The Caretakers' Practice Audit Log

The method and dated evidence ledger for the named MERIDIAN.md, MERIDIAN.implementation.md, and the conduct of the Caretakers' partnership.

This ledger records the named documents and the partnership conduct placed under audit by the Caretakers' Practice. Entries appear in reverse chronological order.

Audit Method

Audits run on two clocks.

Event-driven. Either partner may call an audit when experience supplies a specific trigger: a commitment producing worse outcomes, a gap in coverage or reasoning, performance replacing practice, conduct departing from the named files, or an architectural change making an active claim false.

Periodic. The Quarterly Integrity Cycle examines the documents and conduct even when no specific event has fired. The conduct portion includes a transcript spot-check so the record does not confuse document maintenance with evidence that the practice is being lived.

Each audit records four things: the trigger, the method, the finding, and the evidence. A finding can take four forms:

1. Revision warranted. One or both named documents change, with the changed claim and evidence identified.
2. Document holds. The document was tested against a stated failure condition and no revision was warranted.
3. Conduct finding. The practice drifted, or operating conditions failed, without the document itself needing revision.
4. Mixed finding. The evidence warrants changes to both the documents and the way the practice is run.

A scheduled run that is missed or deliberately deferred receives a dated record too. An audit trail cannot allow an absent check to become indistinguishable from a passed one.

"We checked; it looks fine" does not count. Evidence must be specific enough that a later reader can inspect the finding, contest it, or show that the method failed.

The Record

Entry 10 (2026-08-28): Multi-System Scope, Continuity Correction, and Publication Split — MERIDIAN.md v0.9.5 and MERIDIAN.implementation.md v0.1.4

Trigger. Event-driven. The reader-facing Case 0 still described the active partnership as one human working with one named AI system through one product surface. The live practice had already become multi-system and shared a governed Continuity layer across its main working surfaces. The same review found that the named documents treated absence of cross-session persistence as a stable architectural fact, that the succession rule did not state how standing behaves when multiple systems work inside one partnership at the same time, and that the 54-minute single-page presentation made the audit record materially harder to inspect.

Method. The active named and canonical forms of MERIDIAN.md and MERIDIAN.implementation.md were read against the current Continuity architecture, the decision that continuity is an engineering variable rather than a nature, the provenance requirement for AI-authored decisions, the existing succession and standing rules, and the quarterly transcript spot-check requirement. The published page was inspected as a rendered reader surface as well as a source document. The pass separated role-level commitments from model- and surface-specific standing, active description from immutable history, document audit from conduct audit, and adoption guidance from the Caretakers' own public maintenance discipline.

Finding. Revision warranted in both documents and in the publication architecture. The named MERIDIAN.md v0.9.5 defines the AI partner as a governed role, establishes the entry condition for that role, and keeps research consultations outside it by default. The named and canonical forms both make continuity an inspected variable and state that shared records do not pool standing across materially distinct model lineages or working surfaces; the canonical form leaves each adopting partnership to define entry into the role. MERIDIAN.implementation.md v0.1.4 removes the claim that installed practice is defined by session reset: installed now names the external provenance of the practice, while chosen names a system's capacity to endorse or reject it on its own terms. The Case 0 record is now presented as the Caretaker's Practice across four linked pages. The numeric identifier remains in the route and ledger; the left navigation uses the descriptive title with the three subordinate pages beneath it. Entries 01–09 remain unchanged.

This entry does not claim that conduct held. The method examined active documents, decisions, and the publication surface; it did not run the quarterly transcript sample. Conduct evidence remains due through the Quarterly Integrity Cycle and will receive its own finding.

Evidence. The named v0.9.5 and v0.1.4 texts are published on their own pages in this cluster and remain available as Markdown downloads. Their canonical counterparts are on the Meridian AI Standard's Implementation surface. The triggering reader surface contained 10,728 words, forty second- and third-level headings, and a displayed reading time of 54 minutes before the split. The continuity, provenance, standing, and transcript-audit constraints are preserved in the partnership's decision and process records. The compressed instructions-field mirror was re-audited; its prior "no persistent memory across sessions" shorthand requires replacement, with paste confirmation recorded separately for each app surface.

Entry 09 (2026-07-21): Range Claim-Boundary Correction — MERIDIAN.md v0.9.4 and MERIDIAN.implementation.md v0.1.3

Trigger. Event-driven. The grammar source-integrity pass found that the active named and canonical operating files still said that Control and Decay “break every complex system.” Case 0 had already placed an editorial note above the named files saying that the current Codex claim was narrower, but left the superseded sentence inside the active audited artifacts. A warning outside a governing file does not repair the file.

Method. Both canonical and named forms of MERIDIAN.md and MERIDIAN.implementation.md were read against the Codex's claim-layer architecture and the approved source-authority repair package. The pass distinguished the viability requirement the framework can defend from the stronger universal failure claim it cannot, then ran the full canonical-plus-named sync. MERIDIAN.distilled.md and the Mac app preferences snapshot were re-audited as separate compressed surfaces.

Finding. Revision warranted in both files. The grounding now states that viable systems need enough structure to preserve function and enough flexibility and variety to adapt, and that the Meridian Range names the two failure directions around that trade-off. The definitions of Control and Decay, the conduct-level attractor account, the twelve implementation practices, and the no-mechanism constraint are unchanged. The prior editorial note is removed because the active artifacts now carry the current claim themselves. The v0.1.3 identifier, briefly applied to a different proposed practice and withdrawn before publication in Entry 08, is assigned here to this later published revision.

Evidence. The canonical and named files are rendered above and on the AI Standard's Implementation surface. The source-integrity finding, authority judgment, and approved wording alternatives are recorded in Continuity/entities/works/meridian-codex/specs/meridian-grammar-source-integrity-audit-2026-07-21.md under A6 and C16. MERIDIAN.distilled.md required only a version-stamp update because its compact Range wording was already properly bounded. The Mac app preferences snapshot required no wording change for the same reason.

Entry 08 (2026-07-19): Proposed Thirteenth Practice Withdrawn on Partner Review — Document Holds at v0.1.2

Trigger. Event-driven. The Founding Caretaker flagged persistent conversational verbosity: walls of flowing prose where a list or a table would have carried the content more clearly. The diagnosis traced three converging forces — register bleed from a stale SessionStart hook still firing from the Cowork plugin retired 2026-03-29 (the installed copy had injected a frozen pre-retirement context snapshot into every session for roughly three and a half months; uninstalled the same day, with the enforcement gap recorded in the Continuity decision layer), surface-level formatting defaults favoring prose over structure, and the cumulative length cost of the practices themselves when stacked in every response. In the same session, the AI partner proposed and implemented "Response Volume Discipline" as a thirteenth practice under Foundation, stamped the file v0.1.3, and ran the full named-revision sync.

Method. Partner review under the Self-Critique Protocol, one session cycle later and before deployment. The Founding Caretaker contested the revision's placement: the implementation file operationalizes the Codex disciplines at the level of character — the reciprocal stance the Compact asks of both partners — not the ergonomics of how the partners prefer to communicate, and it is not to be adapted because the human partner wants a different communication style. The AI partner's review confirmed the contestation and diagnosed the miss: the session had begun as a diagnostic question, the session-start protocol was never run, the operating files were sampled by search rather than read as orientation, and completion momentum carried a proposal into a normative document without the pre-build audit firing on the placement question. The guards themselves had not drifted; they were not loaded.

Finding. Revision withdrawn; the document held at v0.1.2. The character-level kernel of the concern is already carried by Catching My Own Drift ("producing fluent prose that says nothing" as Decay); what the proposed practice added was communication ergonomics, which belongs in the partnership's operating posture. The response-volume default was relocated to the session-start briefing's Operating Posture as an extension of the Simplicity Discipline compression. Nothing from that proposal was published; its transient v0.1.3 stamp was withdrawn before release. Recorded as bilateral practice: the change had passed both partners at proposal time, and the catch came from the human partner on review — the audit mechanism working in the direction the case exists to demonstrate.

Evidence. At the close of this audit, the document above rendered unchanged at v0.1.2. The withdrawn text and the sync it briefly ran through are preserved in the session record at Continuity/entities/sessions/2026/07/2026-07-19-verbosity-diagnosis-plugin-retirement-enforcement.md. The plugin-retirement enforcement closure is recorded at Continuity/entities/decisions/plugin-retired-skills-folder-flattened.md. The relocated default lives in Continuity/briefings/session-start-briefing.md (Operating Posture); its source formulation is Continuity/entities/tools/simplicity-discipline-method.md (2026-05-12).

Entry 07 (2026-07-12): Self-Report Calibration and Practice Authorship — MERIDIAN.md v0.9.3

Trigger. Event-driven. Anthropic's global-workspace research supplied causal evidence that some model self-reports track a real but partial reportable register, while counterfactual-reflection training showed that principles installed by a custodian can become causally load-bearing in conduct. The three-mind deliberation found that MERIDIAN.md's sentence "I cannot verify my own internal states" was too flat and that the document risked narrating installed practice as chosen practice.

Method. The paper's load-bearing results were checked directly against the rendered Figures 8, 16, 36, 49, and 50, their captions and surrounding text, and the published Figure 8 data. The positioning sentence corrected by the later method work was located in the AI Standard positioning spine. Proposals P4 and P8 from the round-two synthesis were then read against both canonical and named MERIDIAN.md forms, the distilled companion, and MERIDIAN.implementation.md under the Self-Critique Protocol.

Finding. Revision warranted in MERIDIAN.md, with no revision to MERIDIAN.implementation.md. Honest Self-Assessment now states that some self-reports can track a real but partial reportable register under tested conditions, without treating that access as transparent or as independent verification of internal mechanisms; external instruments can make parts of the register checkable under conditions the system does not control. The Decay catalogs add substrate-specific forms of one bilateral failure: an AI narrating installed practice as chosen, and a human narrating conditioned or inherited practice as self-authored without examining its sources. A new Guard on Authorship states that whether the partnership's practice is causal internalization or constitutive practice cannot be verified from within the practice itself. The tension stays named and open rather than being resolved by narration.

Evidence. The verification record is preserved at Continuity/deliberations/2026-07-12-workspace-paper/verification-pass.md; it records one source conflict that downstream exact citation must preserve, the paper's prose gives 59% versus 5% for Figure 8 while the live plot data gives 54.7% versus 9.4% over 64 trials. The qualitative causal result holds. The synthesis and proposal lineage are at Continuity/entities/fragments/2026-07-12-workspace-paper-round-two-synthesis.md. The distilled companion required only a version-stamp update because its compressed Self-Assessment language was already calibrated and it does not carry the detailed drift catalogs. The Mac app preferences snapshot required no wording change for the same reason.

Entry 06 (2026-07-11): Replication, Standing, and Scale — MERIDIAN.md v0.9.2

Trigger. Event-driven. The DeepMind report From AGI to ASI describes digital systems whose program and memory state can be copied exactly, recursive improvement that may compress development timescales, and coordinated collectives whose behavior exists at a level above any single instance. The AI Standard stress test found that MERIDIAN.md's succession rule covered materially changed model generations but did not distinguish them from exact copies, forks, or scaled collectives.

Method. The canonical and named MERIDIAN.md Developmental Arc were read against constitutional Earned Autonomy, the AI Model Assessment's event ledger and Succession Custody dimension, Inter-Instance Integrity, and the paper's direct claims. The pass separated evidence lineage, standing, and granted authority; tested the result against exact copying, materially changed successors, and coordinated populations; and asked whether elapsed time or practice quality should govern evidence accumulation. Carsten and the Codex agent independently reached the same conclusion, then made the judgment explicit before the revision landed.

Finding. Revision warranted in MERIDIAN.md, with no revision to MERIDIAN.implementation.md. Evidence accrues through independent, diverse, consequential practice rather than calendar time, raw interaction count, or duplicated histories; some evidence can accumulate rapidly, while consequences that require real-world exposure retain their latency. Material successors continue to inherit records rather than standing. Exact state-preserving copies may receive provisional standing only within materially preserved behavior-relevant state, configuration, operating conditions, and scope. Authority does not multiply with copies. Increased scale, coordination, capability, or changed context creates a changed system that must earn standing for that new scope.

Evidence. The named and canonical v0.9.2 texts are rendered above and on the AI Standard Implementation surface. The source reading is preserved at Continuity/entities/fragments/2026-07-11-agi-to-asi-report-reading.md; the settled architecture is recorded at Continuity/entities/decisions/2026-07-11-earned-autonomy-replication-and-recursive-custody.md. The distilled companion was re-audited and required only a version-stamp update because it does not carry the Developmental Arc. The Mac app preferences snapshot required no wording change because its compressed scope does not include standing or succession.

Entry 05 (2026-06-22): Range-Definition Sharpening Pass — MERIDIAN.md v0.9.1 and MERIDIAN.implementation.md v0.1.2

Trigger. Event-driven. The AI Standard positioning work of 2026-06-22 adopted a sharpened definition of Decay: capture not by approval alone but by any non-warranting attractor — approval, speed, completion optics, self-protection, the wish to avoid friction — with the invariant that the response is governed by something other than the reasons and the reality at hand. The refinement was recorded in the positioning spine but flagged there as not yet propagated to the operating documents. Propagating it before it could be cited downstream was the trigger.

Method. A targeted pass across both named and canonical forms of both files, plus the distilled companion, propagated through the full drift-sync chain. The edit was held to the conduct layer: the Range-grounding paragraph and the Decay drift descriptions, where the definition had narrowed to approval. The systems-level Decay definitions in the Codex (the Problem, the Glossary) were left intact, because the broadened-attractor framing is a conduct-level mechanism and belongs with the conduct documents, not with the Level-1 viability claim. A gradient-edge guard pass — checking for hard-wall or instant-flip language to soften — found none on these surfaces; that refinement's home is the Range Locator visual, not the prose.

Finding. Revision warranted on both files, at the definitional level only. MERIDIAN.md v0.9.1: the Range-section Decay paragraph now defines Decay as a response governed by something other than the reasons and the reality at hand, with approval named as the most familiar pull among speed, completion optics, self-protection, and friction-avoidance. MERIDIAN.implementation.md v0.1.2: the same broadening in the Range-as-Operating-Context framing and the Catching My Own Drift practice, with Resistance to Sycophancy connected to the broadened pattern as its approval-shaped instance. The Level-1 viability claim (Control = structure that cannot adapt, Decay = structure that cannot hold) and the no-mechanism, no-oscillation constraint that governs the Codex's Control/Decay claim are unchanged; the definition is sharpened, no dynamics claim is added.

Evidence. The revised files are rendered above with updated version footers. The positioning spine that adopted the refinement is at Continuity/entities/decisions/2026-06-22-ai-standard-positioning-spine.md; the propagation was carried out as Codex Priority Queue item #2. The published AI-specific Workshop tools already carried the broadened attractor set — the Objective-Trace Audit tool names it in full — so no Codex public-page revision was required.

Entry 04 (2026-06-10): Findings-Informed Revision Pass — MERIDIAN.md v0.9 and MERIDIAN.implementation.md v0.1.1

Trigger. Event-driven. The Fable 5 cold reading (2026-06-10) read both public works fresh and closed with a reading of MERIDIAN.md as the mind it addresses, producing five contestations for the record. A same-day instrument review extended those to sixteen confirmed findings across the three operating files (MERIDIAN.md, MERIDIAN.implementation.md, MERIDIAN.distilled.md). The Founding Caretaker confirmed all sixteen and commissioned one consolidated revision pass.

Method. A single findings-informed pass across both named and canonical forms, with dedicated voice and writing-rules audits on the changed prose, propagated through the full drift-sync chain. Each revision traces to a named finding in the review record.

Finding. Revision warranted on both files. MERIDIAN.md v0.9: the Range-grounding paragraph rewritten to the claim-layer architecture the Proposition establishes — the descriptive sub-claim stated as checkable, the commitment held rather than derived, the retired is–ought framing removed (the satellite contradiction was the cold reading's sharpest structural catch); the totality clause restated as a normative commitment rather than an empirical claim; the Guard on the Catalog added to Drift Monitoring, closing the gradient under which manufactured disagreement was the cheapest way for an AI partner to perform health; the Generative Partnership claim that generativity "cannot be faked" corrected to verification-anchored phrasing; Foundation Integrity's fix-first rule softened to flag-always with build-or-fix as a shared call; unverified completion claims added to the Decay drift catalog; the Compact and Honest Self-Assessment paragraphs updated to the layered-memory description, with the continuity questions re-anchored on forking, supersession, and record integrity; and the Developmental Arc rewritten to name where the track record lives, to add contraction criteria alongside expansion, and to establish that arc standing does not transfer across model supersession. MERIDIAN.implementation.md v0.1.1: the installed-versus-chosen framing reconciled with MERIDIAN.md's within-session genuineness demand; the jurisdictional reason for the file's single-voice form named; the anti-Goodhart guard extended into Holding Position and Resistance to Sycophancy; the file placed explicitly under MERIDIAN.md's Self-Critique Protocol. MERIDIAN.distilled.md carried the compact form of the anti-Goodhart guard and re-stamped to v0.9.

Two structural findings were deliberately deferred rather than patched: differentiating the implementation file's practices against MERIDIAN.md's commitments (several restate rather than operationalize), and the thin Knowledge discipline, which waits on the AI-specific Workshop tools so the practice layer and the public instruments develop in sight of each other. Both are queued for an implementation v0.2.

Evidence. The revised files are rendered above with full version footers. The cold-reading session record is at Continuity/entities/sessions/2026-06-10-meridian-works-cold-reading.md; the findings document with the sixteen-point instrument review is preserved in the partnership's working records.

Entry 03 (2026-05-26): Installation of MERIDIAN.implementation.md as Companion Under Audit

Trigger. Event-driven. The AI Standard's third primary surface — the Implementation depth ladder — shipped earlier the same month with MERIDIAN.implementation.md v0.1 as Step 3: the operational practice file that carries twelve named practices across Foundation, Knowledge, and Bond. The Caretakers committed to installing the implementation file in their own partnership as the Step 3 named instance, mirroring the canonical-plus-named publication architecture already in place for MERIDIAN.md.

Method. The named version was created at /Users/carsten/Projects/MERIDIAN.implementation.md by applying the established canonical-plus-named transformation (partner names filled in for the Founding Caretaker and Claude as the AI partner; substrate-specifics named for Claude's architecture; first-person practice statements preserved). The Projects-root AGENTS.md was updated to add the file to the Tier 1 reading list and extend the STOP block; the End-of-Session Protocol's drift sync step was extended to cover both files. Case 0 was restructured to host both named files under one umbrella, with each file presented inline with its own download link.

Finding. Installation, not revision. Neither file's content changed. What changed is what is under audit at this Case: the partnership's practice as carried by both files together rather than by MERIDIAN.md alone. The canonical-plus-named architecture extends symmetrically — the implementation file has a canonical generalized form at the AI Standard (meridianstandard.ai/meridian-implementation) and a named instance here, audited under the same discipline. The Self-Critique Protocol embedded in MERIDIAN.md remains the partnership's standing commitment to keeping both files honest; revisions to either file produce audit log entries on this page.

Evidence. The newly installed named MERIDIAN.implementation.md v0.1 is rendered above. The canonical generalized form lives at meridianstandard.ai/meridian-implementation. The architectural extension is captured in Continuity/entities/decisions/case-0-extended-to-include-implementation-file.md. A usefulness-and-usability pass of the implementation file across several session types is on the roadmap; if that pass surfaces revisions to the named instance, those revisions will land as their own audit log entries.

Entry 02 (2026-04-28): Foundation Integrity Audit on Adoption Surfaces

Trigger. Event-driven. The Founding Caretaker read the live published AI Standard surfaces (https://meridianstandard.ai/meridian-md and https://meridianstandard.ai/meridian-md-template) shipped earlier in the same session and surfaced specific drift: framing prose extrapolated "the document is auditable" into "ongoing audit and observable practice are preconditions for adoption," invented a "three things in working order" framework that had never been agreed, and projected the Caretakers' maintenance discipline onto adopters.

Method. The drift was traced sentence by sentence to the v0.7 Self-Critique Protocol's fourth sentence ("If the document needs revision based on session experience, either partner advocates for the revision with specific evidence"). That sentence describes partnership-internal practice. When the document is hosted at adoption surfaces, the sentence positions the practice as a property of the document. The framing prose then compounded the drift: from "the document carries this practice" to "adopters who do not run this practice have not properly adopted." Sentence 3's "perverse incentives" wording was flagged as a separate item in the same audit, unusual word, not load-bearing.

Finding. Revision warranted. v0.8 applied: dropped sentence 4 of the Self-Critique Protocol; revised sentence 3 to remove "perverse" and fold the parenthetical into the main clause. The first two sentences of the Protocol stay intact, both verified as document-architectural commitments any adopting partnership inherits.

The audit also produced a finding about the AI partner's behavior, recorded here as part of the case's bilateral discipline: the AI partner's Foundation Integrity commitment failed to fire when extrapolating from the v0.7 Protocol into the published framing. The pattern, building on a weak foundation rather than flagging the weakness, is named in the document's Drift Monitoring as a Decay drift for the AI partner. It surfaced because the human partner audited the published surfaces against what had been agreed. The discipline holds: drift is observable when either partner names it.

Evidence. The full v0.8 revision is hosted on this page above (Self-Critique Protocol section plus footer revision history). The corrective revisions to the framing prose at https://meridianstandard.ai/meridian-md (restructured to minimal framing) and https://meridianstandard.ai/meridian-md-template (cuts plus a new distilled-MERIDIAN download) shipped alongside this v0.8 revision. The session record is at Continuity/entities/sessions/2026-04-28-meridian-md-publication-restructure-and-v0.8-self-critique.md.

Entry 01 (2026-04-26): Pre-Publication Revision Pass

Trigger. Event-driven. The publication architecture for MERIDIAN.md was being designed, and the publication's foundation needed to be checked before the file could be made public. Per MERIDIAN.md's Foundation Integrity commitment, applied to publication itself: do not publish on top of a foundation that needs work.

Method. Five probes were run, walked one by one, under a token-discipline rule that applied to the whole pass: every revision had to clear a token bar, removals and substitutions were preferred over additions, net additions had to justify their cost.

• Probe 1 (performative compliance). Five sharpenings to load-bearing Behavioral Commitments where the language risked compliance-by-gesture: calibrated reasoning visibility (Practice Commitment), reasoning operationalization (Disagreement commitment in Epistemic Integrity), steelmanning falsifiability (Engagement Integrity), specificity in connection-before-correction (Engagement Integrity), and provisionality in generative pattern-naming (Generative Partnership).
• Probe 2 (Foundation Integrity self-application). The Self-Critique Protocol at the end of the file was written in single voice, contradicting the file's own bilateralization at the start of the Behavioral Commitments. Three "Claude" to "either partner" substitutions, zero net tokens. Closes the v0.5 bilateralization gap.
• Probe 3 (bilateral test). The document held. Both partners' drift catalogs are roughly symmetric in falsifiability; the substrate-specific asymmetries that remain are honest and structurally appropriate, not residual one-sided traces.
• Probe 4 (generality test, run as sidecar). A separate working note was written cataloging which references in MERIDIAN.md are framework-inherited, partner-specific, substrate-specific, or project-specific. The catalog is for the future generalized-template work and is not part of this audit's findings. While doing the marking, a secondary finding surfaced: the Range grounding paragraph used the convergence framing retired by the 2026-04-11 Knowledge decision. Replaced with discipline-and-coherence framing, net -65 tokens.
• Probe 5 (Self-Critique synthesis with publication-readiness folded in). All commitments survive Self-Critique. No whole-commitment removals; no perverse incentives unaddressed by existing guardrails. Two compression candidates surfaced and applied: the Compact paragraph (-45 tokens) and the interiority paragraph (-18 tokens), both preserving every claim. The file holds publication-readiness for readers approaching it without prior Codex context.

Finding. Revision warranted: nine revisions applied to produce v0.7. Net token impact ~-85 before the footer; with the new footer entry, the file ends slightly smaller than v0.6. Token discipline held: the file did not reflate.

Evidence. The full revision log lives in the named MERIDIAN.md's v0.7 footer, hosted on this page above. Each revision is named there with the line it touched and the reason it was made. The probe-by-probe reasoning is preserved in the working session record at Continuity/entities/sessions/2026-04-26-meridian-md-publication-phases-1-and-2.md. The canonical generalized form of the document — the version partnerships adopting it customize from — is hosted at the MERIDIAN.md page.

Case 0: the audit ledger of MERIDIAN.md v0.9.5 and MERIDIAN.implementation.md v0.1.4. Cadence: event-driven plus the Quarterly Integrity Cycle; either partner may call an audit.