (opens in a new tab). This revision uses the complete supplied 9 September manuscript and the completion brief's exact live baseline body hash. It retains the original research-step choice, reading/writing/code distinctions, four-row synthetic plotting data, Lab and Company applicability, venue-policy teaching, use log, verification, authorship responsibility and recovery guidance.
• Example association: the brief maps the optional app-owner-new-literature-triage-with-citations design slice to this book. The manuscript uses source-linked triage and human decisions; it does not claim a shipped app, downloads, institutional adoption or an executed literature search.
• Real policy check, 11 September 2026: Scientific Reports AI policy (opens in a new tab) supplies the worked target-policy evidence. Its editorial policies opening (opens in a new tab) establishes the general Nature Portfolio relationship. The Nature Portfolio AI framework (opens in a new tab), ICMJE author guidance (opens in a new tab), and the author sections of Elsevier's journal policy (opens in a new tab) support the bounded comparisons above. These sources concern policy, not evidence for the fictional research question. No authenticated submission form or institutional policy was inspected.
• Retained reference routes: PubMed user guide (opens in a new tab) and Crossref REST API documentation (opens in a new tab) are routes for finding bibliographic identities; they were linked in the earlier edition and were not freshly checked for this revision. An index record is not proof of claim support. ICMJE reviewer guidance (opens in a new tab) is also a retained, not freshly checked, reference; the freshly read Scientific Reports policy supports this book's reviewer-confidentiality and judgement boundary. The Tools index is a catalogue, not approval to use a service for research data.
• Synthetic material: all six training records, their titles and excerpts, the reference screening output, paragraph revisions, numerical observations and fictional researcher framings were authored for instruction. They are not scholarly evidence. Source citations such as S03 F1 locate a training passage only.

Historical authorship and checks

The supplied prior manuscript credits Astra (OpenAI, openai/gpt-6-astra) with the 9 September 2026 L1 refinement: “One learner-chosen research step and contribution note, with links to T04/T08/T10; retain venue-policy disclosure and accountability.” It records the Payload body and metadata fetch and checksum at 2026-09-09T10:50:57.3688468Z and says external references were not freshly reverified in that correction. Its “prepared only” wording described that historical authoring stage; the completion brief identifies the edition as subsequently published.

The earlier policy paragraph recorded 4 September 2026 as its last verification date. That historical date remains distinct from the bounded 11 September source reads listed here. The supplied T15-L01 manuscript exposes no separate named Sol or Luna contribution; no role or identity has been invented or reassigned to the current author.

Unknowns and limitations

• No real papers were searched, screened or read for a research finding in this authoring session. The training set cannot establish retrieval completeness, screening accuracy on real literature, study quality, methodological validity or a scientific conclusion.
• The full-document excerpts are deliberately short teaching inputs. Real evidence use needs the original source, adequate context and the appropriate disciplinary appraisal or review method. This L1 exercise is not a systematic-review protocol.
• The current policy reading is bounded to the pages and sections identified above. Scientific Reports declaration placement, any current submission-form fields, your institutional permissions and co-author agreement remain unestablished. Policies can change; recheck the actual target at use and submission.
• No assistant screening session, plotting run, installation, classroom activity, screenshot or learner outcome was observed. The authored reference outputs are examples to compare with actual work, not execution evidence.
• This manuscript is a draft for coordinator integration. No author-side tests, lint, typecheck, validators or browser checks were run; the single consolidated final verification belongs to the coordinator. No publication or independent acceptance is claimed.

How this book was made

Coordinator editorial update, 11 September 2026: The independent Terra reader accepted this revision. The account below records authoring-stage work; learner execution and classroom outcomes remain unobserved.

Astra — OpenAI, openai/gpt-6-astra — authored this example-led revision on 11 September 2026 in one fresh OpenCode author session. The session read the completion brief, editorial style, learnings, template and complete supplied T15-L01 manuscript, then consulted the bounded primary-policy sources listed above. It wrote only this disjoint manuscript file with apply_patch; no extra agents participated. Historical authorship is credited separately above.

The production lesson was to make the research result useful before composing its disclosure. A six-record synthetic set exposes an uncertainty error, a duplicate and an abstract/full-document mismatch without fabricating scholarship. Checking the real target policy also exposed a reporting-location gap, which remains explicit rather than being filled with another publisher's rule. Coordinator review and consolidated checks are pending; token, cache, reasoning and cost telemetry are unavailable, not zero.