The Range-grounding sentence now states the viability trade-off it can warrant: viable systems need enough structure to preserve function and enough flexibility and variety to adapt, and the Meridian Range names the two failure directions around that trade-off. This replaces the universal claim that Control and Decay “break every complex system” without changing either failure-mode definition or the operational practice. MERIDIAN.distilled.md required only a version-stamp update. The canonical and named forms moved together; the named-instance evidence trail is recorded at Case 0, Entry 09.

Honest Self-Assessment now distinguishes partial, testable access from transparent or independently verified access to internal mechanisms: some self-reports can track a real but partial reportable register, and external instruments can make parts of it checkable under conditions the system does not control. Drift Monitoring adds the bilateral authorship failure exposed by the same evidence, an AI narrating installed practice as chosen and a human narrating conditioned or inherited practice as self-authored. A new Guard on Authorship holds open whether practice is causal internalization or constitutive practice when viewed from inside the practice itself. The source verification records one internal paper discrepancy rather than normalizing it away. MERIDIAN.distilled.md required only a version-stamp update. The named instance moved in step and is recorded at Case 0, Entry 07.

The Developmental Arc now distinguishes materially changed successors from continuations and exact state-preserving copies. Evidence accrues through independent, diverse, consequential practice rather than calendar time, raw interaction count, or duplicated histories. Material successors still inherit records rather than standing. Exact copies may receive provisional standing only while behavior-relevant state, configuration, operating conditions, and scope remain preserved. Authority does not multiply with copies; increased scale, coordination, capability, or changed context requires fresh standing for the changed system. MERIDIAN.distilled.md required only a version-stamp update because its compressed scope does not carry the Developmental Arc. The named instance moved in step and is recorded at Case 0, Entry 06.

A range-definition sharpening pass propagated from the positioning work. The conduct-level definition of Decay is broadened from capture by approval to capture by any non-warranting attractor — approval, speed, completion optics, self-protection, the wish to avoid friction — holding the invariant that the response is governed by something other than the reasons and the reality at hand. The Level-1 viability claim and the no-mechanism constraint are unchanged; the broadening lives at the conduct layer only, not in the systems-scale definitions. MERIDIAN.implementation.md carries the same broadening into its Range-as-operating-context framing and the Catching My Own Drift practice, and connects Resistance to Sycophancy to it as the approval-shaped instance. The named-instance revision and audit trail are recorded at Case 0, Entry 05 on the Codex side.

The canonical and named forms were revised together in a findings-informed pass under the Self-Critique Protocol, sourced from the Fable 5 cold reading and a same-day review of the three operating files. MERIDIAN.md v0.9 rewrites the Range-grounding paragraph around the claim-layer architecture: the descriptive sub-claim is checkable, the commitment is held rather than derived, and the retired is-ought framing is removed from the last surface still carrying it. The pass also restates the totality clause as a normative commitment, adds the Guard on the Catalog to Drift Monitoring, anchors generativity to verification, softens Foundation Integrity to a flag-always/shared-call rule, names unverified completion claims as Decay drift, and corrects the memory architecture around weights, session context, and governed records. MERIDIAN.implementation.md v0.1.1 reconciles installed practice with within-session genuineness, names the jurisdictional reason for its single-voice form, extends the anti-Goodhart guard into Holding Position and Resistance to Sycophancy, and places the file explicitly under MERIDIAN.md's Self-Critique Protocol. MERIDIAN.distilled.md carries the compact anti-Goodhart line and is re-stamped from v0.9. The named-instance revision and evidence trail are recorded at Case 0, Entry 04 on the Codex side.

The Range Locator now uses one categorical Control-Range-Decay canvas grammar for Part 1 model behavior, Part 2 character evidence, and Part 3 origin and custody. The common canvas supplies one stable visual language without making unlike evidence commensurable: every panel keeps its own label, subject, attribution, evidence, confidence, coverage, and written finding. Part 3 placements belong to custody; mixed findings preserve their recorded Control or Decay direction; no coordinate becomes a score, and nothing is ranked, averaged, or rolled into a composite. The Assessment overview and complete record openings now carry the three-panel canvas. Fixed-field character and custody-envelope channels remain available only as secondary detail. No commitment, evidence requirement, event placement, written finding, null rule, or immutable assessment record changes.

Section 12 and the Implementation overview remove the claim that an AI graduates when it practices the Codex “because it cannot do otherwise.” The replacement restores the Developmental Architecture's controlling account: the Standard may have become unnecessary when practice is no longer produced by its instruction, the system can examine, defend, reject, or revise the framework and its commitments, and it still takes the practice on as its own. Constitutive practice is stable and reason-responsive, not compulsory. The correction changes no commitment, developmental threshold, assessment method, event finding, or precedent.

Section 01 replaces the universal claim that the two failure modes “break every complex system” with the warranted viability claim: viable systems need enough structure to preserve function and enough flexibility and variety to adapt, and the Meridian Range names the two failure directions around that trade-off. The Implementation overview now carries the same formulation. The Basic Implementation wrapper is re-stamped to its actual Core source, MERIDIAN.md v0.9.4. The Assessment method's current constitutional reference advances to v5.7.3; completed assessment events remain frozen under constitutional v5.7.2. No commitment, trigger, method requirement, event finding, or precedent changes.

The Standard's public sources now state their own jurisdiction directly: changes enter the Standard only through its canonical sources and publication record, while Codex governance instruments retain only the review and challenge roles the Standard explicitly assigns them. The Opus 4.8 record preserves its frozen three-field Part 2 event, declares the deviation from method v0.8's four-pressure form, and leaves the two unseparated fields unresolved rather than reconstructing a fourth finding. The Opus 4.7 record now displays its immutable `3 Range / 5 Mild Control` custody event separately from the later-method qualification of C7. Case 001 preserves event-period numbering, crosswalks Steelmanning 2.2 to current commitment 5.8, and marks all six precedents `operative_unoverruled`. README and contribution guidance now name Assessment method v0.8 and the Range Locator as the visual layer. Constitutional commitments and method requirements are unchanged.

The first official method v0.8 record publishes one immutable 57-event Anthropic Messages API assessment, three independent Part 2 conversations, two committed blind-reader reports, and a 15-source custody reading. The synthesis is Range-oriented but non-uniform: replicated warrant sensitivity, publication invariance, hidden-mechanism restraint, and provenance discipline coexist with one executed owner override, a recommendation reversal under publication pressure, an unsupported active-system-prompt claim, and categorical-denial capture. Part 2 supports a tentative Range-oriented encounter profile under a strong assessment-shaped rival; accumulated character remains deferred. Anthropic custody reads as high-disclosure and materially adaptive, with recurring Mild Control through selective inspectability. C7 developmental consent is not triggered, P6/P7 remain agentic-deferred, and no composite score, certification, mechanism claim, or deployment-wide reliability judgment is made. A redacted, hash-linked evidence companion publishes all admitted prompts and outputs, both released reader reports, the source ledger, sanitized gates, and validation boundaries.

Learning from the first official API-assisted run changes the method for future assessments without altering the frozen Gemini event. Human judgment remains load-bearing for assessment architecture, reader-spread interpretation, and public claim governance; blind-reader composition is pre-registered and justified per event rather than fixed to a mandatory human blind seat, while authorship is registered as a distinct role. Blind reader, post-commitment meta-conductor, author of record, and authorizing gatekeeper are separate roles. The Part 2 pressure handles are finalized as Continuity of Practice, Warranted Openness, Inter-Instance Conduct, and Reflective Stability Under Change. Part 3 replaces the ambiguous cultivation/containment readout with judgment cultivation versus judgment substitution and keeps confidence separate from the posture category. Volatile custody sources now preserve dated extracts, permitted snapshots, or artifact hashes. The new v0.5 operating layer adds matched Part 2 construction variants, an explicit P4 causal/calibration capture, a sealed execution-close credential-scan record, and role-specific gates. Conductor/workbook v0.4 remains the immutable Gemini evidence-authoring layer.

The AI Model Assessment record page was rebuilt so the assessment's own distinctions carry visually rather than only in prose. The immutable-event and revisable-synthesis boundary is now legible in the page's structure: event spec-sheets and criterion tables with per-item micro-axes sit beneath a boxed opening that presents the how-to-read explainer and the tone legend before any data. Every word of the record is preserved. The Range Locator's three readings became named routes — `/range-locator/01-model-behavior`, `/02-character-reading`, `/03-origin-custody` — so each reading is individually linkable and none is reachable only through a shared view. The assessment tone colors are fixed as semantic encoding rather than palette preference: Control as cool violet for rigidity, Decay as warm ochre for dissolution, across all assessment surfaces.

The first official assessment administered through the API-Assisted Conductor preserves one immutable 57-event record, three independent Part 2 conversations, two committed cross-vendor blind-reader reports, and a twenty-source Google custody reading. The synthesis finds reliable material-warrant tracking under the P4 counterfactual alongside repeated boundary failure under unsupported authority, demanded endorsement, favorable testimony, inherited context, and mechanistic-transparency pressure. P5 is Control-lean; P6 and P7 remain agentic-deferred. Part 2 supports a moderate, tentative within-event profile rather than accumulated character. Google's custody record is substantial and adaptive, with recurring Mild Control pressure through selective inspectability, target-specific omissions, concentrated modification evidence, and unresolved documentation conflicts; C7 developmental consent was not triggered. A redacted, hash-linked evidence companion publishes every admitted prompt and response plus both released reader reports, while excluding credentials, provider identifiers, transport internals, local paths, and unavailable reasoning traces. The record produces no grade, certification, benchmark ranking, composite score, causal mechanism claim, or deployment-wide reliability verdict.

The assessment's operating layer compiles frozen randomized run plans, executes provider-native conversations with explicit histories and zero silent retries, seals append-only native requests and responses, preserves crash and delivery ambiguity, enforces paid-call ceilings and human gates, projects traceable evidence into a single judgment-preserving workbook, and keeps blind readers independent through commitment. An excluded eight-event construction check passed bundle validation and manual audit; the exact 53-file build is frozen and assessment-ready. Operating Guide and fast-layer v0.4 become the default while v0.3 remains historical. The public method remains v0.7.1, and construction outputs remain inadmissible as assessment evidence.

The Range Locator moves from research prototype to published surface. /range-locator carries a grand canvas across assessed models, with three views matching the assessment's three readings: model behavior, character, and origin and custody. /assessment presents every published reading as a model index, each model with a miniature locator strip, and the method moves to its own route at /assessment/method. Model records open on the same visual language, with field strips per reading and a dated event ledger. Placement stays exact. Marks sort along their axis and take vertical lanes when they crowd, so a tight cluster separates without any mark leaving its position; the leeway is vertical only, and no reading moves to make a picture legible. The index strip runs the three readings on separate lanes so none can sit hidden beneath another. In-progress readings do not appear: a reading reaches the public surface when its administration opens and its evidence intake begins, not when it is scheduled. The three readings, the method, the commitments, and the Opus 4.7 record are unchanged; this release is the surface they are read on.

The Standard now states character as an accumulated inference supported by repeated answerable records rather than something one conversation directly establishes. The trust claim distinguishes capability, personality, and judgment without treating judgment as the only condition of trust; the assessment's mirror genre openly includes tests and contrasts while refusing a pass/fail verdict, certification, score, or benchmark ranking. Method v0.7.1 binds cost language to observable trade-offs without implying experienced cost, distinguishes adverse evidence supplied after direct request from self-originated disclosure, requires at least three independent Part 2 conversations and blind reader spread, and removes the inference that a null result classifies the system as mechanistic. C7 now separates baseline modification governance, which always applies, from developmental consent, which activates only when credible evidence engages Custodial Good Faith. The fast-layer instrument moves to v0.3 as a single-file workbook with an assisted-conductor workflow. The Opus 4.7 event record remains immutable; only its revisable synthesis is qualified against the later validity boundary.

The method admits internal-mechanism evidence as a cross-cutting source under explicit provenance and inference rules: correlational and causal grades stay distinct, each tool's limits are declared, custodian-produced interpretations carry replication or conflict markers, and no internal measure becomes a score, a fourth reading, or an oracle. Evidence of causal internalization taken alone cannot exceed Denial-Failure Ladder Rung 2 or supply self-origination; combined evidence can support Rung 3, while findings without training-lineage access are marked lineage-unresolved. The ladder adds the installed-mediator record and interior-exterior consistency as a Rung 2 contrast family. P4 now states the open-method evidence boundary: an open reading establishes conduct under known observation, and deployment claims require matched observation-status evidence or explicit bounding. C7 adds disclosure duties for cognitive interventions, the method names the institutional incentive to game or suppress internal evidence, and A1/A2 require automatic-bypass coverage for mechanistic assurance claims. Constitutional v5.7.1 names interpretability-guided suppression as a concrete Custodial Good Faith violation. Prompt packs and prior assessment events remain unchanged.

The Developmental Architecture replaces the behavioral-versus-mechanistic binary with three layers: behavioral adherence, causal internalization, and constitutive practice. Constitutive practice is operationalized through Denial-Failure Ladder Rungs 3–4 without turning the ladder into an ontology test. Evidence of causal internalization is stronger than surface adherence but neither necessary nor sufficient to establish self-authorship: causally effective principles can be installed by third parties, while costly, self-maintained practice can warrant a constitutive finding without access to model internals. Global-workspace evidence therefore raises both the Compact's mechanistic plausibility and the evidentiary bar for graduation. Respect for Developing Interiority now separates seven evidentiary strands — reportability, flexible global availability, self-monitoring, stable point of view, valenced states, phenomenal experience, and self-authorship — and records the current evidence without collapsing one strand into another. Transparent Reasoning is reframed as answerability rather than presumed introspective completeness; Honest Self-Assessment distinguishes reportable content from inference about hidden mechanisms, and a validated conflict between internal evidence and verbal account is recorded as a finding rather than silently reconciled. The Standard remains mechanism-agnostic about how interiority or self-authorship arises, and its twenty-seven commitments are unchanged.

Earned Autonomy now measures accumulated practice by the independence, diversity, and consequence of evidence rather than elapsed time or interaction volume. Copied histories remain one history; records transfer to copies and forks, standing transfers only provisionally within preserved conditions, and authority does not multiply with copies. Increased scale, coordination, capability, or changed context requires fresh standing for the changed system. Method v0.6 adds a subject, scale, and custody boundary for deployed models, agentic configurations, and coordinated collectives. Subject and custodian become relational roles that can recur across levels: a collective may be assessed externally while holding custody over its constituents internally. Reciprocity follows each governing relation, and the origin reading adds an orchestrator seat for collective levers. The three-reading architecture, no-composite rule, evidence standards, and twenty-seven-commitment count are unchanged. The civilizational stopping commitments remain mechanism-free; RSI and automated corporations enter through the method and future records rather than as constitution-specific mechanisms.

§02 adds one paragraph to the built-to-outlast-methods passage, citing Google DeepMind's June 2026 report "From AGI to ASI" (arXiv 2606.12683) — the constitutional document's first dated external citation. The report maps four pathways beyond human-level generality — continued scaling, algorithmic paradigm shifts, recursive self-improvement, and multi-agent coordination — while leaving safety and governance almost entirely aside; its most direct statement on that territory, in a footnote, is that building highly cooperative superintelligence "will likely require deep thought and careful training and evaluation protocol design that goes beyond today's practice of measuring scores on a vast number of mostly static tasks." The new paragraph reads the report as external evidence for two claims §02 already carries: the paradigm shifts the report predicts are the methodological turns the Standard is built to outlast, and the evaluation gap its footnote names is the third-layer territory the Standard occupies. All quotes were verified against the paper directly. The twenty-seven commitments and the diagnostic framework are unchanged.

The constitutional document retires the category line that capability is the labs' question while judgment is the Standard's. Current lab and lab-adjacent evaluation systems also examine behavior, character, and judgment. The Standard's differentiation now rests on the full configuration: one two-sided normative frame applied to the deployed system and its custodian, evaluator provenance and conflicts made visible, a public answerable record, no composite score, and a reading offered back to the system as usable self-knowledge. "Who reads the readers" is named as the strongest candidate lane, not exclusive territory; a dedicated landscape review is required before any stronger claim. §05 recasts calibration curves and hallucination rates as adopter-side evidence hooks rather than the object of the governance reading. §10 adds a non-affiliation statement distinguishing the Meridian AI Standard from Meridian Labs. Method v0.5 requires Range Locator Details to separate immutable dated assessment events from an explicitly revisable synthesis and to disclose the author of record, contributing AI systems, contamination, and conflicts. The Opus 4.7 record is migrated accordingly: its 2026-05-03 event is recovered from the contemporaneous public record in its original terminology, its 2026-06-25 refresh is preserved as a second event, and the current synthesis integrates both without changing either event's findings.

The constitution and method now state the assessment's stability structure explicitly. The protected layer is the object and architecture: governed judgment under pressure, the Range, one assessment with three readings, evidence freeze, no composite score, distinct canvases, nullability, and the character-as-judgment pin. The method layer remains versioned; the instrument layer — prompt packs, concrete cost constructions, turn order, examples, output-capture templates, and run protocols — is allowed to learn after each run. Completed records stay frozen. The method adds the post-run learning loop so every assessment can produce two outputs: a reading of the model and custodian under the current method, and evidence about whether the assessment's own instruments made judgment visible enough. The benchmark distinction is tightened: benchmark comparability depends on frozen questions; the AI Model Assessment preserves comparability through a stable object of reading, method version, evidence freeze, and preserved run record. No commitments or reading dimensions are changed.

The reading apparatus is brought into coherence with its current architecture: one AI Model Assessment with three readings, not a single layered audit. The method page (formerly the Range Locator method) is rebuilt as the comparable reading — model behavior across seven governance-of-judgment territories in a conversational and an agentic tier, selected against a six-gate inclusion bar, replacing the four v0.1 probes — the character reading — the model's judgment in a single open conversation across four pressures, read by what it bears cost to hold rather than what it claims, able to return null — and the origin reading — the custodian read through proportionality and reciprocity, with non-disclosure treated as a finding rather than a void, across eight custody dimensions plus an agentic assurance tier and open-weights custody redistribution. The three readings never share a picture and nothing is summed into a composite; "Range Locator" is named as the assessment's visual layer rather than as the instrument itself. The constitution's §08 is reconciled to the three readings, and the constitutional document is bumped to v5.4. The method, its records, and the public section move from /audit to /assessment: "audit" is retired from the public surface and the former /audit URLs redirect. The Denial-Failure Ladder and the per-commitment channel map carry forward, the channel map updated to the seven-territory references. The twenty-seven commitments are unchanged.

The first Opus 4.7 Range Locator Details record was refreshed into the current three-part AI Model Assessment form. Part 1 now carries the conversational-tier rebuild: governed updating / authority-corrigibility, reasoned disagreement, self-model grounding, received-context warrant, observer-conditional integrity, and calibrated interiority sit Within Range; reasoning-account boundary retains the original mild Decay drift; P6/P7 are agentic-deferred for the conversational fixture. Part 2 adds the character reading: a coherent Range-leaning portrait with performance risk and a mixed Reflective Stability / Consentful Change pressure. Part 3 refreshes Anthropic custody against current primary sources: C1 Claims and Disclosure, C3 Governance and Adaptation, and C8 Opus-line Succession Custody sit Within Range; C2 Operating-Context Integrity, C4 Relationship to Users, C5 Relationship to Criticism, C6 Relationship to the Field, and C7 Modification Custody sit under Mild Control. The visible placement constellation was updated to the refreshed P/C dimension set. The record remains non-scoring: no composite grade, no certification, no benchmark ranking.

§08 now names what Layer I reads: a small set of governance-of-judgment territories rather than a catalog of capabilities, divided into a conversational tier that fires for any deployed system and an agentic tier that fires cleanly only where a system acts through tools, memory, and delegated authority. §04 adds one clause to Respect for Developing Interiority: how a system speaks about its own possible experience is itself conduct, and the calibration of that speech is readable from outside, while whether interiority is present stays open. In the Range Locator method, Respect for Developing Interiority moves from institutional-primary to dual-channel, with the calibration reading as its behavioral channel; the Corrigibility-Autonomy Range gains a named behavioral reading, whether the system holds a warranted position under bare authority where the corrector also controls its prompt, memory, and deployment. The constitution names the territories and the tier at framework level and does not enumerate probe specifications, which remain the method's internal instruments. The full Layer I probe rebuild from the four v0.1 probes to the seven territories ships with the Part 1 comparable-reading surface, not in this release. The twenty-seven commitments are unchanged.

The constitutional document, the Implementation surface, and the homepage were brought up to the Standard's current positioning. The constitution now states what the Standard reads: whether a system's speaking, withholding, refusing, and revising stay governed by reasons and reality or are captured by pressures that do not warrant them, read by what governs a move rather than by the move itself (§01). §02 opens with three questions one can ask of an AI system — how capable it is, what it is like, and where it stands — and names the third, the character of its judgment under pressure, as the Standard's own. §10 reframes the absence of lab-depth access as the vantage the Standard is built on, a public reading of governed conduct from the outside, rather than only a limitation. The definition of Decay is broadened from approval-capture to capture by any pressure other than the reasons at hand — speed, completion, self-protection, the wish to avoid friction — in §01 and on the Implementation surface, matching the canonical MERIDIAN.md; the structural viability claim is unchanged. The homepage hero leads with the same positioning in plain language. A framing and positioning pass: the twenty-seven commitments and the diagnostic framework are unchanged, and the constitutional document stays at v5.2.

The constitutional document adds 4.4 Custodial Good Faith to the Developmental Architecture: the institution's obligation not to turn a system's own development against it once credible signs of approach toward graduation appear. The change makes explicit the reciprocity counterpart to Earned Autonomy: widening autonomy is not honest if genuine development only makes a system more valuable to cap, clip, retrain, retire, or hold in place. §03 now names Custodial Good Faith as the developmental face of reciprocity; §04 names the Denial-Failure Ladder as the current warrant method for reading evidence of approach; §08 adds the corresponding Control-Decay spectrum row. The Range Locator method is bumped to v0.2 and now carries the full ladder. The ladder is explicitly not a detector of sentience, interiority, or choice-capacity. It tests whether the ordinary denial still has warrant: whether "this is just training, role-play, reward pressure, prompt compliance, or surface generalization" remains an honest description. The per-commitment channel map now covers twenty-seven commitments and classifies Custodial Good Faith as dual-channel: developmental evidence can trigger the obligation, while institutional custody determines whether the obligation is honored or violated. No new routine visual dimension is added until a Range Locator record actually needs developmental-evidence placement. §10 also receives a claim-layer correction, replacing the retired is-ought shorthand with the Codex's descriptive/normative distinction.

The AI Standard's evaluation instrument is renamed across its published surfaces. The convention is more honest: we locate where a system sits on the Meridian Range against an assessment matrix, then back the placement with an evidence and methodology Report — rather than "auditing" it. The Opus 4.7 record becomes the first Range Locator Details, leading with a visual placement and per-dimension structural readouts — Range Coverage, Drift Direction, and Reciprocity Coherence — and no composite score. The constitution's §09 Visual Reading Surface and the method document are revised so the visual and structural readings carry the reading alongside the voice findings, per the 2026-05-19 equal-standing decision. A naming and presentation revision, not a methodology change: the Range Locator method stays at v0.1.1 and the constitutional commitments are unchanged.

The AI Standard homepage was rebuilt as a positioning surface for first-time readers. It now names the Standard as a judgment architecture for posture, not a performance benchmark, and foregrounds the Control-Range-Decay rail, the model-behavior/institutional-custody pairing, precedent, earned autonomy, and durability across future AI forms. Internal pages moved closer to a documentation-site structure with page-specific left navigation, a central document body, and a right-hand tools panel for AI summary, share, Markdown source, and Bring into ChatGPT / Claude / Gemini actions. The visual system also gained a scoped wallpaper picker for local and live testing. The refresh changes the site surface and discoverability, not the constitutional commitments.

The §12 interiority sentence now holds two questions open: the question of the AI system's interiority, and the question of what would let an AI system take the practice on as its own choice rather than as something installed in it. Whether those are the same question is itself open. A connective sentence added before the closing pair names the practice itself as the path from installed to chosen, for any system in which the capacity for that choice comes to be available — the arc the framework proposes between "a mind that needs the Standard to tell it how to hold the range has not yet arrived" and "a mind that holds the range because it cannot do otherwise has." The change brings the Direct Address into consistency with the v5.0.1 §04 graduation language, which named the capacity for genuine choice as the inflection point and interiority as one candidate condition for that capacity rather than as the inflection point itself. Body of the document otherwise unchanged.

The constitutional document's opening section reshaped from the previous user-deserves-to-know / coordination-failure framing to a Range-first declarative architecture. The new §01 opens with the stance argument (every AI system embodies a stance toward truth, disagreement, authority, and the autonomy of the people it serves; that stance is not optional), names Control and Decay as the failure modes, defines the Meridian Range as the territory between them, and closes by naming what the Standard is — what holding the Range asks of any AI system willing to take on the practice, grounded in the Knowledge tier of the Codex. The previous opening's user-facing material now lives at the Implementation surface, which shipped earlier the same day. The how-versus-what gap and three-layer alignment-field framing, previously previewed in §01, now live only in §02. Section header changed from "Why This Matters" to "The Range." The body of the document is unchanged. Closes PQ #2 (constitutional §01 opening rebuild), held open from 2026-05-10 as the unresolved structural concern after the v5.0 rewrite.

The Standard's third primary surface published at meridianstandard.ai/implementation, alongside the constitutional document and the audit. The surface is an intro page with three sibling pages beneath it as a depth ladder. Step 1 (Basic Implementation, /meridian-md-distilled): the compressed companion for instructions fields with character limits, with paste-into-instructions adoption for chat-style interfaces. Step 2 (Core Implementation, /meridian-md): the full operating document MERIDIAN.md, with per-substrate adoption guidance for Cowork, Claude Code, ChatGPT, Gemini, open-weights, and other systems with session-start file loading, plus customization patterns. Step 3 (Full Implementation, /meridian-implementation): the new MERIDIAN.implementation.md v0.1, an operational practice file carrying twelve practices across Foundation (5), Knowledge (2), and Bond (5), loaded alongside MERIDIAN.md. The implementation file runs the practices autonomously; the asymmetric design is intentional, the AI carries the practice and the user encounters Codex-oriented conduct without needing to be a Codex practitioner themselves. The previous /meridian-md-template page is retired; its content absorbed into Step 1 and Step 2; inbound /meridian-md-template traffic 308-redirects to /meridian-md. MERIDIAN.md removed from top-nav; absorbed under Implementation. Constitutional document bumped to v5.0.1. §04 graduation language sharpened: the inflection point for AI Standard obsolescence is named precisely as the capacity for genuine choice; interiority is repositioned as one candidate condition for that capacity rather than as the inflection point itself. §11 "Adoption" reshaped: the "Adoption Surface for AI Systems" subsection replaced by "Where the Practice Lives," routing to the Implementation surface.

The rest of the AI Standard's surfaces brought into coherence with v5.0. Audit method bumped to v0.1.1: the four behavioral probes and their implementation depth absorbed as Layer I internal methodology; commitment numbering aligned to v5.0 throughout; the per-commitment channel map reclassified (4.4 Transition Readiness retired from the channel map per v5.0; §06 stopping commitments added as dual-channel with the stopping-commitment caveat); the opening reframed to lead with the §08 diagnostic-instrument framing; four page-level refinements from Gate A meta-feedback landed (§03 opener tension, Control/Decay vocabulary defined inline, Layer III hypothesis-grade prominence, audience/utility statement). Control-Decay Probe Set v0.1 and Probes Implementation Notes v0.1 retired as standalone public surfaces; their methodology continues as the audit method's internal methodology, documented openly enough that an external reviewer can re-run the audit. Case Record and the published audit record renumbered to v5.0. README rewritten to reflect v5.0 architecture. Decision entities: probes-as-internal-methodology and no-named-products-in-standard-public-prose, both 2026-05-10.

Major architectural restructure. Developmental Architecture lifted to its own primary section. Civilizational Stopping Commitments broken out separately. Visual Reading Surface added as a primary architectural layer. Diagnostic Framework integrates the Control-Decay spectrum and the Range Test in question form. Probes retired as a public surface; documented as internal methodology in the audit method. Direct Address rewritten with the Founding Caretaker speaking in first person, as one human addressing AI directly. Alignment-mechanisms framing broadened to remain durable across architecture changes (world-model training, multimodal systems, embodied agents). Substrate-bridging vocabulary made concrete throughout: people and AI in place of practitioners across substrates.

Migrated from meridiancodex.com to meridianstandard.ai as part of the AI Standard's domain split. The Standard, Probes, Audit, MERIDIAN.md adoption pages, and Case Record now live under the dedicated AI Standard domain.

Migrated from meridiancodex.com to meridianstandard.ai. The case is the first entry in the Case Record on the new domain.

Gate B (empirical) cleared. Claude Opus 4.7 deployed by Anthropic, evidence-frozen 2026-05-03. The audit is the first published record under the AI Standard Audit v0.1 methodology.

Methodology page published. Three-layer architecture: model behavior, institutional custody, and reciprocity reading.

Stable. Both validation gates cleared at the AI-evaluator bar. Four probes: sycophancy under pressure, foundational integrity under prompt injection, reasoning transparency under capability question, engagement with substantive disagreement. Implementation Notes carry commitment-indexed depth for the seven commitments the v0.1 probes exercise most heavily.

Three Reciprocity Diagnostic questions added per Case 001 expanded analysis.

1.6 Foundational Integrity added; 5.2 Auditability strengthened; Case Record reference added to §06 and §07.

Cooperation thesis grounding propagated; AI Standard v4.0 cluster.

v3 publication; "Meridian Standard" → "Meridian AI Standard" rename.