Version History

Versioning Policy

The ARGUS version number follows three levels.

• Major version (X.0.0): an analysis produced under the previous version is no longer reproducible or comparable under the new one. This level covers the merging, removal or renumbering of a step, the introduction of a new class of flaw, a change to the standard of proof, and a change to the overall procedure.
• Minor version (X.Y.0): addition of a test, section or optional appendix that leaves unchanged the judgments made under the previous version.
• Patch version (X.Y.Z): clarification, drafting rule, precaution, correction of wording.

A published analysis must always state the full version number under which it was produced. This policy is declared from V4.0.0 onwards and is not applied retroactively: earlier entries are retained as published.

V5.0.0 — August 2026

Why a major version. This version was prepared as V4.1.0, that is, as an addition of tests leaving unchanged the judgments made under the previous version. It is not. Several of its changes mean that an analysis produced under V4.0.0 is no longer comparable with one produced under this version, which is exactly the criterion for a major version declared above.

• The relevance threshold in 0.a has been raised. Texts that V4.0.0 classified as strongly relevant may now fall under partial relevance and enter analysis only after validation of a scope.
• The external-verification counters no longer count the same objects, since counting is now conditioned on production of the source and distributed across two independent axes.
• The absence qualifications in 3.E now pass through two successive gates whose order is constrained, and a new qualification appears, the non-attributable absence, which did not exist before.
• The activation threshold for Appendix 4 changes criterion: the evidentiary function assigned to a figure replaces its indispensability to the conclusion.
• ARGUS Light moves from eleven to twelve operations: the absence test in 3.E enters the short version, including its Stage 2 and its two gates. A Light analysis produced under V4.0.0 is therefore not comparable with a Light analysis produced under this version.
• The substitution test in 7.a changes method: it now concerns minimal correction of the flaw rather than removal of the faulty element. This is a particularly consequential change because 7.a determines the final severity qualification: the same flaw may now receive a different rank from the one it would have received under V4.0.0, especially an omission, which cannot be corrected by removal.
• The scale in 7.c moves from four to five levels and changes what it explicitly measures: the degree of interpretive closure. “Oriented information” is separated from advocacy, and the term “propaganda” is reserved for Levels 3 and 4. The same text may therefore receive under V5.0.0 a different level from the one the V4.0.0 scale would have produced.

A change in the overall procedure is added to these: as soon as relevance is not strong, a mandatory stop precedes the analysis and submits a scope for validation, whereas V4.0.0 allowed the analysis to proceed. For a corpus, this stop becomes a single validation placed before the first text. Strong relevance does not trigger a stop, but the scope decision must still be justified.

Step 0 and relevance

• New integrity check of the object at the start of Step 0: analysis of a manifestly truncated document is interrupted and flagged before the relevance test. A list of indicators is provided, together with the distinction between truncation and a deliberate excerpt, and three declared consequences if the user chooses to proceed.
• The relevance threshold in 0.a is recast. The five criteria no longer have equal weight: three are central, two corroborating, and strong relevance requires at least one central criterion. The previous threshold, any two criteria out of five, was satisfied by almost any document with a title and framing and therefore filtered nothing.
• The stop condition in 0.a is repaired, as 0.a bis made it possible to bypass it by attaching to partial relevance by composition a procedure that continued. The stop is now independent of the form of partial relevance and consists in submitting a proposed scope in three elements, not asking for permission. Three outcomes are recalled to the user, and an exemption is possible but declared at the head of the analysis. Seven executions under V4.0.0 had never triggered this clause. The exemption now distinguishes two cases: a scope already fixed by the user, treated as validated without a new request, and a waiver of validation without a user-defined scope, in which the scope remains set unilaterally by the analyst and is declared as such.
• What the stop is not: it concerns scope only and may not solicit either the severity of findings or what the user wishes to find.
• Clarification to the third point of 0.a bis: the prohibition on transfer between regimes works in only one direction. A section outside the scope may still be used to establish a flaw inside the scope when its data bear directly on a claim within that scope and alter its truth, reach or representativeness. That section nevertheless remains internal to the declared object: using it through this bridge constitutes neither external verification nor an external primary source within the meaning of 3.B and feeds none of the external-check counters.

External verification and sources

• Counting condition for an external verification in 3.B: a verification is counted only if the analysis produces the source actually consulted in a form allowing it to be found again, with its date. Knowledge remembered by the model, even when presented as verified, does not constitute external verification.
• Verifications are declared on two independent axes: their status — successful, partial, unsuccessful — and their mode of access — confrontation with a provided source, or external search. An external verification is external to the analysed object, not necessarily found online. The counting unit is defined by the verificatory proposition undertaken: the sub-elements needed to decide it are not split into separate checks; if some are established and others are not, the check is partial. The partition cannot be redone after the fact to transform a partial check into one successful and one unsuccessful check.
• New test in 3.B, fidelity to an available primary source: decisive claims in a secondary text are confronted with the source it comments on, with seven qualifications including “strengthened”, which identifies a claim made more certain than its source while remaining literally defensible. The severity of a fidelity finding is determined in 7.a by the substitution test, not by the qualification itself. Three of the seven qualifications are not flaws. A cross-reference is added to Rule 13 and a corresponding item to the Step 8 consistency check.

Absences, appendices and corpus

• Clarifications to Stage 2 of 3.E: accessibility of information is assessed at the publication date of the analysed object rather than at the date of analysis, with a distinction between a modality flaw and an omission flaw; qualifications pass through two successive gates whose order is not optional, and the criterion of the second cannot be used to refuse the first; treatment of an absence with an indeterminate result, which cannot be qualified as structural.
• Clarification to the Appendix 2 checkpoint: a characterisation without a criterion of application does not satisfy the definition test, which is therefore positive. The case in which the boundary test alone is positive had not been covered.
• Clarification to the activation conditions of Appendix 4: a figure is load-bearing whenever the text offers it in support of a central proposition, even if the conclusion could survive without it. The criterion is the evidentiary function assigned to the figure, not its indispensability.
• Clarification to Test G of Appendix 4: the test also applies to a count relayed by the text when the text presents it as established and draws its own conclusion from it, but not to a count explicitly attributed to a source while preserving its modality.
• New C.1 bis in Appendix 3: integrity check of each corpus document, then relevance triage before analysis of the first text, one line per text, proposed scope only for texts whose relevance is not strong, and a single validation covering both scopes and corpus composition. The user may restrict the analysis as a whole to only some texts. Any reduction is rewritten into C.1, otherwise the representativeness and second-order circularity tests would concern a corpus that was not analysed. The triage states how far each text is suited to ARGUS, not what it lacks.
• Clarification in C.3 of Appendix 3: distinction between anchoring risk caused by position in the analysis sequence and anchoring risk caused by the chronology of the objects when the order of treatment differs from publication dates.
• C.5 of Appendix 3 split into C.5a and C.5b, separately numbered. The clause required a structural separation between legitimate aggregation and interpretive contribution; a single section containing two registers did not satisfy it.
• ARGUS Light: the 3.E absence test enters the short version, which moves to twelve operations. The unaided pass in the preamble identifies Step 1, the absence test and epistemic symmetry as the three highest-yield questions, and the short version omitted one of them; the check on qualification of absences in Step 8 thereby also gains its object. The integrity check of the object and the stop in 0.a apply without simplification. Rule 13 also applies, including minimum external verification and the fidelity test; only the developed examination of 3.B is omitted. The short version abbreviates the analysis, not the decision that precedes it; the absolute rules continue to govern the retained operations without reintroducing explicitly excluded steps, apart from the evidentiary exception in Rule 13.
• The first gate in 3.E now has four outcomes when an absence is established — strategic omission, mere point of weakness, non-attributable absence, qualification not determinable as things stand — after which the second gate may produce structural omission. “Indeterminate result” is not a competing qualification but an additional property. The non-attributable absence, where information was not available at the publication date, is not a flaw and does not carry forward to 7.a. The “not determinable” outcome is bounded: a reasonably necessary check must have been attempted and produced a partial or unsuccessful result, or the unavailability of any check must be declared; the missing information must be named. Question 2 of Stage 2 is reversed so that both criteria of the first gate have the same polarity: two positive answers establish the omission.
• Appendices 2 and 4 reclassified as conditional appendices. They were presented as advanced options even though activation is mandatory once their threshold is crossed. Availability of the analyst’s time is removed from the activation conditions of Appendix 2: the cost of an analysis does not exempt a check made mandatory by the protocol.
• Severity under Test A of Appendix 4 is separated: an impossible load-bearing figure invalidates the quantitative proposition it purports to establish, while its severity for the central thesis is determined in 7.a through the substitution test. The word “fatal” is reserved for the 7.a scale. The previous automatic severity contradicted the new load-bearing threshold, which includes a figure the conclusion could survive without. Three remaining formulations of the old criterion, “figures on which the thesis depends”, are aligned with the new one.
• Appendix 3 activates before the first text rather than after individual analyses, as a consequence of C.1 and C.1 bis. An article accompanied by its primary source is not a corpus: the source is a control item within the meaning of 3.B.
• Blind rereading in C.3 is defined as such: a fresh conversation, without the other texts or analyses already produced. Returning to the first text after having read all the others was precisely the anchoring the test is intended to detect.
• Register C.5b opens with an explicit status declaration, “interpretive contribution external to the corpus”, while first-person wording remains an admissible marker. These markers are enunciative rather than modal: Rule 14 forbids attenuating a finding, not saying who carries it.
• Classification of a text as propaganda in 0.b is declared descriptive of the reading contract and not adjudicative: the degree is established in 7.c, which is not bound by any 0.b classification.
• Step 5: a gap between announced programme and content establishes a self-legitimating function and does not allow a conclusion about the author’s sincerity. The protocol does not impute intention where it observes only an effect.

Execution consistency

• The scope declaration in 0.a bis is carried out in every analysis, including one of strong relevance, where it takes the form of a single line. Only the stop and validation remain conditional. Under V4.0.0, 3.K and Step 7 referred back to a scope that a strongly relevant text had never declared.
• Explicit distinction between partial relevance and weak or nil relevance, based on the central criteria in 0.a: both triggered the same stop but called for different follow-up without a criterion separating them.
• Three verification numbers declared separately — successful, partial, unsuccessful checks — and “number of external verifications”, used on its own, refers only to successful checks. A single count had been required for objects that the counting condition separates, and the Step 8 consistency check necessarily flagged a discrepancy.
• Automatic activation of Appendix 2 relies on the boundary test alone, whether or not the text supplies a definition; an explicit user request remains the other activation mode. The case of a defined term whose scope nevertheless determines the conclusion had not been covered. A reading convention is added: a test is positive when it reveals the flaw, not when the text meets the requirement.
• The long version of Step 8 develops points (a) to (c), adds an explicit epistemic-symmetry check and retains points (d) and (e), mandatory in both regimes. The regime recommended for published analyses previously contained less mandatory content than the short version, and the 3.B cross-reference to “Step 8(d)” had no object there.
• Two indicators of incompleteness in Step 0 — an announced programme whose parts are missing and a standfirst announcing an absent demonstration — count as such only together with a material sign of interruption. On their own they created a presumption of truncation which, through the first of the three declared consequences, then neutralised the Step 5 finding from which they had been derived.
• ARGUS Light and Appendix 3 cannot be combined, because Appendix 3 presupposes a 7.c level that the short version excludes. The operation list in Appendix 1 is declared the executable scope of the short version. An item in the consistency check whose object was not produced is checked as not applicable without becoming a section to fill in; a single line covers points excluded from the short version.
• In a corpus, a text of weak or nil relevance is proposed for exclusion rather than excluded automatically: the outcomes are those of 0.a. The same text submitted alone could be analysed after validation.
• The word fatal is reserved for the 7.a scale. Test A may invalidate a quantitative proposition; Test G establishes the absence of demonstration of a non-reproducible count. Neither finding is carried into 7.a as such: its severity is determined there by the substitution test.
• The activation conditions of Appendix 4 are restricted to load-bearing elements, because previously the list and the non-activation clause could both be satisfied at once.
• The relevance criteria in 0.a are assessed over the whole submitted text: strong relevance presupposes a text that is argumentative throughout, while an object with identifiable argumentative parts falls under partial relevance by composition and therefore the stop. Practical consequence: most reports, dossiers and books now pass through a scope validation that V4.0.0 did not require.

Final rereading revisions

• Reach of the absolute rules in the short version: they govern operations within the scope and do not by themselves reintroduce an excluded step. Rule 13 is the sole exception. Without this doctrine, Rule 11 reinstated the register examination of 3.G and Rule 17 reinstated 0.f and 7.c, which the short version excludes.
• Test A of Appendix 4 bounded: a result exceeding the reference set establishes impossibility only if the counted unit is non-repeatable. For a set that renews, the bound is increased by adding to the initial size an upper bound on renewal established from the text or an admissible verification, never from the analyst’s own estimate. Without an established upper bound, the calculation is written out but impossibility is not answerable for lack of a bound. The unit, bound and time window are named before the conclusion.
• Test A: neither the cost of calculation nor uncertainty about the order of magnitude makes it unanswerable. It may, however, be unanswerable if a minimum datum is missing — reference set or bound, counted unit, time window or correspondence between the figure and its reference — and that datum cannot be established by an admissible verification. Absence of a reference falls under Test B only when it is genuinely a denominator issue. Listing these data is not an excuse not to search: a datum that a simple check can establish is not missing within the meaning of this clause.
• Test G of Appendix 4: a non-reproducible count does not establish the sub-thesis, without thereby establishing its falsity. If the count is produced by the text, it becomes an unproven and non-reproducible quantitative claim; if relayed, it retains only the status of a reported claim. The effect on the central thesis is determined in 7.a and is not automatic. V4.0.0 made it automatic. Its domain is bounded to documentary counts obtained by selecting or querying a corpus, media coverage or documentary database; it does not extend to ordinary statistical, administrative, demographic, scientific or operational measurements or series merely because they rely on a database.
• Indeterminacy of result in 3.E is a property added to the first-gate qualification rather than a competing qualification. Its only effect is to bar the structural qualification.
• Appendix 2 defines the strategic empty signifier by the function of vagueness rather than by the author’s intention, and Step 1 defines the straw man by the reconstruction performed rather than by a deliberate omission. Intention belongs to Step 4 and its three-level progression, which these two definitions had short-circuited.
• The 7.b successes section no longer has to be “as developed” as the failures section: it develops every solid element in proportion to their number and importance. Equal length forced the complacency that Step 7 otherwise prohibits.
• Level 0 in 7.c is renamed “non-propagandistic information”. “Neutral information” conflicted with the Fundamental Principle, which states that an opening device is never neutral.
• A text of weak or nil relevance analysed at the user’s express request carries a declared statement at the head. This is a derogation, not a fourth outcome of the stop, and applies to an isolated text as well as in a corpus, where C.1 bis had introduced it only for the latter.
• In C.5b, every new factual claim is sourced under the counting condition in 3.B. The register in which the analyst speaks in their own name is not one in which evidence is relaxed.
• The drafting example in 7.c and the scale itself comply with Rule 15: the two forms of Level 2 are on indented lines without secondary bullets. Because Step 7.c requires reproduction of the scale, leaving secondary bullets would have forced a non-compliant output.
• The specific rules of Appendix 4 are distinguished in the clause reserving “fatal” for 7.a: Test A may produce an invalidating finding for a quantitative proposition, Test G an absence of demonstration, and neither is carried directly into the scale.

Final executability revisions

• ARGUS Light now executes Tests A and G before the conclusion: a quantitative control capable of modifying 7.a cannot occur after the judgment it is supposed to inform.
• The Light scope clause no longer says that the absolute rules “do not depend on any scope”. Their reach is defined by the following paragraph: they govern retained operations without reintroducing an excluded step, Rule 13 remaining the explicit exception.
• Appendix 2 no longer has a third discretionary activation mode. It activates when the boundary test is positive or at the user’s request; outside those two cases, the analyst may propose activation but does not impose it.
• The 7.a substitution test is stated as minimal correction of the flaw: remove an unjustified claim, restore an omission, replace a distortion, or repair a logical leap. An omission cannot be “removed”; it is corrected by restoring the absent element.
• In 3.E, past accessibility may be not determinable. This outcome is not imputed to the text, does not carry forward to 7.a, and names the information that would be needed to decide. In a corpus, another text counts as accessible information only if it had already been published at the date of the object examined.
• In a corpus, C.1 bis constitutes execution of the integrity check and 0.a relevance qualification for each text. Individual analyses restate these results and the scope, then continue at 0.b without recalculating the triage or replaying the stop.
• The global level in 7.c applies to the ARGUS scope taken as a whole, not to sections excluded in 0.a bis.
• Two remaining unestablished frequency formulations in Appendix 2 are neutralised in accordance with Rule 19: the opening “can” install an empty signifier and that use “may constitute” an indication of intent, without asserting frequency.
• The long version of Step 8 develops (a) to (c), adds epistemic symmetry and retains (d) and (e); the developed formulations replace the corresponding short answers and do not create three additional responses.
• The threshold of around ten checks in 3.B becomes a signal of change of scale, not an exemption from Rule 13. Simple checks concerning decisive elements remain mandatory; additional checks fall under separate fact-checking work.

Final convergence and execution tests

• The 3.E outcome “first-gate qualification not determinable as things stand” is reserved for an investigated indeterminacy: the reasonably necessary check produced a partial or unsuccessful result, or the analysis justifies that no check was accessible. An uninvestigated uncertainty cannot be delivered under that label.
• The substitution test in 7.a distinguishes the quantitative case in which a correct value is established — it then replaces the faulty value — from the case where no correct value is available, which leads to removal of the quantitative proposition rather than invention of a plausible value.
• In a corpus, C.2 restates in each individual analysis any incompleteness limitation declared during the pre-check and its three consequences; the integrity check and relevance triage from C.1 bis are not replayed.
• The Step 8 consistency check has been aligned with the three counters in 3.B, fidelity to a primary source, and the boundaries of absence qualification. In the short version, points without an object are checked without recreating a reporting form.
• Synthetic execution scenarios were finally used before publication. They exposed four interfaces that are now explicit: the atomic unit of an external check, separation of the two exemption regimes in 0.a, the status internal to the object of the bridge from an out-of-scope section, and the domain boundary of Test G. These clarifications are intended to improve execution reproducibility and add no new class of flaw.

Contradiction and drafting

• Added to the Level 2 differential grid: coverage divergences, meaning findings raised by only one of the two analyses. A finding not mentioned is neither conceded nor contested, and the adjudication rule does not apply until the silent analysis has stated whether it rejects it after examination or did not examine it.
• Wording corrections applied to the protocol itself under its own rules. One inaccurate incompleteness indicator was removed. Five unestablished empirical claims about classes of texts, authors or traditions were removed under Rule 19, in 3.F, in two tests in Appendix 4, and in the Lineage and Distinctive Contribution section: the methodological content depended on none of them. Two enumerations announcing a fixed number of elements were made independent of that number. Residual errors corrected: a reference to a Step 3.L that no longer exists, a sentence in 7.c claiming that a global level averages heterogeneous regimes before specifying that it is not an average, an imputation of bad faith replaced by a description of its effect, and a sentence in Step 8 whose inverted syntax reversed its meaning.

Revisions from the first real integration test — RC10

• Fidelity test against a primary source: a supplied, official, contemporary or directly relevant primary item no longer activates the test by itself. The text must be linkable to that source through citation, naming, summary, commentary, criticism, or established lineage of the checked claim. Mere concordance of date, wording or figure is not enough; without lineage, the item remains an ordinary external verification.
• The three counters in 3.B are qualified as an internal execution trace, not an inter-analysis metric. Two independent analyses may start from different checking propositions; what must be reproducible is the partition announced and maintained within each execution.
• 3.E bounds accessibility to the absent information itself. The mere availability of an interlocutor, their physical presence, or the possibility of asking them a question does not establish that a particular piece of information was easily accessible at the publication date.
• Test B of Appendix 4 distinguishes the denominator from its absolute value. A share of “9% of global sales” has a stated denominator even without the global volume; a missing volume may affect reconstructibility without becoming a denominator flaw. For a change or reduction, the comparison base or period must be identifiable.
• 7.c adopts a five-level scale: 0 open information / non-propagandistic; 1 oriented information; 2 advocacy / engaged journalism; 3 propaganda; 4 locked propaganda. The four threshold transitions are made explicit: structuring orientation, active defence of a conclusion, protection against contradiction, then self-immunising closure.
• The forms “frontal” and “sophisticated” are removed from sub-level status and become descriptors of rhetorical mode. Origin or function — journalistic, activist, corporate, institutional, state, para-state — becomes a second orthogonal descriptor. Neither form nor origin automatically determines the level.
• Corresponding adjustments are applied to the “propaganda” genre in 0.b, the Step 8 consistency check, Rule 13, ARGUS Light, and the Lineage and Distinctive Contribution section.

Revisions from the second real integration test — RC11

• C.1 bis now explicitly separates corpus composition from the ARGUS scope internal to each text. Fixing the list or order of texts does not automatically validate an internal division; for a strongly relevant text, declaring the scope equal to the whole text is the normal execution of 0.a bis and does not trigger the formula for a scope set unilaterally. The two exemption formulas in 0.a are declared mutually exclusive for a single decision.
• 3.E fixes the case of texts with the same date or in the same edition: another corpus text proves accessibility of information only if its effective prior public availability is established, or if a distinct earlier source had already made the information accessible. Mere identity of date, publication in the same edition or presumed availability within the newsroom is insufficient to ground a strategic omission.
• The proportionality signal in 3.B beyond around ten checks is assessed per individual object. A corpus total may be reported as an inventory, but its accumulation does not by itself trigger the signal, and cross-text consistency checks in Appendix 3 are not artificially added to those of an individual text.

V4.0.0 — July 2026

• New Appendix 4, arithmetical and statistical control: seven tests applied only to the figures carrying the thesis, including the reproducibility-of-counts test. Introduces a class of flaw that previous versions could not detect, the impossible figure correctly attributed.
• New preamble section, Function of the Protocol: aid and learning tool, with the practice of the unaided pass.
• New preamble section, Reliability of the Analysis and Contradiction: three levels of increasing cost, short counter-check, blind cross-audit with differential, documented human arbitration. Adjudication rule to contain an AI’s tendency to defend its own analysis.
• Reorganisation of the preamble: “Function of the Protocol” is placed first, before the Fundamental Principle, because the unaided pass must be carried out before submitting the text. The “Reliability of the analysis” paragraph inserted into the Fundamental Principle in V2.1.3 is moved and merged with “Reliability of the Analysis and Contradiction”. It remains present in the user guide and FAQ. An explicit note now distinguishes sections addressing the user from those addressing the AI executing the protocol.
• New Step 0.a bis, composite texts: distinction between partial relevance by degree and by composition, declaration of scope, prohibition on transferring judgment between regimes.
• Performative contract added to 0.c: the standard of rigour a text prescribes enters its reading contract.
• Effective insertion of the “When elements are insufficient” clause at the head of Step 7, announced in the V3.0.1 entry and absent from the text then published. Completed by its application to the three sub-steps and by three boundaries, including the distinction between insufficiency of the text and indecision of the analyst.
• Added to 3.B: the test for recontextualised quotations.
• Added to 3.F: the test for unfalsifiability in use, distinct from formal immunisation.
• Added to 3.H: symmetry of the interest-based reading.
• Added to 3.K: the test for distribution of rigour and the term “displaced rigour”.
• Added to 7.c: sub-scope levels for long and composite texts, with two safeguards.
• Added to Step 8: points (d) deference and (e) series conformism.
• Added to 0.a bis: the obligation to justify the scope decision in both directions. Treating a long text as a single argumentative whole must be justified on the same footing as excluding a section, and the statement “no section excluded” is admissible only with a reason.
• Preliminary declaration of the exact object analysed, in one line at the head of the analysis, with the rule that every external verification bears on this object. Added after an analysis confused a web publication with its revised republication.
• New Rule 19: the analyst applies Tests 3.C and 3.G to their own prose. Prohibition of comparative praise or blame based on an unestablished claim about a class of texts, and prohibition of charged terms the finding does not require. The corresponding point is added to the Step 8 consistency check. The protocol examined the register of the analysed text without examining its own.
• Added at the head of Step 8: a mandatory consistency check of the analysis itself, in five points: agreement between header and produced appendices, agreement between declared and conducted external verifications, non-contradiction between Step 7 and Steps 3 to 6, a recomputed value required for every calculation declared exact, and a single consistent description of the same fact. ARGUS subjected the analysed text to an internal consistency test without subjecting itself to one.
• New section, Lineage and Distinctive Contribution.
• ARGUS Light raised to 11 steps: Tests A and G of Appendix 4 are included.
• External verification: Rule 13 and the scope of 3.B move from an exceptional regime to a minimum obligation. A decisive element that a simple check could establish must be checked. Three conditions added: symmetry of verification, reporting of unsuccessful searches, and mandatory declaration of the number of verifications conducted, including when it is zero.
• Objective activation condition for Appendix 2, as a checkpoint placed at the end of Step 2: a decisive term of the central thesis with no definition, criterion or enumeration in the text, whose boundary changes the conclusion. The decision not to activate must be justified.
• Clarification in 3.K: in a composite text, the interpretive-closure-by-complexity test bears on the argumentative scope declared in 0.a bis, not on technical sections.
• Addition of the present versioning policy.

V3.0.1 — July 2026

• Added a “When elements are insufficient” clause to Step 7.
• New Rule 18: prohibition against producing an artificially complete analysis.

V3.0.0 — July 2026

• Merged Steps 3.E and 3.K into a single test of absences (identification + strategic qualification).
• Refocused Step 3.K (formerly 3.L) on target-audience adaptation, without repeating absences.
• Reversibility clause between 0.f and 3.K: mandatory return to the audience hypothesis at the end of Step 3.
• Mental substitution test for the severity scale (7.a).
• Step 8 made mandatory (minimum short version required).
• Principle of proportionality added after the Fundamental Principle.
• Note against double penalisation between 3.I and 3.E.
• New Appendix 3 — ARGUS-Corpus for the coordinated analysis of multiple related texts.
• ARGUS Light updated: Appendices 2 and 3 excluded, Step 8 mandatory in the short version.

V2.3.1 — July 2026

• Added a cautionary sentence in Step 7.c: the term “propaganda” should not be assigned solely because of commitment or asymmetry.
• Revised 7.c scale: Level 2 renamed “Advocacy or camp propaganda”, with clarified sub-levels.
• Enriched Step 0.b with textual regimes: advocacy / manifesto, camp-based activist text, propaganda.

V2.3.0 — June 2026

• Added the Principle of Rhetorical Destination.
• Added Step 0.f (target audience, actual audience and credibility contract).
• Added Step 3.L (target-audience adaptation test).
• Enriched Step 7.c (target audience consideration, sub-levels, effective/decorative counter-argument).
• Added Rules 16 and 17.

V2.2.0 — June 2026

• Added source audit (0.d): analysis of the diversity, reliability and possible bias of cited sources.
• Added production context (0.e): identification of the media outlet, its economic model and its links to stakeholders.
• Added informational circularity test (3.I): verification that sources do not all come from the same ideological ecosystem.
• Added comparative treatment analysis (3.J): systematic comparison of vocabulary and treatment of different actors.
• Added strategic omissions test (3.K): identification of accessible but absent information that would harm the thesis.
• Added qualification of the degree of propaganda (7.c): four-level scale, from neutral information to state propaganda.
• ARGUS Light: explicit statement that advanced steps (0.d, 0.e, 3.I, 3.J, 3.K, 7.c) are not included in the short version.

V2.1.4 — June 2026

• Locked the normative link between 0.c and 3.B: it is forbidden to relax the standard of proof during the analysis.
• Refined the straw-man test (Step 1.2) using the principle of charity.
• Added a safeguard clause to Rule 14: the ban on loose modifiers must not penalise an author who expresses legitimate nuance.
• Strict definition of structural omission (Step 3.E): the missing element must contradict the conclusion or require a major overhaul of the causal mechanism.

V2.1.3 — June 2026

• Removed the prompt-based verification mechanism in favour of a warning about AI reliability.
• Warning inserted at the end of the analysis, in the Fundamental Principle, the user guide and the FAQ.
• Removed the reference to formatting from the warning.

V2.1.2 — June 2026

• Corrected the verification mechanism: the AI no longer self-assesses; it invites the user to trigger a verification.
• Numbered lists authorised in Rule 15.

V2.1.1 — June 2026

• Added a recommendation for a conformity check (replaced in V2.1.2).

V2.1.0 — June 2026

• Added Rule 15 (formatting). Added Appendix 2 (empty signifiers).

V2.0.0 — June 2026

• Added textual genre, standard of proof, severity scale, mandatory qualities section, Step 8 self-critique, ARGUS Light.

V1.1.4 — June 2026

• Added Rule 14 (against analytical euphemisation).

V1.1.3 — June 2026

• Added Step 3.H (epistemic symmetry).

V1.1.2 — June 2026

• Added Step 0 (relevance test).

V1.1.1 — June 2026

• Added Rule 13 (methodological scope).

V1.1 — June 2026

• Restructured strategic-intent inference.

V1.0 — June 2026

• Initial version in 6 steps and 12 rules.