Status: COMPLETE — all 6 phases shipped; PR #107 ready for merge (matrix 64/64 + design space 5 dimensions + 10 open questions + 4-anchor snapshot integrity verified) Date: 2026-05-13 Issue: #106 Author: Cheng-I Wu Scope: Discovery — NO schema commitment, NO implementation Discovery output is a PR adding this document; PR merge closes #106. Implementation issue (deferred) blocked-by #106.
This discovery is split into three execution phases. Each phase ships as a PR commit.
| Phase | Deliverable | Status |
|---|---|---|
| Phase 1 | Doc scaffold + 4 anchor snapshots locked (Wayback + SHA-256) + manifest | shipped (PR #107 c1) |
| Phase 2 | Fill cells for PRISMA-trAIce + ICMJE (32 of 64 cells) | shipped (PR #107 c2) |
| Phase 3a | Fill cells for Nature Portfolio (16 of remaining 32 cells; 48 of 64 total) | shipped (PR #107 c3) |
| Phase 3b.1 | Fill cells for IEEE (remaining 16 cells; 64 of 64 total) | shipped (PR #107 c4) |
| Phase 3b.2 | Deliverable 2 design space (5 dimensions A–E + trade-off matrix) | shipped (PR #107 c5) |
| Phase 3b.3 | Deliverable 3 open questions (6 baseline + 4 surfaced) | shipped (PR #107 c6) |
Phase 1 explicitly does NOT fill any cells. Cells carry phase: 1-scaffold placeholder until Phase 2/3. This is intentional — Phase 1 builds the provenance infrastructure so Phase 2/3 cell-filling can cite verbatim with byte-level integrity.
ARS currently has a minimal ai_disclosure field (boolean / narrative). Three downstream pressures suggest this may need structural expansion:
docs/design/2026-05-12-ars-v3.7.3-claim-faithfulness-and-contaminated-source-spec.md) records source provenance + citation anchors. Per-stage AI disclosure has been hypothesized as a complement on the production side, but should be tested rather than assumed.Honest about uncertainty: discovery output may conclude that the right answer is to align ARS to a single anchor verbatim, construct a hybrid, defer policy mapping to a renderer layer, OR retain the current minimal ai_disclosure field. All four outcomes are valid discovery results.
A prior draft attempted to propose a concrete schema directly. Codex pre-open review flagged 5 P1 structural problems. These were symptoms of skipping discovery. This document is the discovery reset.
These are reserved for the follow-up implementation issue, which is blocked-by this discovery's completion.
All explicit-mandate / explicit-recommend / conditional-mandate cells filled in Phase 2/3 MUST cite a snapshot from this provenance bundle. Snapshot manifest is canonical; live URLs may drift after capture.
Manifest: docs/design/snapshots/2026-05-13-ai-disclosure-discovery/manifest.yaml
| Anchor | Wayback ID | Local SHA-256 (truncated) | Scope |
|---|---|---|---|
| PRISMA-trAIce | 20260513075443 |
f95fc59f… |
SLR-specific living guideline |
| ICMJE | 20260513075516 |
52f9e6bc… |
Medical journals, broad |
| Nature Portfolio | 20260513075542 |
cf691cba… |
Nature Portfolio (publisher-level guidance) |
| IEEE | 20260513075605 |
3ab8db50… |
IEEE author guidelines |
Verification protocol (per manifest): re-compute SHA-256 of local mirror before writing a mandate/recommend cell; mismatch → re-fetch from wayback_url + re-hash before continuing.
Each cell carries:
| Field | Required? | Rule |
|---|---|---|
source_strength |
yes | explicit-mandate / explicit-recommend / conditional-mandate / implicit / not-addressed / unknown |
verbatim_quote |
yes for mandate/recommend/conditional cells | Paraphrase forbidden; ≤ 50 words verbatim from primary source |
inference_passage |
yes for implicit cells |
The passage(s) used for inference, verbatim |
locator |
yes for mandate / recommend / conditional / implicit cells | Section / item / heading within source (e.g., "PRISMA-trAIce Table 1 M6.a"). not-addressed cells use — because the read scope is the entire source (per the not-addressed rule below); unknown cells use — because the source was not fully read. |
expected_value_type |
yes for mandate / recommend / conditional / implicit cells | boolean / structured / narrative. not-addressed / unknown cells use —. |
conditional_trigger |
required for conditional-mandate |
Plain-text trigger description |
snapshot_ref |
required for mandate / recommend / conditional / implicit cells | {anchor_slug}:wayback={wayback_id}. not-addressed / unknown cells do not require snapshot_ref (the read-scope is the full snapshot, recorded once at the per-anchor matrix header). |
not-addressed rule: a cell may be not-addressed only when the matrix author has read the entire primary source AND can affirmatively state the field is absent. Cells where the source has not been fully read must be unknown.
unknown ceiling: matrix may close with up to 10 of 64 cells marked unknown (subjective ceiling, adjustable upward per #106 acceptance criteria if author finds source genuinely hard to fully read — document why).
These 16 fields anchor the 4×16 = 64-cell minimum matrix. Additional fields may be added during Phase 2/3 if material to ARS's design decision; subtractions require rationale.
Snapshot: prisma-trAIce @ wayback 20260513075443
Source-strength caveat: PRISMA-trAIce is a pre-Delphi proposal (Holst et al. 2025, §Methods: "this work is explicitly intended as a well-founded proposal that provides an immediate solution and can serve as the foundation for a subsequent, community-driven formal consensus process"). It has not undergone formal consensus. No item carries explicit-mandate strength — the framework explicitly invites community refinement. Items within the framework use reporting-guideline directive verbs ("describe", "report", "include") classified here as explicit-recommend within the proposed framework. The framework itself is conditional-mandate on "AI tool used as methodological tool in an SLR" — outside that condition the framework does not apply. Cells below distinguish item-level recommend from framework-level conditional via the conditional_trigger column.
| # | Field | Source strength | Verbatim quote (or inference passage) | Locator | Value type | Conditional trigger |
|---|---|---|---|---|---|---|
| 1 | AI tool name | explicit-recommend | "For each AI tool or system used: a. Specify the name, version number (if applicable), and developer/provider." | Table 1, M2.a | narrative | AI tool used in SLR |
| 2 | AI tool version | explicit-recommend | "Specify the name, version number (if applicable), and developer/provider." | Table 1, M2.a | narrative | AI tool used in SLR; "if applicable" softens to recommend-when-versioned |
| 3 | AI tool developer / manufacturer | explicit-recommend | "Specify the name, version number (if applicable), and developer/provider." | Table 1, M2.a | narrative | AI tool used in SLR |
| 4 | Stage / phase of use | explicit-recommend | "For each AI tool, clearly describe: a. The specific SLR stage(s) where it was applied (e.g., search, screening, data extraction, Risk of Bias assessment, synthesis, drafting)." | Table 1, M3.a | structured (SLR stage enum) + narrative | AI tool used in SLR |
| 5 | Specific task within stage | explicit-recommend | "b. The precise task(s) the AI was intended to perform at each stage." | Table 1, M3.b | narrative | AI tool used in SLR |
| 6 | Affected manuscript sections / content locator | implicit | (inference passage) Items T1, A1, I1, R1, D1, D2 each prescribe a manuscript section for disclosure but the framework does not require an explicit per-task content-locator field. | Table 1 row headings (Title / Abstract / Introduction / Results / Discussion) | narrative | — |
| 7 | Date(s) of use | not-addressed | (Read entire checklist Table 1 + Methods §; no item requests date-of-use disclosure. Confirmed against M1–M10, R1–R2, D1–D2, T1, A1, I1.) | — | — | — |
| 8 | Prompts | explicit-recommend | "For each LLM/GenAI tool used, report: a. The full prompt(s) employed for each specific task. If prompts are extensive, provide a detailed description of their structure, key instructions, context provided..." | Table 1, M6.a | narrative (or repository link) | LLM/GenAI tool used (M6 header: "Prompt Engineering (if any)") |
| 9 | Human oversight method | explicit-recommend | "Describe the process of human interaction with and oversight of the AI tool(s) at each stage: a. How many reviewers interacted with/validated the AI outputs for each task? b. Did reviewers work independently when validating AI outputs? c. What were the qualifications or training of reviewers..." | Table 1, M8.a–g | narrative | AI tool used in SLR |
| 10 | Human responsibility statement | not-addressed | (Read entire Methods §; M8 covers oversight process but no item assigns explicit author-responsibility statement for AI output correctness. Distinct from ICMJE which mandates this separately.) | — | — | — |
| 11 | Performance evaluation method | explicit-recommend | "Describe methods used to evaluate the AI tool(s) performance for the specific tasks within the review (if applicable and feasible). This may include: a. The reference standard used for evaluation... b. The metrics used..." | Table 1, M9 | narrative | AI tool used in SLR; "if applicable and feasible" softens further |
| 12 | Performance evaluation results | explicit-recommend | "Report the results of any performance evaluations of the AI tool(s) for the specific tasks within the review (as described in P-trAIce M9). Include quantitative results (see M9) and measures of agreement between AI and human reviewers if assessed." | Table 1, R2 | structured (quantitative metrics) + narrative | M9 performance evaluation was conducted |
| 13 | Limitations / known failure modes | explicit-recommend | "Discuss any limitations encountered in using the AI tool(s) (eg, technical issues, biases identified, challenges in prompt engineering, unexpected outputs, limitations in AI performance for specific sub-tasks)." | Table 1, D1 | narrative | AI tool used in SLR |
| 14 | Disclosure location | implicit | (inference passage) Each Table 1 item is row-categorized by manuscript section (Title / Abstract / Introduction / Methods / Results / Discussion), implying section-of-record per item. No single "disclosure_location" field; location is structural property of the checklist itself. | Table 1 row groupings | structured (manuscript section enum) | — |
| 15 | Copyediting exemption predicate | not-addressed | (Read entire checklist; PRISMA-trAIce scope is "AI as methodological tool in an SLR"; copyediting exemption is not in scope — distinct concern handled by venue-level policies, not this framework.) | — | — | — |
| 16 | AI-generated image / figure / content rights | implicit | (inference passage) "Describe how data handled by AI tools (input, output, intermediate data) was managed and stored, and any measures taken to ensure data privacy, security, and compliance with copyright or terms of service, especially when using third-party cloud-based AI tools." M10 covers copyright/terms-of-service compliance for data handled by AI tools — touches AI-generated content rights indirectly without dedicated field. | Table 1, M10 | narrative | AI tool used in SLR |
Snapshot ref: all explicit-recommend / implicit cells above cite prisma-trAIce:wayback=20260513075443 (sha256: f95fc59f…).
Cell count: 16 cells filled; source-strength distribution: 0 explicit-mandate / 10 explicit-recommend / 0 conditional-mandate / 3 implicit / 3 not-addressed / 0 unknown.
Snapshot: icmje @ wayback 20260513075516
Source-strength caveat: ICMJE Recommendations are adopted (not draft) and used by 1000+ journals as baseline. Strength language is consistent: "should require", "should describe", "must ensure", "is not acceptable". The page on AI Use by Authors (§V.A) is short — just two paragraphs — and is deliberately framework-level, not field-by-field. Many disclosure fields ARS tracks (tool version, prompts, performance metrics, date of use) are not addressed by ICMJE because ICMJE delegates section-level detail to individual journals. Where ICMJE uses "should require" addressed at journals, the effective strength to authors (post journal adoption) is explicit-mandate; where it addresses authors directly with "should describe" / "must ensure" / "is not acceptable", strength is recorded as written.
| # | Field | Source strength | Verbatim quote (or inference passage) | Locator | Value type | Conditional trigger |
|---|---|---|---|---|---|---|
| 1 | AI tool name | implicit | (inference passage) "the journal should require authors to disclose at submission whether they used AI-assisted technologies (such as LLMs, chatbots, or image creators)". Tool identity is implied by "such technology... how they used it" but no field-level requirement to name the specific tool. | §V.A, paragraph 1 | narrative | AI-assisted technology used |
| 2 | AI tool version | not-addressed | (Read entire §V.A; no mention of version-level granularity. ICMJE delegates this detail to individual journal instructions.) | — | — | — |
| 3 | AI tool developer / manufacturer | not-addressed | (Read entire §V.A; no mention of developer/manufacturer disclosure.) | — | — | — |
| 4 | Stage / phase of use | implicit | (inference passage) "Authors who use such technology should describe... how they used it". "How they used it" implies stage but no enumerated stage taxonomy. | §V.A, paragraph 1 | narrative | AI-assisted technology used |
| 5 | Specific task within stage | implicit | (inference passage — same source as #4) "how they used it" implies task-level description without explicit field. | §V.A, paragraph 1 | narrative | AI-assisted technology used |
| 6 | Affected manuscript sections / content locator | not-addressed | (Read entire §V.A; "the submitted work in the appropriate section if applicable" prescribes a disclosure section but not per-content-affected-location.) | — | — | — |
| 7 | Date(s) of use | not-addressed | (Read entire §V.A; no date-of-use disclosure required.) | — | — | — |
| 8 | Prompts | not-addressed | (Read entire §V.A; no prompt-disclosure requirement. ICMJE delegates this to journal-level policy.) | — | — | — |
| 9 | Human oversight method | explicit-recommend | "Authors should carefully review and edit the AI-generated content as the output can be incorrect, incomplete, or biased." | §V.A, paragraph 1 | narrative | AI-assisted technology used |
| 10 | Human responsibility statement | explicit-mandate | "Therefore, humans are responsible for any submitted material that included the use of AI-assisted technologies." | §V.A, paragraph 1 | narrative or boolean | AI-assisted technology used |
| 11 | Performance evaluation method | not-addressed | (Read entire §V.A; no performance-evaluation method disclosure required.) | — | — | — |
| 12 | Performance evaluation results | not-addressed | (Read entire §V.A; no performance-evaluation results disclosure required.) | — | — | — |
| 13 | Limitations / known failure modes | not-addressed | (Read entire §V.A; output limitations mentioned in oversight context — "output can be incorrect, incomplete, or biased" — but no requirement to disclose specific limitations encountered.) | — | — | — |
| 14 | Disclosure location | explicit-recommend | "Authors who use such technology should describe, in both the cover letter and the submitted work in the appropriate section if applicable, how they used it" | §V.A, paragraph 1 | structured (closed enum: cover-letter + manuscript-section) | AI-assisted technology used. Note: ICMJE's "should describe" reads as recommend at the language layer; adopting journals may upgrade to mandate via their own instructions. |
| 15 | Copyediting exemption predicate | not-addressed | (Read entire §V.A; no copyediting carve-out. ICMJE treats all AI-assisted technologies uniformly within §V.A — distinct from Nature §AI authorship which carves out "AI assisted copy editing" at no-disclosure strength, and from IEEE Paragraph 2 which carves out editing/grammar-enhancement at downgrade-to-recommend strength. Both Phase 3a/3b verifications confirmed the carve-out is anchor-specific.) | — | — | — |
| 16 | AI-generated image / figure / content rights | explicit-mandate | "Humans must ensure there is appropriate attribution of all quoted material, including full citations." + "Referencing AI-generated material as the primary source is not acceptable." | §V.A, paragraph 1 | narrative + negative-constraint flag | AI-assisted technology used (text or image). Both clauses are mandate strength ("must ensure" + "is not acceptable"). |
Snapshot ref: all explicit-mandate / explicit-recommend / implicit cells above cite icmje:wayback=20260513075516 (sha256: 52f9e6bc…).
Cell count: 16 cells filled; source-strength distribution: 2 explicit-mandate / 2 explicit-recommend / 3 implicit / 9 not-addressed / 0 unknown.
Cross-anchor observation: ICMJE's distribution skews toward not-addressed (9/16) by design — it's deliberately framework-level, delegating field-level granularity to adopting journals. This contrasts with PRISMA-trAIce (9 explicit-recommend) which provides item-level granularity within its SLR-only scope. The two anchors are complementary, not competitive: ICMJE sets the policy floor (must disclose, must take responsibility); PRISMA-trAIce sets the content ceiling (these are the specific things to report when AI is used).
Snapshot: nature @ wayback 20260513075542
Source-strength caveat: The Nature Portfolio AI editorial policy is adopted (not draft) and applies across all Nature Portfolio journals. The author-facing surface is structured as four short sections: AI authorship, Generative AI images, AI use by peer reviewers, Editorial use. ARS's matrix covers author-side obligations only (sections §AI authorship + §Generative AI images); peer-reviewer-side use and Springer Nature's own editorial use are out of scope per §2.2. Strength language is mixed: "do not currently satisfy our authorship criteria" + "unable to permit" + "must be labelled" carry mandate strength; "should be properly documented" + "should be disclosed" + "does not need to be declared" carry recommend strength. The policy is framework-level for text (Methods-section documentation is called for via should-language, with no field-level granularity — no version / no developer / no prompt / no performance metric requirement) and prohibition-level for images (generative AI images banned by default with three carve-outs; non-generative ML manipulations carry a recommend-strength caption-disclosure rule). Many text-side disclosure fields ARS tracks are therefore not-addressed; image-side rights field carries a hard explicit-mandate carrying a negative-constraint + labelling-mandate composite.
| # | Field | Source strength | Verbatim quote (or inference passage) | Locator | Value type | Conditional trigger |
|---|---|---|---|---|---|---|
| 1 | AI tool name | implicit | (inference passage) "Use of an LLM should be properly documented in the Methods section (and if a Methods section is not available, in a suitable alternative part) of the manuscript." Documentation requirement implies tool identity disclosure but no field-level name requirement. | §AI authorship | narrative | LLM used (non-copyediting) |
| 2 | AI tool version | not-addressed | (Read entire §AI authorship + §Generative AI images; no version-level granularity required. Nature delegates Methods-section detail to author and journal-level instructions.) | — | — | — |
| 3 | AI tool developer / manufacturer | not-addressed | (Read entire §AI authorship + §Generative AI images; no developer/manufacturer disclosure required.) | — | — | — |
| 4 | Stage / phase of use | implicit | (inference passage) "Use of an LLM should be properly documented in the Methods section". "Use" implies stage/task description in Methods but no explicit stage taxonomy. | §AI authorship | narrative | LLM used (non-copyediting) |
| 5 | Specific task within stage | implicit | (inference passage — same source as #4) "properly documented in the Methods section" implies task-level description without dedicated field. | §AI authorship | narrative | LLM used (non-copyediting) |
| 6 | Affected manuscript sections / content locator | implicit | (inference passage) "in the relevant caption upon submission" for non-generative ML image tools; for LLM text use, Methods-section documentation implies which sections used AI but no per-content-affected-location requirement. | §Generative AI images (caption rule); §AI authorship (Methods rule) | narrative | AI tool used |
| 7 | Date(s) of use | not-addressed | (Read entire §AI authorship + §Generative AI images; no date-of-use disclosure required.) | — | — | — |
| 8 | Prompts | not-addressed | (Read entire §AI authorship + §Generative AI images; no prompt disclosure required. Methods-section documentation requirement does not enumerate prompts as a required element.) | — | — | — |
| 9 | Human oversight method | explicit-mandate | "In all cases, there must be human accountability for the final version of the text and agreement from the authors that the edits reflect their original work." | §AI authorship | narrative | LLM / AI-tool used in manuscript text ("In all cases" generalizes the rule beyond the copyediting carve-out) |
| 10 | Human responsibility statement | explicit-mandate | "an attribution of authorship carries with it accountability for the work, which cannot be effectively applied to LLMs" + "there must be human accountability for the final version of the text" | §AI authorship | narrative or boolean | LLM used in manuscript |
| 11 | Performance evaluation method | not-addressed | (Read entire §AI authorship + §Generative AI images; no performance-evaluation method disclosure required.) | — | — | — |
| 12 | Performance evaluation results | not-addressed | (Read entire §AI authorship + §Generative AI images; no performance-evaluation results disclosure required.) | — | — | — |
| 13 | Limitations / known failure modes | not-addressed | (Read entire §AI authorship + §Generative AI images; §Generative AI images mentions general legal-copyright and research-integrity risks but does not require authors to disclose tool-specific limitations or failure modes. AI-tool limitations are characterized in §AI use by peer reviewers — out of scope for author-side matrix.) | — | — | — |
| 14 | Disclosure location | explicit-recommend | "Use of an LLM should be properly documented in the Methods section (and if a Methods section is not available, in a suitable alternative part) of the manuscript." + (for non-generative ML on images) "should be disclosed in the relevant caption upon submission" | §AI authorship + §Generative AI images | structured (closed enum: Methods + image-caption + suitable-alternative) | AI tool used; location varies by use type |
| 15 | Copyediting exemption predicate | explicit-recommend | (carve-out) "The use of an LLM (or other AI-tool) for "AI assisted copy editing" purposes does not need to be declared." (predicate, partial verbatim) "AI-assisted improvements to human-generated texts for readability and style... [but] do not include generative editorial work and autonomous content creation." | §AI authorship | structured (boolean carve-out + predicate definition) | AI use limited to copyediting-as-defined |
| 16 | AI-generated image / figure / content rights | explicit-mandate | "Springer Nature journals are unable to permit its use for publication." + "All exceptions must be labelled clearly as generated by AI within the image field." | §Generative AI images | structured (default-prohibit + 3-carve-out enum + labelling-required flag) | Generative AI image/video proposed for publication (default-deny); non-generative ML caption rule handled at #14 as recommend-strength |
Snapshot ref: all explicit-mandate / explicit-recommend / implicit cells above cite nature:wayback=20260513075542 (sha256: cf691cba…).
Cell count: 16 cells filled; source-strength distribution: 3 explicit-mandate / 2 explicit-recommend / 0 conditional-mandate / 4 implicit / 7 not-addressed / 0 unknown.
Cross-anchor observation: Nature Portfolio's distribution skews to not-addressed (7/16) for the same structural reason as ICMJE — it is a framework-level policy that delegates field-level granularity to Methods-section author practice and journal instructions, not a reporting-guideline with item-level disclosure schema. Nature's distinctive contribution to the 4-anchor set is the image-rights mandate package (#16: default-deny + 3 carve-outs + labelling-required), which is more prescriptive than ICMJE's text-only attribution rule (ICMJE #16 forbids citing AI as primary source; Nature #16 forbids publishing the image at all unless a carve-out applies). The copyediting exemption predicate (#15) is the second distinctive contribution: Nature is the first anchor in the matrix to define an explicit carve-out boundary ("AI assisted copy editing" with prose definition) — PRISMA-trAIce and ICMJE both leave this not-addressed. Three-anchor hybrid emerging: ICMJE = policy floor (must disclose + must take responsibility), PRISMA-trAIce = SLR content ceiling (item-level granularity), Nature = section-of-record + carve-out shape (Methods location + copyediting predicate + image-rights regime).
Snapshot: ieee @ wayback 20260513075605
Source-strength caveat: IEEE's guideline page (dated April 16, 2024) is adopted (not draft) and applies across IEEE publications. The snapshot HTML mirror is 212 KB, but the substantive policy text is two short paragraphs (~150 words combined; webinar reference and Author Center pointer carry no normative content, video content is out of scope as it is a separate medium without byte-stable transcript). The policy is therefore the tersest of the four anchors: it makes 3 stacked mandates within the opening paragraph (Claim A: disclosure-in-acknowledgments mandate carrying images/figures/code in scope; Claim B: tool-identity-mandate; Claim C: section-locator + level-of-use mandate), then 1 recommend (Claim D: copyediting carve-out with disclosure-still-recommended). Many fields ARS tracks are not-addressed because IEEE delegates field-level granularity to author practice within the acknowledgments section. Notable absences vs the other three anchors: no explicit human-responsibility statement (no ICMJE-style "humans are responsible" sentence), no human-oversight requirement (no Nature-style "human accountability for the final version"), and no image-rights regime distinct from text (images/figures/code fold into the same acknowledgments-disclosure mandate; Nature-style default-deny on generative images is absent). The disclosure-location mandate is closed-enum tighter than Nature/ICMJE (acknowledgments only; no Methods / no cover letter / no caption alternatives).
| # | Field | Source strength | Verbatim quote (or inference passage) | Locator | Value type | Conditional trigger |
|---|---|---|---|---|---|---|
| 1 | AI tool name | explicit-mandate | "The AI system used shall be identified" | Paragraph 1, sentence 2 | narrative | AI-generated content used in article |
| 2 | AI tool version | not-addressed | (Read entire policy page — 2 paragraphs of substantive content; no version-level granularity required. IEEE delegates Acknowledgments-section detail to author practice.) | — | — | — |
| 3 | AI tool developer / manufacturer | not-addressed | (Read entire policy page; no developer/manufacturer disclosure required.) | — | — | — |
| 4 | Stage / phase of use | implicit | (inference passage) "accompanied by a brief explanation regarding the level at which the AI system was used to generate the content". "Level at which the AI system was used" implies stage/extent description but no enumerated stage taxonomy; "level" reads as degree-of-involvement rather than SLR-style stage enum. | Paragraph 1, sentence 2 | narrative | AI-generated content used in article |
| 5 | Specific task within stage | explicit-mandate | "specific sections of the article that use AI-generated content shall be identified and accompanied by a brief explanation regarding the level at which the AI system was used to generate the content" | Paragraph 1, sentence 2 | narrative | AI-generated content used in article |
| 6 | Affected manuscript sections / content locator | explicit-mandate | "specific sections of the article that use AI-generated content shall be identified" | Paragraph 1, sentence 2 | structured (per-section locator) + narrative | AI-generated content used in article |
| 7 | Date(s) of use | not-addressed | (Read entire policy page; no date-of-use disclosure required.) | — | — | — |
| 8 | Prompts | not-addressed | (Read entire policy page; no prompt disclosure required. "Brief explanation" of usage level does not enumerate prompts as a required element.) | — | — | — |
| 9 | Human oversight method | not-addressed | (Read entire policy page; no human-oversight method requirement. IEEE is silent on reviewer count, qualifications, or validation workflow — contrasts with Nature §AI authorship "In all cases, there must be human accountability" and PRISMA-trAIce M8 a–g.) | — | — | — |
| 10 | Human responsibility statement | not-addressed | (Read entire policy page; no explicit author-responsibility statement for AI output correctness. Contrasts with ICMJE "humans are responsible for any submitted material that included the use of AI-assisted technologies" and Nature "there must be human accountability for the final version of the text".) | — | — | — |
| 11 | Performance evaluation method | not-addressed | (Read entire policy page; no performance-evaluation method disclosure required.) | — | — | — |
| 12 | Performance evaluation results | not-addressed | (Read entire policy page; no performance-evaluation results disclosure required.) | — | — | — |
| 13 | Limitations / known failure modes | not-addressed | (Read entire policy page; no limitations / failure-modes disclosure required. IEEE makes no general statement about AI output reliability — contrasts with ICMJE "output can be incorrect, incomplete, or biased".) | — | — | — |
| 14 | Disclosure location | explicit-mandate | "shall be disclosed in the acknowledgments section of any article submitted to an IEEE publication" | Paragraph 1, sentence 1 | structured (closed enum: acknowledgments-section only) | AI-generated content used in article. Note: IEEE's enum is tighter than Nature (Methods or alternative) and tighter than ICMJE (cover letter + manuscript section). |
| 15 | Copyediting exemption predicate | explicit-recommend | "The use of AI systems for editing and grammar enhancement is common practice and, as such, is generally outside the intent of the above policy. In this case, disclosure as noted above is recommended." | Paragraph 2 | structured (boolean carve-out + predicate definition) + downgrade-strength flag | AI use limited to editing-and-grammar-enhancement-as-defined. Note: the carve-out does not eliminate disclosure — it downgrades strength from mandate to recommend while keeping location/content rules intact. |
| 16 | AI-generated image / figure / content rights | implicit | (inference passage) "The use of content generated by artificial intelligence (AI) in an article (including but not limited to text, figures, images, and code) shall be disclosed in the acknowledgments section" — images/figures/code fold into the same acknowledgments-disclosure mandate as text; no separate image-rights regime, no default-prohibition (contrasts with Nature §Generative AI images "unable to permit its use for publication"), no labelling-required flag distinct from the general acknowledgments rule. | Paragraph 1, sentence 1 | narrative (within general acknowledgments disclosure) | AI-generated content (text/figures/images/code) used in article |
Snapshot ref: all explicit-mandate / explicit-recommend / implicit cells above cite ieee:wayback=20260513075605 (sha256: 3ab8db50…).
Cell count: 16 cells filled; source-strength distribution: 4 explicit-mandate / 1 explicit-recommend / 0 conditional-mandate / 2 implicit / 9 not-addressed / 0 unknown.
Cross-anchor observation: IEEE's distribution skews to not-addressed (9/16) on a different basis from ICMJE — ICMJE is framework-level delegating to adopting journals; IEEE is framework-level delegating to author practice within the acknowledgments section. The two converge on the same not-addressed count by very different routes, which is itself a discovery signal: framework-level brevity reaches the same not-addressed floor regardless of delegation target. IEEE's distinctive contribution to the 4-anchor set is disclosure-location closed enum (#14: acknowledgments-only — the tightest enum across all four anchors; PRISMA-trAIce uses 6-section row enum across the checklist, ICMJE uses cover-letter + appropriate-section, Nature uses Methods + image-caption + suitable-alternative). The copyediting carve-out (#15) is also distinctive in its downgrade-not-eliminate construction: IEEE explicitly says disclosure is still recommended under the carve-out, whereas Nature says the carve-out use "does not need to be declared". This is the cleanest split in the 4-anchor set on whether copyediting carve-out means no disclosure (Nature) or softer-strength disclosure at the same location (IEEE). The absence of human-oversight (#9), human-responsibility (#10), and image-rights-regime (#16) means a schema chasing IEEE-only alignment would omit three fields that at least one of ICMJE / Nature treats at mandate strength: #9 is mandate only in Nature (ICMJE #9 is recommend, PRISMA-trAIce #9 is recommend); #10 is mandate in both ICMJE and Nature (PRISMA-trAIce leaves it not-addressed); #16 is mandate in both ICMJE (text-attribution clause) and Nature (default-deny + labelling). IEEE-only alignment loses all three at mandate strength. 4-anchor hybrid emerging: ICMJE = policy floor (must-disclose + must-take-responsibility) / PRISMA-trAIce = SLR content ceiling (item-level granularity) / Nature = section-of-record + carve-out shape (Methods location + copyediting predicate + image-rights regime) / IEEE = location closed-enum tightener + carve-out downgrade-not-eliminate construction.
COPE / CSE / WAME / JAMA / PLOS — partial expansion allowed; each cell column added must follow the same provenance rules. Snapshot capture deferred until Phase 3 author elects to include.
Enumeration of design choices across 5 dimensions (A–E), with trade-off matrix per option across at least 4 axes (alignment depth / schema simplicity / backward-compat burden / integration cost). Per #106, options are credible alternatives — not a quota — and choosing among them is OUT of scope for this discovery.
Each dimension below presents the §4-matrix evidence that bears on the option set (advisory, not prescriptive), then enumerates options, then offers a trade-off table. Where the matrix gives an asymmetric signal across the four anchors, the asymmetry is noted but not resolved.
Trade-off rubric (consistent across all 5 dimensions):
H (3–4) / M (2) / L (0–1).H simple (single field / closed enum), M moderate (structured object), L complex (event log / multi-field separation).H low burden (legacy stays valid), M moderate (migration tool with auto-mapping), L high (mass-edit required).H low cost (fits existing schemas), M moderate (one schema/agent touched), L high (multi-component change).Higher letter = better trade-off on that axis. No option is universally H/H/H/H — that is the discovery point. The dimensions are largely orthogonal but some option pairs interact (noted under each table).
Matrix evidence:
if applicable and feasible) and M6 (if any) attach conditional applicability to a recommend-strength field. Conditional structure exists in the matrix even before mandate strength.Options:
ai_used: true | false for each stage. Tool name / version / prompt / etc. are not modeled.{ ai_used: bool, tool_name: str, version: str?, prompts: [str]?, oversight_method: str?, … } with required sub-fields gated by ai_used: true.{ ai_used: bool, narrative: str? }. ARS's current shape extended with structured boolean.disclosure_tier: none | minimal | extensive with tier→fields mapping defined externally.disclosure_events: [{ stage_id, tool_id, task, timestamp, claim_anchor_id?, performance_metric_ref?, … }] linked to v3.7.x trust-chain anchors.Trade-off table:
| Option | Alignment depth | Schema simplicity | Backward-compat burden | Integration cost | Notes |
|---|---|---|---|---|---|
| A1 Boolean per stage | L | H | H | H | Satisfies no mandate beyond "ai used yes/no"; cannot carry IEEE #5/#6 task+section locator, ICMJE #10 responsibility statement, Nature #16 image rights. Renderer would have to invent fields. |
| A2 Structured object | M | M | L | M | Carries ICMJE/IEEE/Nature text-side mandates. Cannot carry one-sentence-→-many-facets pattern without redundancy (Nature #9 = #10). Required sub-fields gated by ai_used create flat schema with 8–12 optional fields. |
| A3 Hybrid bool+narrative | L | H | H | H | Today's ARS shape. Auditability low — narrative is free text. Satisfies no anchor at mandate strength beyond #14 location (only if narrative happens to include location). |
| A4 Per-stage tier | L–M | M | M | M | Tier mapping captures Nature not-addressed-heavy vs PRISMA-trAIce field-rich asymmetry, but the tier→fields rule lives outside the data — auditability depends on the external rule being versioned. Tier semantics drift if not pinned. |
| A5 Event-log | H | L | L | L | Carries v3.7.x trust-chain alignment, one-sentence-→-many-facets pattern natural (one event, multiple facet_ids), conditional applicability via applies_when predicate, IEEE #6 section locator native (event linked to claim anchor). High cost everywhere: schema is the most complex, migration from legacy boolean is mass-edit. |
Interaction notes: A1↔C1 (Boolean per stage + single policy field) collapses 4-anchor matrix to a single bit + a single string. A5↔E1 (Event-log + single ARS render) creates a renderer whose output may exceed what venue policy requires; an A5↔E4 (Event-log + schema-only) combination keeps the audit trail rich and lets the user format for venue manually — at the cost of no native render.
Matrix asymmetry not resolved by this dimension: the 4 anchors disagree at the field level about which structured sub-fields belong in the schema (PRISMA-trAIce demands version + developer + prompts + performance evaluation; ICMJE/Nature/IEEE all leave version + developer not-addressed). A1 underfits, A2/A5 risk overfit to PRISMA-trAIce-shaped data, A4 only works if tier is pinned to a specific anchor.
Matrix evidence:
Options:
other_label escape hatch — fixed list with a single fallback bucket carrying user-provided free text.systematic-review mode uses PRISMA-trAIce 6-stage enum; other modes use a 4-stage writing enum (research / drafting / revision / copyediting).research → search → screening for SLR; production → outlining → drafting for academic-paper).Trade-off table:
| Option | Alignment depth | Schema simplicity | Backward-compat burden | Integration cost | Notes |
|---|---|---|---|---|---|
| B1 Closed enum | M | H | L | M | Aligns with PRISMA-trAIce only; ICMJE/Nature/IEEE are agnostic so any closed enum is compatible at the policy layer, but a PRISMA-shaped enum may surprise non-SLR authors. Migration from legacy free-text narrative requires mass-classification. |
| B2 Open enum | L–M | H | H | H | Maximally compatible — every anchor is agnostic, every legacy entry survives. Auditability low — same stage may appear under 5 different labels across the corpus. |
| B3 Enum + escape hatch | M | M | M | M | Captures PRISMA-trAIce 6-stage alignment for SLR mode while letting other modes use other_label. Audit still soft because escape hatch encourages unsystematic use. |
| B4 Two-track by mode | H | L | M | M | Highest alignment depth: SLR mode satisfies PRISMA-trAIce field-by-field; other modes satisfy ICMJE/Nature/IEEE's stage-agnostic policy. Schema is most complex — two enums in one schema gated by mode. Mode-detection logic must be reliable. |
| B5 Hierarchical | M–H | L | L | L | Top-level satisfies ICMJE/Nature/IEEE; substage satisfies PRISMA-trAIce. Schema complexity is highest of the five options. Future expansion (COPE / WAME / JAMA) plugs in as new substage groups. |
Interaction notes: B1↔A2 (PRISMA-shaped closed enum + Structured object) keeps both dimensions at their simplest, and aligns well inside SLR mode; outside SLR the alignment comes from A2's anchor-mandate coverage, not from B1 (B1's stage labels do not match ICMJE/Nature/IEEE language). B4↔C5 (Two-track + Versioned policy profile) is the highest alignment depth combination — but introduces two cross-cutting axes (mode + policy profile) that must compose without conflict.
Matrix asymmetry not resolved by this dimension: the IEEE "level" semantics (degree-of-involvement, not stage) is not addressed by any of the 5 stage-taxonomy options. A schema that needs to satisfy IEEE #4/#5 at mandate strength must add a separate level_of_involvement field independent of stage taxonomy. This may belong in Dimension A (value model) rather than B.
Matrix evidence:
not-addressed. A schema that supports the carve-out without specifying which anchor variant is in force is ambiguous; a schema that hardcodes one variant under-aligns with the other.Options:
declaration_mode: strict | lenient | none — schema records the policy stance, not field-level structure.exempted_uses: ["copyediting", …] — schema records what is exempted, not what is required.{ ai_used_any: bool, disclosure_required: bool, exempted_uses: [], policy_aligned_with: anchor_id }. Each axis a separate field.{ policy_profile: anchor_slug, profile_version: date, profile_source: locator_url }. Policy is named, versioned, and source-anchored.Trade-off table:
| Option | Alignment depth | Schema simplicity | Backward-compat burden | Integration cost | Notes |
|---|---|---|---|---|---|
| C1 Single declaration_mode | L | H | H | H | "strict / lenient" hides which anchor's strictness; ambiguous at audit time. Copyediting carve-out is one bit, losing eliminate-vs-downgrade distinction. |
| C2 Carve-out list | M | H | H | M | Captures Nature/IEEE #15 carve-out shape natively. Cannot express Nature #16 default-deny regime — that is a mandate, not a carve-out. |
| C3 Multi-field separation | M | L | M | L | Each axis explicit; carve-outs + alignment anchor + strictness composable. Schema has 4+ correlated fields that may drift (exempted_uses set without disclosure_required: false is inconsistent). Validator burden is high. |
| C4 Defer to renderer | H | H | H | M | Schema stays minimal; policy mapping lives in renderer skill. Aligns with all 4 anchors simultaneously because schema does not commit to one. Audit-quality depends on renderer correctness, not data correctness — bug in renderer breaks all venue alignments at once. |
| C5 Versioned policy-profile | H | M | M | L | Each entry pins to a named profile (icmje-v2024 / nature-2025-01 / ieee-2024-04 / prisma-traice-pre-delphi); changes traceable. New profiles plug in. Profile registry is a separate artifact to maintain. |
Interaction notes: C4↔E2 (Defer + multiple per-policy renders) is the most policy-faithful combination — facts in data, policy in renderer, one renderer per anchor. C4↔E1 (single render) hides the policy choice behind a single rendered output. C5↔A5 (Versioned profile + Event log) gives temporal + policy provenance simultaneously; cost is the most complex schema in the design space.
Matrix asymmetry not resolved by this dimension: ICMJE's policy-routed strength (adopting journals upgrade) is not addressable by C1–C5 alone. A schema may need an effective_strength_via field separate from the policy_aligned_with field to distinguish ICMJE's source-level recommend from a specific journal's mandate uptake. Whether this matters for ARS depends on whether ARS authors target named journals (which would override anchor defaults).
Matrix evidence:
ai_disclosure field is boolean/narrative. Existing literature_corpus entries and bibliography entries may carry no AI disclosure at all.Options:
legacy: true marker, renderers degrade gracefully) — legacy entries flagged, renderer emits a "legacy disclosure" note.Trade-off table:
| Option | Alignment depth | Schema simplicity | Backward-compat burden | Integration cost | Notes |
|---|---|---|---|---|---|
| D1 Strict reject | M | H | L | M | Forces clean data from day-1. Breaks existing pipelines until users migrate every entry. Over-strict relative to anchor scope (no anchor requires retrospective disclosure). |
| D2 Hybrid (storage-yes / render-no) | L | M | M | M | Data survives; render hides legacy. Renderer output may be misleading — looks like no AI use when in fact disclosure status is unknown. |
| D3 Soft + legacy marker | M | M | H | H | All entries survive, renderer surfaces "legacy disclosure unavailable" honestly. Schema adds 1 boolean + acknowledgment text. Closest to anchor scope — forward-looking policy with explicit unknown surfacing. |
| D4 Forced backfill | L | H | L | L | Even more disruptive than D1; backfilling AI use that happened months ago may not be reliably recoverable. Most likely to produce fabricated retrospective disclosures, which violates ICMJE #10 (human-responsibility) in spirit. |
Interaction notes: D3↔A1/A3 (Soft + Boolean or Hybrid bool-narrative) lets ARS keep current shape and add the legacy marker as a single new field. D1/D4↔A2/A5 (Strict or Forced + Structured / Event-log) creates the harshest migration — schema change + retroactive data fill.
Matrix asymmetry not resolved by this dimension: the matrix is silent on whether no AI use is itself a disclosable statement. PRISMA-trAIce M1 implies "if AI tool used"; ICMJE/Nature/IEEE all condition on AI use. A schema that defaults to ai_used: false for legacy entries may be assigning a positive fact (no AI used) that was never recorded by the original author. Dimension D options interact with this concern: D3's legacy: true marker partially addresses it; D1/D4 do not; D2 hides the ambiguity.
Matrix evidence:
Options:
Trade-off table:
| Option | Alignment depth | Schema simplicity | Backward-compat burden | Integration cost | Notes |
|---|---|---|---|---|---|
| E1 Single ARS-native | L | H | H | M | Aligns with at most one anchor — others see render that mismatches their location/format requirements. Lowest user friction once user accepts ARS format. |
| E2 Multiple per-anchor | H | M | H | L | Each anchor's renderer is correct. Maintenance burden grows with anchor set (4 today; if expansion adds 5 candidates from §4.7, that's 9 renderers). Each renderer is a separate code path that can drift from anchor policy updates. |
| E3 ARS render + user template | M | M | H | M | Default render satisfies one anchor; advanced users override per-venue. Two-track UX may confuse users into thinking default render is universally venue-safe (it isn't). |
| E4 No renderer (schema only) | M | H | H | H | Schema-only output can in principle support any anchor (since user does mapping), but in practice depends entirely on author labour — alignment-depth in delivered manuscripts is unverifiable at the ARS layer. Honest about the fact that no single render fits all venues. Lowest cost to ARS infrastructure, but every user reimplements rendering per venue. |
Interaction notes: E4↔C4 (No renderer + Defer policy) is the most discovery-honest combination: schema records facts, user maps facts to venue. E2↔C5 (Multiple renders + Versioned policy profile) gives the most venue-correct combination but at the highest maintenance cost. E1↔A1 (Single render + Boolean per stage) is the lowest-cost combination but aligns with no anchor at mandate strength beyond bare disclosure presence.
Matrix asymmetry not resolved by this dimension: Nature #16's in-image-field labelling and #14's caption-disclosure cannot be produced by a schema-only output without a renderer — yet a renderer that touches image-field metadata or captions has to integrate with the manuscript-production pipeline (LaTeX / Word / Markdown source), not just emit a separate rendered text block. This is a scope question for the implementation issue (deferred): does ARS render into the manuscript source or produce a separate disclosure artifact?
The 5 dimensions are largely orthogonal, but the per-dimension Interaction notes (§5.1–§5.5) surfaced 13 distinct option-pairs with load-bearing properties. They are consolidated here for easy reference, in the order they first appeared. Pairs labelled X↔Y and Y↔X are treated as the same pair (e.g. C5↔A5 in §5.3 is the same pair regardless of which dimension surfaces it).
| Pair | Source § | Interaction |
|---|---|---|
| A1↔C1 | §5.1 | Boolean per stage + single policy field = collapses 4-anchor matrix to a single bit + a single string; lowest-information combination in the space. |
| A5↔E1 | §5.1 | Event-log + Single ARS render = renderer may emit more disclosure than the venue requires; under-aligns by overshooting. |
| A5↔E4 | §5.1 | Event-log + No renderer = keeps audit trail rich but pushes all rendering to user. |
| B1↔A2 | §5.2 | Closed enum + Structured object = simplest combined schema; alignment outside SLR comes from A2's anchor-mandate coverage, not from B1. |
| B4↔C5 | §5.2 | Two-track + Versioned profile = highest alignment-depth pairing; introduces two cross-cutting axes (mode + policy profile) that must compose without conflict. |
| C4↔E2 | §5.3 | Defer policy + Multiple per-anchor renders = most policy-faithful pairing (facts in data, policy in renderer, one renderer per anchor). |
| C4↔E1 | §5.3 | Defer policy + Single ARS render = hides the policy choice inside a single rendered output; auditability degraded. |
| C5↔A5 | §5.3 | Versioned profile + Event-log = temporal + policy provenance simultaneously; highest schema complexity in the space. |
| D3↔A1/A3 | §5.4 | Soft legacy + Boolean / Hybrid = lowest-friction migration path that preserves current ARS shape; adds the legacy: true marker as a single new field. |
| D1/D4↔A2/A5 | §5.4 | Strict / Forced legacy + Structured / Event-log = harshest migration; schema change combined with retroactive data fill. |
| E4↔C4 | §5.5 | No renderer + Defer policy = most discovery-honest combination; schema records facts, user maps facts to venue. |
| E2↔C5 | §5.5 | Multiple renders + Versioned policy profile = highest venue-correctness; highest maintenance cost (one renderer per profile per anchor). |
| E1↔A1 | §5.5 | Single render + Boolean per stage = lowest-cost combination but aligns with no anchor at mandate strength beyond bare disclosure presence. |
The 13 pairs are not exhaustive — the implementation issue (deferred) would need to enumerate the full Cartesian space (5×5×5×4×4 = 2,000 option combinations) only if a non-orthogonal interaction surfaces during selection.
These questions are open by design — Deliverable 2 enumerated credible options without choosing among them; Deliverable 3 names the questions whose answers should drive that choice. The 6 baseline questions from #106 are kept verbatim and given matrix-derived framing context. Phase 2/3 cell-filling surfaced 4 additional questions that were not in the baseline set and are added below. No question here is rhetorical. Each one is a genuine open issue where the matrix evidence supports more than one defensible answer.
A GitHub Discussions thread will be linked from #106 when this PR merges. Discovery closure requires the thread to be linked and open, not "answered" or "consensus reached" (per #106 acceptance criteria).
Q1 — Should ARS schema force a single declaration policy, or be policy-agnostic?
The 4-anchor matrix gives no clear answer. The four anchors' disclosure-location enums do not intersect (PRISMA-trAIce 6-section / ICMJE cover-letter+section / Nature Methods+caption+alt / IEEE acknowledgments-only). Forcing a single policy means picking one anchor and losing the others; staying policy-agnostic means deferring policy semantics to a renderer layer (Dimension C option C4) and trusting downstream correctness. The matrix is silent on whether ARS authors target named journals — if they do, ICMJE's policy-routed strength (adopting journals upgrade) would make a journal-aware schema preferable to either pure single-policy or pure policy-agnostic.
Q2 — Should stage taxonomy be SLR-centric, general, or hierarchical?
PRISMA-trAIce specifies a 6-stage SLR taxonomy (M3.a); ICMJE / Nature / IEEE all use unstructured stage language. The matrix supports SLR-centric (B1) only for systematic-review mode; for other ARS modes (deep-research / academic-paper), a B2/B3/B4/B5 option may serve better. Open question: should ARS treat systematic-review mode as a special case, or unify under one hierarchical taxonomy (B5) that absorbs SLR as a sub-tree?
Q3 — Is per-stage prompt disclosure mandatory or conditional on policy choice?
PRISMA-trAIce M6 calls for prompts (or detailed structure) when LLM/GenAI is used, but at explicit-recommend strength only (pre-Delphi). ICMJE / Nature / IEEE do not mandate prompts. A schema that mandates prompts unconditionally over-aligns with PRISMA-trAIce (and treats a pre-consensus framework as binding). A schema that omits prompts loses the SLR-specific content ceiling. Open question: should the schema gate prompts on mode == systematic-review AND tool_type ∈ {LLM, GenAI}, or leave both gates to user/renderer?
Q4 — What's the failure mode if disclosure is incomplete? (Pipeline halt / warning / silent log)
This is a runtime-policy question that ties to Dimension D options. D1 / D4 mean pipeline halt or forced backfill (strictest). D2 means render skip (silent). D3 means warning surfaced with a legacy: true marker. None of the 4 anchors specifies a failure mode — they describe what to disclose, not what happens when disclosure is absent. Open question: does ARS prefer to be strict (halt + force user attention), defensive (silent log + audit later), or transparent (warning + render the gap)?
Q5 — Backward compatibility horizon: how long do we accept legacy boolean entries?
ARS has existing literature_corpus entries with boolean/narrative ai_disclosure. The matrix does not impose retrospective disclosure obligation (Dimension D evidence). Open question: is "indefinite" the right horizon (D3 soft acceptance forever), or should the implementation issue set a deprecation date (e.g., 12 months from schema-issue close → reject legacy)? Implementation cost trades off audit honesty.
Q6 — Is a hybrid schema warranted? This is itself a hypothesis to test.
The 4-anchor matrix surfaces three structural facts that argue for some form of hybrid:
- ICMJE / Nature / IEEE all skew to not-addressed (9, 7, 9 of 16) but each contributes a distinct shape no other anchor provides. A single-anchor schema loses two-thirds of the design space.
- PRISMA-trAIce field-rich content does not transfer cleanly outside SLR.
- 4-anchor #14 location enums have empty intersection.
These three facts make "single-anchor verbatim alignment" hard to defend. But "hybrid" itself can mean Dimension A2 (Structured object covering ICMJE/Nature/IEEE text-side) OR A5 (Event-log with anchor-conditional applicability) OR B4 (Two-track by mode) OR C5 (Versioned policy profile). Open question: which hybrid axis (data model / stage taxonomy / policy expression) is the load-bearing one?
These questions did not appear in the #106 baseline. They emerged from specific matrix patterns observed during cell-filling.
Q7 — Does the copyediting carve-out have semantics distinct from boolean exemption?
Two of the 4 anchors address the copyediting carve-out (Nature, IEEE); the other two (PRISMA-trAIce, ICMJE) leave it not-addressed. Among the two addressed anchors, Nature eliminates disclosure ("does not need to be declared") and IEEE downgrades strength while preserving location and content ("disclosure as noted above is recommended"). If ARS schema supports the carve-out, it must choose whether exempted_uses: [copyediting] means "boolean off" (matches Nature) or "strength downgrade preserving location/content" (matches IEEE). Open question: should the carve-out be a single field or a struct with strength + location preserved?
Q8 — Should the schema model "level of involvement" separately from stage?
IEEE #5 carries explicit-mandate strength for a "brief explanation regarding the level at which the AI system was used to generate the content"; §4.6 #4 interprets that level language as degree-of-involvement rather than workflow stage (and classifies #4 as implicit on this basis). The distinction matters because degree-of-involvement (e.g., "full drafting" vs "outline only" vs "paragraph-level suggestion") is not the same axis as workflow stage. None of Dimensions A–E's options model degree-of-involvement separately from stage. If IEEE #5's mandate is in scope, the schema needs a 17th field (level_of_involvement or similar) that may or may not belong in Dimension A.
Q9 — Should image-rights regimes be unified or anchor-specific?
ICMJE #16 (text attribution + no-AI-as-primary-source) and Nature #16 (default-prohibit publication + 3 carve-outs + labelling) handle the same field by structurally different policies. IEEE #16 folds images into the general acknowledgments mandate. PRISMA-trAIce #16 is implicit. Open question: does ARS unify image-rights into a single field (losing Nature's regime-specific shape) or carry per-anchor variants (with anchor-conditional rendering)?
Q10 — Is "no AI use" itself a disclosable statement, or is silence default-OK?
The matrix is silent on whether legacy entries with no AI disclosure should be interpreted as "no AI was used" (positive fact) or "disclosure status unknown" (epistemic gap). All 4 anchors describe forward-looking obligations conditional on AI use. ARS's current schema treats absence of disclosure as silence. Open question: does the new schema require an explicit ai_used: false for the no-AI case, or accept silence as backward-compatible default?
These 10 questions are input to the implementation issue, not closure gates for this discovery. Discovery closes when the matrix + design space + open-questions thread are linked. Whether the implementation issue answers all 10, a subset, or reframes them entirely is a separate decision the community + maintainer thread will surface.
Phase 1 was scaffold + provenance only. It deliberately filled zero cells. The infrastructure it locked:
snapshots/manifest.yaml)Phase 2 fills 32 of 64 cells: PRISMA-trAIce (§4.3) and ICMJE (§4.4). Both anchors fully read end-to-end before any not-addressed cell was emitted.
Cell distribution summary (32 cells total across two anchors):
| Anchor | explicit-mandate | explicit-recommend | conditional-mandate | implicit | not-addressed | unknown |
|---|---|---|---|---|---|---|
| PRISMA-trAIce | 0 | 10 | 0 | 3 | 3 | 0 |
| ICMJE | 2 | 2 | 0 | 3 | 9 | 0 |
| Phase 2 total | 2 | 12 | 0 | 6 | 12 | 0 |
Key observations (advisory; Deliverable 2/3 will weigh these):
unknown cells: both anchors were fully readable in single sessions (PRISMA-trAIce 14 checklist items + Methods §; ICMJE §V.A two paragraphs). The 10-cell unknown ceiling from §4.1 stays available for Nature + IEEE in Phase 3 if needed.not-addressed skew on ICMJE (9/16): expected given ICMJE's deliberate framework-level scope. This is a discovery signal, not a flaw in ICMJE — it means a schema chasing ICMJE-only alignment would have a small core (mandate-only) and large optional space.explicit-mandate in PRISMA-trAIce: by design, the proposal is pre-Delphi. Any schema that treats PRISMA-trAIce items as hard requirements would be over-reading the source.Acceptance for Phase 2 (commit 2 of PR #107): 32 cells filled with source-strength classification + verbatim quote (or inference passage) + locator; SHA-256 of 4 snapshot HTML files unchanged from Phase 1; document parses as valid Markdown.
Phase 3a fills 16 more cells (Nature Portfolio §4.5), bringing the running total to 48 of 64 cells. Nature was fully read end-to-end across §AI authorship and §Generative AI images before any not-addressed cell was emitted. Sections §AI use by peer reviewers and §Editorial use are out of scope per §2.2 (author-side obligations only).
Cell distribution after Phase 3a (48 cells across three anchors):
| Anchor | explicit-mandate | explicit-recommend | conditional-mandate | implicit | not-addressed | unknown |
|---|---|---|---|---|---|---|
| PRISMA-trAIce | 0 | 10 | 0 | 3 | 3 | 0 |
| ICMJE | 2 | 2 | 0 | 3 | 9 | 0 |
| Nature Portfolio | 3 | 2 | 0 | 4 | 7 | 0 |
| Phase 3a total | 5 | 14 | 0 | 10 | 19 | 0 |
Key observations from Nature (advisory; Deliverable 2/3 will weigh these):
not-addressed. Nature is the first anchor to define an explicit carve-out boundary with prose definition. Dimension C (Phase 3) needs to handle whether copyediting carve-out is in scope.Acceptance for Phase 3a (commit 3 of PR #107): 48 cells filled cumulatively; Nature SHA-256 unchanged from Phase 1; document parses; codex gpt-5.5 xhigh review converged to 0 findings in 2 rounds.
Phase 3b.1 fills the final 16 cells for IEEE (§4.6), bringing the matrix to 64 of 64 cells complete. IEEE was fully read end-to-end — the substantive policy text is 2 short paragraphs (~150 words combined) inside <div itemprop="articleBody">. The 212 KB snapshot HTML is mostly IEEE.org site chrome / menu / footer. The webinar reference and Author Center pointer carry no normative content; the linked video is out of scope as a separate medium without byte-stable transcript.
Cell distribution after Phase 3b.1 (64 cells across four anchors):
| Anchor | explicit-mandate | explicit-recommend | conditional-mandate | implicit | not-addressed | unknown |
|---|---|---|---|---|---|---|
| PRISMA-trAIce | 0 | 10 | 0 | 3 | 3 | 0 |
| ICMJE | 2 | 2 | 0 | 3 | 9 | 0 |
| Nature Portfolio | 3 | 2 | 0 | 4 | 7 | 0 |
| IEEE | 4 | 1 | 0 | 2 | 9 | 0 |
| Matrix total | 9 | 15 | 0 | 12 | 28 | 0 |
Key observations from IEEE (advisory; Deliverable 2/3 will weigh these):
not-addressed. This is the cleanest binary split across the 4 anchors and informs Dimension D (legacy disclosure handling): how should ARS treat copyediting-only AI use?not-addressed count converges at 9 for both framework-level anchors (ICMJE, IEEE) by different delegation routes. PRISMA-trAIce (3 not-addressed) and Nature (7 not-addressed) sit below this floor because they specify more fields directly. This suggests there is a structural ceiling on what framework-level brevity can specify, independent of which audience the policy delegates to. Dimension B (Phase 3) stage taxonomy can leverage this: a stage taxonomy that aligns with framework-level scope will be not-addressed-heavy regardless of which framework is chosen.unknown cells across the entire matrix: All four anchors were fully read end-to-end. The 10-cell unknown ceiling from §4.1 went unused.Acceptance for Phase 3b.1 (commit 4 of PR #107): 64 of 64 cells filled cumulatively; IEEE SHA-256 (3ab8db50…) unchanged from Phase 1; document parses; codex gpt-5.5 xhigh review converged to 0 findings in 2 rounds (Round 1 P1 = cross-anchor "mandate in ICMJE and Nature" overstatement fixed to "at least one of ICMJE/Nature treats at mandate strength").
Phase 3b.2 builds §5 Deliverable 2 design space: 5 dimensions A–E (Value-type / Stage taxonomy / Policy expression / Legacy handling / Renderer target), each with 4–5 credible options enumerated, each option scored across 4 trade-off axes (alignment depth / schema simplicity / backward-compat burden / integration cost). §5.6 consolidates 13 distinct option-pairs surfaced as load-bearing in the per-dimension Interaction notes. No option is selected.
Acceptance for Phase 3b.2 (commit 5 of PR #107): 5 dimensions, 23 options, 4-axis scoring per option, 13-pair interaction inventory; document parses; codex gpt-5.5 xhigh review converged to 0 findings in 3 rounds (Round 1 = 6 findings incl. 3 P1; Round 2 = 2 P2; Round 3 = 0).
Phase 3b.3 surfaces §6 Deliverable 3 open questions: 6 baseline questions from #106 (Q1–Q6) verbatim with matrix-derived framing context, plus 4 additional questions (Q7–Q10) surfaced from Phase 2/3 cell-filling patterns not in the baseline set. §6.3 scope note clarifies these 10 questions are input to the implementation issue, not closure gates for this discovery.
Acceptance for Phase 3b.3 (commit 6 of PR #107): 6 baseline questions verbatim + 4 surfaced; each question has matrix-evidence context paragraph and remains genuinely open (more than one defensible answer per matrix); no implicit option selection. Discovery acceptance criterion from #106 is "thread linked and open" — the GitHub Discussions thread will be linked when this PR merges.
docs/design/2026-05-12-ars-v3.7.3-claim-faithfulness-and-contaminated-source-spec.md (locator infrastructure prior art)