# Adversarial review of the historical IL graph Date: 25 September 2026. Object reviewed: the delivered historical transcription, its source bindings, geometry and interface. Criteria: source fidelity, preservation of historical distinctions, no invented inferential relations, readable and inspectable presentation, reproducible construction. The anti-sycophancy protocol is applied to findings rather than used as a certificate of quality. ## Findings that changed the artifact ### 1. The newer archive missed explicit annex numerals The archive’s annex inventory said that the final table supplied visual ordering but no explicit numeric 7/8 labels. Direct extraction of the original EMF shows `7` at the ethics row and `8` at the metacognition row. The source image confirms this. Meanwhile, both manuscript bodies give metacognition 7 and ethics 8. **Consequence:** normalizing to a single numbering convention would falsify the historical record. The delivered data retain separate body and annex numbers, expose a numbering selector, preserve both ethics headings, and use `A`-prefixed competency codes to identify annex numbering unambiguously. The underlying manuscript files were not edited. ### 2. The source crosswalk is not a uniform endorsement graph Annex 3 has X marks, empty cells and a hatched NILIS column. A six-model synthesis does not warrant connecting every model to every synthesized competency. The earlier visual inventory is insufficient to infer missing cell values. **Consequence:** only the 39 actual X marks produce model–competency connections. Blank and hatched cells remain in each competency’s record and the source table. NILIS remains a source-model node and retains its many actual Annex 2 occurrences. Hatching is not interpreted as “NILIS contributes nothing.” ### 3. Source indicators are not the author’s fifteen synthesized competencies Annex 2 groups quoted source-model indicators by pillar. Annex 3 separately states the synthesized competencies. Treating the first as a complete author-supplied decomposition of each competency would add a mapping absent from the historical record. **Consequence:** all 181 occurrences remain attributed to their models and classified under their source pillar. No fine indicator-to-competency edge is invented. Repeated codes stay separate when they occur in different cells or pillars. ### 4. Removing a sequence is not enough to encode nonlinearity The first graph correctly avoided a pillar-1 → pillar-2 → … chain, but a hub-and-spoke PRI still failed to represent the stronger manuscript statement that a person may navigate from one operational pillar to another at any point. Absence of a sequence is weaker than mutual navigability. **Consequence:** the six operational pillars now share one plane and a faint **presentation-only navigability mesh** joins every pair. Those fifteen lines are explicitly not historical ledger assertions. They visualize a single global prose claim rather than laundering it into fifteen separately sourced pairwise claims. The canonical PRI→pillar relations remain unchanged. ### 5. Initial geometry obscured some competency labels In the first rendered competency view, the metacognition position lay too close to the finding/access competency cluster. In the initial crosswalk, full competency names competed for space between the pillar and source-model columns. **Consequence:** the competency layout now separates those clusters. The crosswalk uses the annex codes, with complete wording in the selected record and index. Long connections in the sparse historical-context view were also routed laterally so their default projection does not look like a chain passing through unrelated intermediate nodes. At the inspected 1440-pixel desktop viewport, all 29 crosswalk nodes have visible labels. The larger views report labels omitted for collision avoidance; they do not omit nodes or records. ### 6. A view change could leave an invisible selection in the sidebar The initial implementation kept the model root selected when switching to a view that did not include it. The detail text was still accurate but no longer corresponded to a visible selection. **Consequence:** changing views selects a visible default when necessary. Browser checks exercise this behavior in all six views. ### 7. Generic validation did not catch a full-schema naming defect The first ledger passed the supplied generic validator but failed full JSON Schema validation: the ledger identifier lacked the `ledger:` prefix, source-model canonical node IDs contained uppercase letters, and the `historical.sourceModel` term type contained a capital letter. These were construction defects, not historical-source defects. **Consequence:** canonical IDs and types were brought into the schema’s required syntax. Historical display labels and source-model names were not lowercased or otherwise altered. Presentation IDs are retained as explicit annotations. Full schema validation was rerun after the repair. `qa/schema-repair.json` preserves the initial failure; `qa/package-checks.json` records the final check. ### 8. Lexical strengthening would change the model The ethics/legal competency does not say that an information-literate person obeys law. Its wording is: “comprend et tient compte des enjeux éthiques et légaux qui entourent l’usage de l’information.” Converting this into legal compliance would add a normative duty absent from the source and would erase the distinction between legality and ethical justification. **Consequence:** exact French competency wording remains authoritative; the English translation remains “understands and takes into account the ethical and legal issues surrounding the use of information.” The HTML contains a reviewer comment explicitly prohibiting paraphrases such as “obey the law” or “comply with the law.” The wire enclosure is documented as a visualization of the source verb “encadre,” not enforcement, permission, legal obedience, or moral correctness. ### 9. Post-PRI reflection requires an inter-iteration reading The manuscript says metacognition exists outside the PRI after the search, where the person reflects on the experience and identifies what could improve a later process. A terminal “after PRI” node preserved location but not the learning loop. **Consequence:** the historical node and canonical metacognition→after relation remain unchanged. A separate dashed, directed **future-iteration guide** now connects post-PRI reflection toward PRI. It is explicitly editorial and refers to a later iteration, not a return inside the same search episode. ### 10. Arbitrary depth was removed The previous layout alternated positive and negative z-values among peer pillars mainly for visual separation. That created spatial differences with no source meaning. **Consequence:** the six first pillars are coplanar. Depth now separates presentation layers: operational structure, synthesized competencies, source provenance, quoted indicators, and manuscript context. Crosswalk geometry stays planar because its source is a table and added depth would imply a relation absent from that table. ### 11. “All” is now explicitly a provenance view The complete view is numerically dominated by quoted source-indicator occurrences. Treating it as the primary model view made raw documentary granularity look like conceptual centrality. **Consequence:** Structure and Competencies are the model-reading views. All is labelled and arranged as complete provenance, with layer separation rather than a degree-like hairball. Source models do not become conceptually more important merely because their source table contains more quoted occurrences. ## 12. Reader-facing bilingual parity and manuscript attribution The 1.1.0 source reader exposed the historical source record primarily through a French-only interface. The main graph was bilingual, but the reader-facing source page was not. The model table also lacked an explicit bibliographic attachment to the two manuscripts requested for attribution. This was a presentation-completeness defect rather than a defect in the canonical model. **Consequence:** revision 1.1.2 keeps the historical model content unchanged but reorganizes the reader surface. The eight-pillar/fifteen-competency model table and its two manuscript references now appear on the main page above the graph, under the title “Integrated information literacy model / Modèle intégré d’infolittératie,” with explicit wording that the model is interconnected and nonlinear rather than a step-by-step procedure. The duplicate model block was removed from `sources.html`. Annex 2 is rendered as compact side-by-side comparison tables matching the manuscript analysis structure while preserving all 181 source occurrences and their anchors. Internal review and canonical-ledger links were removed from the main reader navigation; the underlying review and ledger files remain in the package for build, provenance, and QA use. ## Checks with no substantive transcription finding The 15 competency texts match the final annex’s vector text. The 181 indicator occurrences match their original code and text cells. The seven annex EMFs are byte-identical between the two source copies. The selected body paragraphs match their original XML text. The 5.1 truth-value formulation was not replaced with the assistant’s later criticism or with a modernized version. The generic schema, validator, projector and query module match the supplied foundation byte for byte. The reference stylesheet is extracted unchanged; the additional stylesheet is a separate file. Navigation relations bind to source records and do not license inference. The canonical ledger has no fabricated historical ChangeEvents. ## Executed tests - **39 original-source checks:** exact hashes of both DOCX copies, all 23 selected paragraphs, and each of seven annexes against both manuscripts. Output: `qa/original-sources.json`. - **39 structural/historical tests:** source text and mark checks; namespace, reference and contract validation; deterministic projection; all-view parity; per-pillar occurrence counts; finite, repeatable and noncollapsed XYZ positions. The suite includes negative mutations for renumbering, altered wording, omitted records, fabricated X marks, invented mappings, a false linear sequence, stronger inference, rewritten 5.1 epistemology, missing language labels and contract-hash spoofing. Output: `qa/node-tests.tap`. - **74 Chromium checks:** startup without network or WebGL dependency; browser-visible guide-layer invariants; numbering, language and accent-insensitive search; preservation of original French source text; finite projections and visible selections across all views; the 29-label crosswalk; 221-node / 453-relation parity; the complete index; per-pillar indicator-occurrence counts; expanded view; responsive overflow checks at 1440, 900, 390 and 320 pixels; source-reader model and competency table counts; bilingual reference-language marking; English/French source-reader switching; manuscript-translation exposure; responsive source-reader overflow; source-reader record and image counts; and absence of runtime exceptions or external requests. Output: `qa/browser-tests.json`. **72 dependency-free package checks** verify local links and anchors, byte-identical foundation files, the retained reference stylesheet, deterministic rebuilding, absence of redistributed fonts, the canonical ledger validator, revision metadata, guide/ledger separation, and the lexical-fidelity warning. Output: `qa/package-checks.json`. The optional Python `jsonschema` harness remains included for a full Draft 2020-12 schema pass, but that dependency was unavailable in the execution container for revision 1.1 and was not represented as rerun. The exact counts describe the executed tests, not a general certification of the historical model. Revision 1.1.2 retains the historical-source tests and adds browser checks for main-page model placement, removal of internal reader links, bilingual model parity, compact Annex 2 rendering, preservation of all 181 indicator anchors, and removal of the duplicate model table from the source reader. The current Chromium suite passes 73 checks. ## Visual inspection The screenshots in `qa/` are actual Chromium renderings of the delivered application assets. The desktop French and English pages, all six views, expanded mode and responsive layouts were inspected. The source table showing the 7/8 reversal was also inspected visually rather than inferred solely from text extraction. Changes were made after the first inspection, then the final browser checks and captures were rerun. The final package retains the resulting screenshots, not obsolete pre-fix screenshots represented as current output. The visual harness drives system Chromium through the DevTools Protocol. Exact local CSS and JavaScript bytes are inlined into `about:blank`; application code and source data are unchanged. This avoids dependence on file/HTTP navigation policies while still exercising the delivered renderer and interactions. Native local-link integrity is checked separately. The shipped page has no remote dependencies, fetch-based data loading or ES-module loading requirement. A software 3D renderer is used instead of the original Three.js/CDN runtime. This choice is explicit in the README. It projects real XYZ coordinates and supports perspective orbit/pan/zoom; it is not a static illustration masquerading as an interactive graph. The reused foundation is the canonical construction machinery and visual system, not an unverified claim of runtime identity. ## Disposition The canonical historical ledger remains conservative. Revision 1.1.2 preserves the canonical historical ledger and changes only the reader-facing presentation. Revision 1.1 changed the **lens** rather than fabricating new historical content: nonlinearity, post-PRI learning and semantic depth are rendered in a separate presentation layer; exact source wording retains authority over editorial paraphrase. The remaining gap is intrinsic to the historical material: the prose describes competencies as detailed by components, but the supplied final historical tables do not provide a clean canonical synthesized-component layer that can be added without reconstruction. That layer remains absent rather than invented.