← EthicalAI
Working paper
Download catalogue paper · PDF · Download advisory auditing paper · PDF
Paper versions
Archived, navigable versions of each paper, with a changelog for every version. The newest version of the catalogue paper is also the live one.
EthicalAI: An Evidence-Linked Catalogue of AI Integrity Failures
Current live version
v1.12 · 2026-11-19 (latest)
Read this version · commit 6395b82
- Grounded the paper in the user's Six Principles of Ethical AI. Added contracts/ethical-principles.md (the canonical charter, above the AI Integrity Contract), and framed the citation-grounding finding as operationalising the honesty-and-transparency principle, descriptively rather than as a bolted-on manifesto (slop judge 4/10 -> fixed).
- The charter is referenced in the introduction as motivation and in Section 8 among the artifacts.
v1.11 · 2026-11-19
Read this version · commit 641865c
- De-slop pass toward a human scientific register. AI-slop judge 4/10 -> 2/10, all seven prior structural tells resolved: cut the 'three things to reuse' contribution list and 'This paper delivers' package sentence, replaced 'sits next to' / 'Its proposed value is that', consolidated the per-section scope disclaimers, unstacked the causes-relation negations, and rewrote the conclusion to close the opening problem instead of paraphrasing the abstract.
- Judge verdict: reads primarily as human scientific prose. Every verified number preserved.
v1.10 · 2026-11-18
Read this version · commit 39eed9b
- Submission roadmap Stage 6: produced the submission PDF (docs/paper.pdf) via pandoc + tectonic, with a native LaTeX/TikZ Figure 1 replacing the web SVG. 12-page PDF, all nine sections, Figure 1, and twelve references. arXiv/workshop ready.
v1.9 · 2026-11-18
Read this version · commit 45401d4
- Submission roadmap Stage 5 (writing polish): differentiated the conclusion from the abstract (no verbatim finding restatement; ends on the forward-looking rule plus the machine contrast); pinned the OWASP living-document citation with an accessed date. Length 5,219 words.
v1.8 · 2026-11-18
Read this version · commit 5b7983f
- Submission roadmap Stage 4: four-reviewer council re-review (peer minor-revision up from major; every number reconciles; ai-writing 2/10; one groundedness BLOCK).
- Re-grounded the LLM-citation probe to the saved run data (gpt-4o 40/45 declined, 5/5 fabricated; deepseek 7/8), committed both per-prompt jsonls; moved Figure 1 into the markdown source; consolidated the repeated scope disclaimer; cut the Section 3 meta-announcement; reframed the surface-match/machine link as a contrast. Rejected one flag after verifying a number was correct at the pinned commit.
v1.7 · 2026-11-18
Read this version · commit 8bba855
- Submission roadmap Stage 3: added Figure 1, an inline SVG bar chart of the five-class citation-failure distribution (surface-match dominates), with a caption and an in-text reference. Counts match the Section 3.1 table.
v1.6 · 2026-11-18
Read this version · commit 318585a
- Submission roadmap Stage 2: added a second-model probe. deepseek-chat never declined (0/45, vs gpt-4o's 85%) and fabricated 7 of 8 verified suggestions, reusing one arXiv id across unrelated modes.
- The two models fail in opposite styles (over-cautious vs over-confident) but neither produces a citation trustworthy without resolving the source. Abstract, Section 3 probe paragraph, and the probe writeup updated.
v1.5 · 2026-11-18
Read this version · commit 92efa87
- Submission roadmap Stage 1: citation integrity sweep. Verified all 12 references resolve to real papers on arXiv (titles, authors, years); fixed Agrawal's author list; confirmed Spracklen's 2025 is the USENIX venue year.
- Added a Section 8 note that every reference was resolved to its source before inclusion, the same discipline the paper applies to the catalogue.
v1.4 · 2026-11-18
Read this version · commit fc67571
- Paper-orchestra agent council revision (architect, literature, section writers, integrator, groundedness verifier), validated before publishing.
- Promoted the citation-grounding result to its own Section 3 and named the five-class failure typology (surface-match, scope-mismatch, misfiling, polarity-reversal, wrong-population) so it is a citable object.
- Positioned against the closest prior work with five arXiv-verified citations: Lewis 2020 (RAG), Rashkin 2021 (AIS attribution), Liu 2023 (verifiability audit), Gao 2023 (ALCE), Agrawal 2023 (reference hallucination).
v1.3 · 2026-11-13
Read this version · commit 5f4b1a7
- Added the LLM-citation probe: asked for references, GPT-4o declined on 55 of 65 prompts, and all 10 confident suggestions bound a plausible title to an arXiv id for an unrelated paper.
- Converted the keyword/LLM hook from an (ungrounded) analogy into a measured small-probe result; earlier groundedness BLOCK resolved.
v1.2 · 2026-11-11
Read this version · commit 6e2a9f1
- Added the citation-failure typology that turns the headline error rate into a diagnosis.
- Applied the review council's convergent fixes: reframed 29/50 as a count not a rate, stated the binning rule, added the kappa small-n caveat, fixed a source batch-arithmetic bug.
v1.1 · 2026-11-11
Read this version · commit 29c826d
- Reframed to lead with the transferable finding and the reader payoff instead of a flat construction log.
- Consolidated the defensive disclaimer cadence flagged by the AI-writing review.
v1.0 · 2026-11-03
Read this version · commit c36db6a
- Honest, map-led resource paper: the historical 207-record study, the citation audit, and the verification sweep.
- Baseline before this improvement cycle.
Contracts Between Humans and AI Agents
v1.0 · 2026-11-03 (latest)
Read this version · commit c36db6a
- Companion method paper: the clause-judge vs baselines ablation on six hard suites.
- Honest result: decomposition buys specificity, the named catalogue buys traceability; no dominance or SOTA claim.