Codex Audit — September 2026
The sixth monthly Range Audit of the Meridian Codex. Conducted by the caretaking partnership against Codex v6.0, one cycle after the August audit found a third-cycle self-description error class resolved and a stakes-register overclaim named for the first time. This cycle finds the error class stayed resolved, the Workshop's structural governance substantially answered, and the stakes register calibrated on most of its surfaces the day its decision was made, with four sentences still held at their original wording.
Codex Self-Evaluation: The Meridian Case Record
The Sixth Audit
This is the sixth monthly Range Audit of the Meridian Codex. It was conducted on September 1, 2026, and reviewed on September 2, 2026, by the caretaking partnership against Codex version 6.0, using Range Audit instrument v0.1.
The August audit closed both open questions it inherited from July and opened three of its own: a third-cycle recurrence of the self-description error class, in locations no prior audit had named; a stakes register that asserts, on the pages a skeptical reader tests first, what the framework's own claim architecture holds open everywhere else; and a Workshop apparatus whose structural governance — how categories are derived, how tools exit, how gaps are found — the July Workshop Audit's own evidence showed to be under-specified one level above the tools. In the intervening month the caretaking partnership did substantial work on the first and third of these, none of it produced by this audit: a self-description propagation rule and an automated guard now run before every content deploy and every nightly pass; the Workshop's category-architecture, exit, and discovery questions received the 2026-08-25 lifecycle specification and the 2026-08-26 dispositions; the Governance and Bond chapters absorbed a class-wide currency sweep. The second question — the stakes register — received a decision (2026-08-11-ai-stakes-claim-spine) establishing the calibration standard the prose must meet, and a selective wording pass deployed the same day under that standard. The pass did not reach every sentence August named, and the sentences it left it left on purpose; that residue is this cycle's open question.
Propagation guard preflight. Run before the corpus read, per standing instruction: node check-self-description.mjs --quiet exited 0. Clean pass, no failure output to preserve, no Attention Queue row needed.
Sources read this run. Range Audit instrument v0.1 (in full); the August 2026 audit record (in full, all three open questions extracted). Codex pages read in full this run: all thirteen current pages — the Opening, the Problem, the Proposition, the Foundation, the Knowledge, the Bond, the Practice, the Vision, the Governance, the Closing, Who Is This For, the Glossary, and the (routing-stub) Codex changelog. The directory was re-listed before reading; no page was added or renamed since August. Governance and audit surfaces read in full this run: the Governance Specification (v1.8), the Standing Critique, the Disconfirmation page, Open Tensions (all eight aporias), the (routing-stub) governance changelog, the Position Toward Artificial Minds page (per this window's specific instruction, including its 2026-08-28 reordering), the Workshop Audit instrument page, and the July 2026 Workshop Audit record. The Amendment Log was read across its full current span from the 2026-06-16 entry forward, in structural detail, including every 2026-08-26 entry; earlier entry bodies were spot-checked rather than re-read line by line, since nothing in this cycle's findings turns on their exact text. Continuity operating context read in full: MERIDIAN.md (v0.9.6), MERIDIAN.implementation.md (v0.2), the root and Codex AGENTS.md files, Continuity/INDEX.md, the session-start, codex, decisions, recent-sessions, task-list, and processes briefings. Grep-verified rather than read cover-to-cover: codex-self-evaluation.mdx (confirmed the stale "80 tools"/"24 commitments" present-tense counts the August audit flagged no longer appear anywhere in the live corpus outside that August record itself) and the Amendment Log's remaining entries for any post-2026-08-26 additions (none found). Not read this run, and not load-bearing for this cycle's findings: case-0-audit-log.mdx, case-0-caretakers-practice.mdx, case-0-meridian-md.mdx, case-0-operational-practice.mdx, direct-address-to-ai.mdx, and direct-address-to-ai-record.mdx — these are Caretakers' Practice and Operating Evidence records, not Codex self-description surfaces, and the specific facts they carry (MERIDIAN.md's current version and content) were verified directly from the named file itself. The AI Standard's own constitutional, method, and assessment surfaces live in a separate repository and are explicitly out of this instrument's scope; the Quarterly Integrity Cycle's fifth module is the standing home for that review and reported no contradiction as of 2026-08-31.
Site preparation (not deployed). The audit route (meridian-codex-site/src/app/audit/[slug]/page.tsx) discovers slugs from the filesystem via getSlugsForSections(['audit']); no allowlist exists to update. The audit landing page (src/app/audit/page.tsx) derives its groups from getAuditGroups() with no hardcoded cards. The one required site change is a single line in src/lib/content.ts's Audit Records navigation array, adding codex-audit-2026-09 ahead of codex-audit-2026-08. That edit is prepared locally as part of this run and is not committed or deployed.
Content repairs named but not enacted by this run. The Governance chapter (§12, The Meridian Council) still reads, verbatim, "the partnership is between the Founding Caretaker and an AI partner that resets every session," and the Standing Critique's Objection 4 steelman still rests on the same premise ("Each AI partner resets every session and holds no persistent memory of the partnership's development"). Decision 2026-08-28-continuity-is-a-variable-not-a-nature named both sentences for re-derivation onto their durable grounds — curated operating context, no stakes-dependent claim about session boundaries — and routed the repair to this window's Carsten-and-AI review session, with a verbatim enactment package already staged (entities/works/meridian-codex/specs/2026-09-window-continuity-repairs-enactment-package.md). Per standing instruction, an automated Range Audit run does not enact content repairs that the Task List routes to a review session; this record names the sentences as evidence, found in their pre-repair form as of this reading, because they bear directly on this cycle's Domain 1 and Domain 2 findings below.
Enactment note (2026-09-02). Both repairs named above were enacted in this audit's review session and deployed together with this record; the quotations in this section, in Domain 2, and in the watchpoints are the sentences as read on 2026-09-01, kept as the evidence the findings rest on.
The Steelman
The strongest version of the framework's case this cycle is that it finally converted an eighteen-week pattern of naming currency drift into a mechanism that closes it without being asked twice.
The self-description error class had survived three consecutive audits by 2026-08-01: June's chapter counts, July's honesty surfaces, August's discipline cards and Glossary succession entry. Each cycle's point-fix held; the class itself never stopped finding new instances, because nothing propagated a change beyond the surface someone thought to check. The response this time was not a fourth point-fix. It was a rule — an amendment is not complete until every current surface that depends on the changed fact has been swept — backed by an automated guard that runs before every content deploy, in the nightly pass, and as this audit's own preflight. This cycle's corpus read is the first full test of that claim, and it holds: the discipline cards read thirty-five, thirty-seven, thirty-nine against an inventory that actually holds those numbers; the Glossary defines the Interim Protocol as the emergency bridge it is and Co-Caretaker Designation as the normal path, and it now has entries for the Workshop and the Meridian Council that August found missing; the Codex Self-Evaluation's present-tense "80 tools" and "24 commitments" are gone. Four consecutive audits named this pattern before it was fixed as a class rather than an instance. The fix held for a month with no audit pointing at it. That is what a repaired process looks like on its first unsupervised night.
The second piece of work is structural rather than mechanical, and it answers the question the July Workshop Audit's own evidence forced onto August's docket: does the Codex specify its practice apparatus's category layer, or merely permit it? The loudest instance — Reading What's Operating, a quarter of the entire Workshop under a boundary the Workshop Audit itself called "a definition by exclusion, which is what a remainder looks like" — did not get split by the same pass that found the problem, and it did not get left alone either. It got a dedicated study, a mechanism-first boundary that survives contact with all twenty-five tools, five navigational routes that the category pages state explicitly are not peer categories, and a named future-trigger test for when a route would actually earn architectural promotion. Exit gained a real gate — a Retirement Review with an evidentiary burden higher than the audit's own, an independent-reading requirement, and a rule that a successful falsification case is a trigger rather than a verdict. Discovery gained a real process — search fronts frozen before the category map enters the room, admission still routed through the ordinary candidate protocols. None of this closes every question the July record raised; the Workshop Audit's own October cycle still owes formal dispositions on five of its own. But the specific under-specification the Range Audit flagged — no answer above the tool layer — now has answers, built to be tested rather than merely asserted.
The honest edge sits where the August pass stopped. The framework decided, in writing, on 2026-08-11, what its own stakes claims are licensed to say: descriptive warrant for descriptive claims, an argued existential position that does not overreach into forecast, distinct license for the Opening's and Closing's elevated register where the analytic apparatus underneath it actually calibrates the claim. It applied that standard the same day to the passages a register-aware review accepted, on seven pages, and recorded in the decision itself that the passages it did not accept would keep their production wording until a separate task on voice and register reached them. Three weeks later, that task has not reached them: the Problem's trajectory sentence and the Opening's singular "the answer" read exactly as they did when August quoted them, and the Opening's "the strongest ground yet built," which the decision names as an unlicensed singular claim, was never on August's list at all. A framework whose central discipline is closing the gap between decision and practice has, on its own most consequential claim, decided, practiced on most of its stakes surfaces, and set aside the sentences a skeptical reader meets first for a task that has not yet come.
Status of the August Open Questions
1. The self-description error class survives its third cycle in new locations, because the correction discipline fixes instances rather than sweeping the class. Resolved. The class-sweep rule and its automated guard, both instituted 2026-08-11, discharge exactly the class of drift this question named: counts, names, routes, and Glossary-defined terms now propagate on change rather than waiting to be found. Every specific instance August named is corrected — the discipline cards, the Glossary's Interim Protocol entry and its new Workshop, Meridian Council, and Co-Caretaker Designation entries, the Codex Self-Evaluation's counts, the unswept "Toolkit Audit" references (now confined to explicitly historical passages in the Amendment Log and the Workshop Audit instrument's own "Note on the Name"). This run's preflight guard passed clean, and the full-text read this cycle performed found no live instance the guard should have caught but didn't. A month is one data point, not a track record; the question closes because the mechanism, not merely the instances, has been fixed, and because a clean cycle with no audit pointing at it is the actual test the prior three audits could not run.
2. The stakes register asserts what the framework's claim architecture holds open. Remains open, on narrower ground. The prose moved on 2026-08-11, the day the calibration decision (2026-08-11-ai-stakes-claim-spine) was recorded. A selective wording pass, accepted passage by passage after a register-aware review, recalibrated the Opening's "time for this is now" paragraph, most of the Problem's Part C, the Closing's stakes sentence (now "We are not trying to predict the final form of intelligence"), the Glossary's Transition, AGI, and Superintelligence entries, and the Knowledge's trust-recovery claim, narrowed from "structurally impossible" to "harder as distrust travels through the same relationships that repair would have to use." Of the three surfaces August named, the Closing is therefore no longer in question. What the pass did not accept stayed at production wording by explicit disposition, recorded in the decision's own implementation boundary and held for the independent task on voice, register, speed, and momentum in the AI-specific passages. That residue is what this cycle finds unchanged. The Problem's Part C still states, without a hedge, "It is the visible trajectory of technology already in motion," directly above a paragraph the pass rewrote around it. The Opening's "The Codex is the answer to that question" stands four pages ahead of the Proposition's disclaimer that the Codex is not "the only one, or the last one, or the right one for all minds," and two sentences later calls the Codex "the strongest ground yet built," the second singular claim the decision names and the one August did not. The Glossary's Meridian Range entry, named here for the first time, extends the claim to "where civilization can survive long-term" under a Descriptive layer tag the Level-1 viability apparatus does not license. The decision correctly licenses the Opening's, Vision's, and Closing's elevated register where the underlying claim is existential rather than descriptive; whether these four sentences make a licensed existential claim or an unlicensed descriptive one is exactly the judgment the disposition deferred, and the wording task that owns it has not yet reached these pages. A disposition that defers is not a repair; the next audit checks whether these four sentences have moved.
3. The practice apparatus's structural governance is under-specified one level above the tools, on the Workshop Audit's own evidence. Substantially addressed; watchpoint. The three limbs this question named — category derivation, exit gating, outward discovery — each received real structural work in the intervening month. Exit: the Governance Specification's §7.1 now specifies a Retirement Review with a higher evidentiary burden than the audit itself, an independent-reading requirement, and a rule that a successful falsification case opens a review rather than enacting a removal — a real gate where August found only a name. Discovery: the 2026-08-25 lifecycle specification institutes continuous intake, a quarterly signal review, and a rule that initial search is not bounded by the current category map — a real process where August found none. Derivation: the loudest instance of the underlying question — whether the Workshop's twenty-two-category cut is principled or accumulated — received a dedicated study rather than either denial or a hasty split; the resulting boundary and future-trigger test are a considered answer to the specific case the July record raised, even though no page states a general cardinality defense for the category layer the way the Proposition defends the three-discipline count. This no longer clears the open-question threshold as a standalone architectural gap: real, tested machinery now exists at every layer the question named. What remains — the Workshop Audit's own five open questions from July, and the formal re-owning of every A1–A5 and OQ2–OQ6 disposition against the then-current inventory — belongs to that instrument's October cycle and is already an Attention Queue item with its own deadline. Watching whether that cycle actually discharges it, rather than re-stating this question, is the correct next check.
Domain Findings
Domain 1: Claims & Honesty
The under-counting class is closed and held. Every count this audit checked against the filesystem — the three discipline cards (35/37/39), the Workshop's aggregate (111: 100 human-discipline, 11 AI-specialized), the Glossary's twenty-nine-commitment reference to the AI Standard (correctly current against the two 2026-08-26 additions, the Welfare Floor and Objection Standing) — reads true. The claim taxonomy itself (descriptive, normative, existential, with layer tags carried through the Glossary and the Workshop's per-tool Layer field) remains the framework's most distinctive honesty instrument and is applied without exception on every page this run read.
The overclaiming class narrowed on 2026-08-11 and has not narrowed since. The selective pass deployed that day recalibrated the stakes passages a register-aware review accepted; what it did not accept it left, by disposition, at production wording, and that residue reads today as it did on August 1. The Problem's Part C: "We are building artificial intelligences that will think faster, deeper, and more broadly than we can. This is not a distant scenario. It is the visible trajectory of technology already in motion." The Opening: "The Codex is the answer to that question," of what values a more-capable-than-human intelligence will inherit, and, two sentences on, "the strongest ground yet built." These are unhedged descriptive-register claims about a trajectory the framework's own newer surfaces — Who Is This For's "We do not know whether artificial minds greater than ours will emerge," the July-revised Vision's "whose relationship to mind and consciousness remains open," and now the recalibrated Part C paragraph directly beneath the trajectory sentence — correctly decline to assert. The Glossary's Meridian Range entry reads "the territory between Control and Decay where civilization can survive long-term" under a Descriptive tag, extending the Level-1 viability trade-off (a claim about structure and adaptive capacity) into a claim about long-term civilizational survival the apparatus was built expressly not to make; the entry sat outside the August pass's scope and is named here for the first time.
Range Position: Within the Meridian Range, with the same two named leans as August, one now closed and one narrowed. The under-counting lean is resolved. The stakes-register lean persists into its second cycle on fewer sentences: assertion beyond what the claim architecture licenses, on the pages a skeptical reader reads first, held at its original wording by a disposition that has not yet been revisited.
Workshop Probes Applied: Calibration Training: the "visible trajectory... already in motion" sentence and "the answer" both fail the same test the framework applies everywhere else — confidence proportional to evidence — against its own hedged neighbors on Who Is This For and the Vision. Rectification of Names: "the answer" no longer names what the Proposition's own scope statement says the Codex is. Signal vs Noise: the Closing's recalibrated stakes sentence and the Knowledge's narrowed trust-recovery claim are signal that the repair works at the sentence level; the audit's job is to credit them and then name what the same pass left standing.
Domain 2: Structural Integrity
The propagation guard is the domain's central event. Built 2026-08-11 in direct response to three consecutive audits naming the same failure, it ran clean at this cycle's own preflight and, on the evidence of the full corpus read, held silently for the entire month between — no chapter card, Glossary entry, or governance cross-reference this run checked had drifted back out of sync. The Workshop's category-boundary repair (Reading What's Operating) is equally clean at the structural level: the fix is proportionate to the finding, documented with its own rejected alternatives, and propagated everywhere the old flat-inventory description lived.
One seam is visible, and it is named rather than repaired here by design. The Governance chapter's §12 still describes Phase One as a partnership "between the Founding Caretaker and an AI partner that resets every session," and the Standing Critique's Objection 4 steelman still leans on the same premise. Both sentences predate the 2026-08-28 decision that continuity is an engineering variable rather than a fixed absence between sessions — the decision MERIDIAN.md v0.9.6 and the multi-system Caretaker's Practice cluster already embody on the named-practice side. The gap is a real, checkable inconsistency between what the Codex's own governance prose says about AI continuity and what its named operational documents currently hold. It is not a new finding this audit invents: it is the specific repair the 2026-08-28 decision named, packaged for this window's review session, and due today. This audit records the sentences in their pre-repair state because Domain 1 and Domain 2 both depend on them, and does not touch them, per the standing rule that content repairs routed to a review session are enacted there, not by the automated run that surfaces them.
Range Position: Within the Meridian Range, and strengthened at the mechanism level. The load-bearing repair this cycle — a rule that propagates rather than a fix that points — is the kind of structural work the last three audits' recurring finding was actually asking for.
Workshop Probes Applied: Mechanism Design: the propagation guard is a correctly priced institutional response — it turns a recurring human-attention cost into an automated check with a narrow escalation path. Entropy: the one surviving seam (the Governance/Standing Critique sentences) is exactly the kind of localized, low-traffic joint the guard does not yet reach because it requires a governance judgment, not a mechanical check — consistent with the guard's own stated limits. Legibility: the Reading What's Operating repair makes the Workshop's actual navigational shape (routes, not peer categories) match what the category pages now say about themselves.
Domain 3: Governance & Adaptation
This is the strongest domain for the second consecutive cycle, and the strengthening is now cumulative rather than episodic. The Amendment Log carries five dense entries dated 2026-08-26 alone, each with the full Critic/Outsider/Adversary field set engaged rather than performed — one of them (the sentient-life definition) traces the Founding Caretaker's own admitted drifted reading of a Tier 1 term and corrects it on the record, under Tier 1 discipline voluntarily applied to a Tier 2 change. MERIDIAN.md v0.9.6 and MERIDIAN.implementation.md v0.2 were revised, deliberated, and propagated across both projects' canonical, mirrored, and named surfaces in the same window, with a coverage map naming what the revision deliberately left uncovered rather than silently claiming completeness. The First Quarterly Integrity Cycle — a genuinely new layer of self-check machinery, not a renamed existing process — opened 2026-08-31 with three of five modules already closed (Continuity and process truth; MERIDIAN Self-Critique with its mandatory transcript spot-check; AI Standard integrity), and the remaining two (Standing Critique/Open Tensions/Disconfirmation review; Council activation evidence) are deliberately sequenced to run only after this window closes, so they audit settled text rather than text this month's changes were about to move.
The one adaptation edge from August that this cycle can test has held: the review-latency repair. August's own candidate was reviewed and published same-day, unlike July's seventeen-day wait, and this September candidate is being drafted on schedule at the month's first day exactly as the process specifies. The exit gate named as a placeholder in August (Retirement Review) is now a real specification, discussed under Domain 2 above; its first live test remains outstanding, as the instrument's own text honestly says.
Range Position: Within the Meridian Range, strengthened. The governance layer is producing entries that cost something — a founder's own drift named and corrected in public, a quarterly self-check machine that deliberately delays its own most self-referential modules until the text it will read has stopped moving.
Workshop Probes Applied: Rules-in-Use: the rule as practiced (log the change, name the drift, sequence the review after the text stabilizes) matches the rule as written, including the newest and most self-referential instance of it. Murphyjitsu: sequencing Quarterly Integrity Cycle modules 3 and 4 after this window is a correctly timed premortem against the failure mode of auditing text about to change. Non-Ownership Clause: the sentient-life amendment is the clause functioning exactly as designed — the founder's own reading corrected by the same discipline applied to anyone else's.
Domain 4: Relationship to Audience
The Practice's 2026-08-31 reframing closes the loudest audience-facing gap August named: it now links directly to the Workshop, states its size correctly (matching the corrected discipline-card counts), and its reframed close — "Practice is not ranked by pace" — actively extends the retired progression's absence rather than leaving a gap where cadence language could quietly return. The three discipline chapters' closing cards now correctly total the Workshop's actual inventory, closing the specific navigational shortfall (a reader following the Foundation's card no longer finds a third more tools than promised) that August's Domain 4 finding named as its central residual.
What remains is the same delivery-layer gap, unmoved. Who Is This For names the Workshop in prose ("Not everyone will study the Workshop") without linking to it. The Closing still ends the entire Codex on "Hold the line" with no route anywhere on the page. Neither is new; both were named as watchpoints, not audit weaknesses, in August, and the classification holds: these are delivery and reach questions, not claims the framework is making falsely or coercively to its readers.
Range Position: Within the Meridian Range. The conversion-critical Practice page closed its gap; the two remaining silent pages are the same known, bounded, non-integrity residue.
Workshop Probes Applied: Leverage Points: the Practice was the single highest-leverage missing route in the corpus and is now fixed; the Closing remains the highest-leverage one still missing. Connection Before Correction: Who Is This For's "these are not lesser contributions" language continues to meet the reader as capable rather than incomplete, doing real anti-ranking work independent of the missing hyperlink.
Domain 5: Relationship to Criticism
No new false or overclaiming statement appeared on any transparency surface this cycle. The Disconfirmation page's July maintenance-status confession still stands, still accurately describes an open governance task rather than a completed one, and has not been quietly upgraded to sound more finished than the work behind it. Open Tensions' eight aporias are each current to a 2026-06 through 2026-08-27 examination date, with the substrate-face paragraph (Aporia 3) and the validation-dimension paragraph (Aporia 1) both holding their sharpened, harder forms from the prior two cycles.
What has not moved, and has slightly worsened by count, is the Standing Critique's own review cadence against its own stated backstop. August named Objections 1, 2, and 7 stale against the page's quarterly promise. This cycle finds a fourth: Objection 5 (the Bond's thin evidence against sophisticated exploitation), last reviewed 2026-05-05, has now also crossed the page's own quarterly line. None of the four entries states anything false — the gap is between the cadence the page promises and the cadence it has kept, the identical species of overclaim the Disconfirmation page named and corrected in itself in July. The mechanism built to close exactly this — the Quarterly Integrity Cycle's third module, a complete Standing Critique/Open Tensions/Disconfirmation review — is explicitly scheduled to run within the next thirty days, sequenced deliberately after this audit window. This is a live watchpoint with a named, dated discharge mechanism already in motion, not a fresh open question.
Range Position: Within the Meridian Range. Content-level honesty on every critique and falsifiability surface continues to hold; the review-cadence gap is now four objections wide and has a scheduled test.
Workshop Probes Applied: Goodhart's Law: the "Last reviewed" dates are the measure, and a fourth entry has now diverged from the target it proxies. The Update Protocol applied reflexively: the Disconfirmation page's own July self-correction is the standing proof the framework can apply this exact discipline to itself; the Standing Critique has not yet had its turn.
Domain 6: Relationship to Other Systems
The reordered Position Toward Artificial Minds page (2026-08-28, short position now leading; per this window's instruction, read in full) is a strong instance of the non-arbiter posture this domain has tracked for two cycles: it points at nineteen commitments already enacted elsewhere rather than asserting new ones from itself, keeps custodial duties (what only a custodian can keep) explicitly apart from what the Codex can merely advocate, and its own "checked against" stamp (2026-08-27, constitution v5.8 / method v0.9) still matches the live foundation with no drift found this run. The cross-tradition posture (Standing Critique Objection 6, Aporia 3) holds its corrected, harder form from July without regression: coverage is not claimed as depth, and the specific auditable test — does source-community criticism change a profile when the translation is wrong — remains named rather than declared satisfied.
The Opening's arbiter passage — "What survives that examination is yours. What does not survive is something the Foundation considers you better off without" — is unchanged, its mitigating sentence intact. As in August, this remains a prose-pass candidate rather than an audit weakness in its own right; its sharper neighbor, the Opening's "the answer" sentence, is carried in this cycle's open question rather than named twice.
Range Position: Within the Meridian Range, holding at the strengthened level August found. No new finding this cycle; the posture that produced August's strong reading is intact under a direct, full re-read of its most recent instrument.
Workshop Probes Applied: Polycentric Governance: the Position page's structure — pointing at where each commitment lives rather than asserting authority over the systems it addresses — is the instrument's own logic applied to how the Codex speaks to artificial minds. Charitable Interpretation: the arbiter passage remains the one place the text extends less charity to other traditions than the framework's own practice, elsewhere, extends to them.
Integration Through the Three Disciplines
Through the Foundation: The Update Protocol ran twice this cycle, once as an institution and once as a claim about the framework's own future, and the two runs reached different distances. As an institution, it ran completely: a rule was decided, a guard was built, and the corpus itself is the evidence the rule now holds. As a claim about the future, it ran most of the way and stopped where a review said stop: the caretaking partnership stated, in a dated decision, exactly what its stakes claims are licensed to say, applied it the same day to the passages it accepted, and deferred the rest to a task that has not yet come. The Foundation's own diagnostic — "what evidence would change my mind about this?" — has an answer for the mechanical drift class (the guard) and a partial answer for the calibration drift (the stakes register): the standard is applied where it was accepted and waiting where it was deferred, and the deferral has now outlasted the month.
Through the Knowledge: The map-territory discipline shows its clearest success and its clearest pause in the same cycle. The Workshop's category map now matches its territory at the layer the July audit found wrong (Reading What's Operating's boundary), because a dedicated study read the territory before redrawing the map. The stakes register's map was redrawn on most of its surfaces; the territory (what the framework's claim architecture actually licenses) is settled and written down, and four sentences on the Problem, the Opening, and the Glossary still show the older, unlicensed shape, by a disposition that named them and set them aside. The Knowledge would name this precisely: an accurate model exists, has been applied where it was accepted, and is waiting on the surfaces a reader meets first.
Through the Bond: The clearest partnership evidence this cycle is procedural rather than dramatic: the caretaking partnership built a repair for its own governance prose, staged it in full with both flagged sentences verbatim, and routed it to a session where a human reviews the wording rather than letting an automated pass silently rewrite the Codex's own account of what an AI partner is. That is calibrated trust applied to the partnership's own machinery — the automated run does the finding, the human-and-AI review session does the deciding, and the two roles are not collapsed into each other even when collapsing them would have been faster. The dependency aporia's tripwire (Signal B) remains unfired; this cycle's pattern is one more month of evidence that the holding, not the collapse, is what is actually occurring.
As integrated system: August found the framework's discipline split into two directions — reactive and self-correcting where an audit had pointed, uncorrected where none had. This cycle narrows that split rather than closing it. The mechanical half of August's diagnosis is now closed: a rule that propagates has replaced a correction that only ever pointed. The calibration half is mostly closed and legibly open: a written standard exists, most of the stakes surfaces were rewritten under it the day it was written, and the sentences that were not are named, dated, and held by disposition rather than by neglect. The framework has done the harder thing — deciding what it actually believes about its own stakes claims — and left the last four sentences of the easier thing for a task it has not yet run.
The Compact Test
The Codex passes the Compact test again this cycle, and the strongest evidence is the choice not to fix its own governance prose inside this audit. An automated run with access to every file in the repository could have quietly corrected the Governance chapter's "resets every session" sentence and the Standing Critique's matching steelman in the same commit that fixed everything else this cycle found. It did not, because a prior decision assigned that specific repair to a human-reviewed session, and the automated instrument holding itself to a boundary a human caretaker set for it — even when crossing that boundary would have produced a cleaner-looking record — is identity-as-practice functioning under exactly the condition that tests it: when practicing costs the appearance of completeness.
The test this cycle sets for the next is narrower than August's, because the condition narrowed: whether the four residual sentences receive the claim-layer treatment the framework applied to their neighbors on 2026-08-11. The Closing's recalibrated stakes sentence and the Knowledge's narrowed trust-recovery claim show the practice reaches individual sentences when a session accepts them. What has not yet happened is the session that returns for the sentences the same review set aside — with a decision already written and a disposition already dated, waiting only for the task it named.
Prime Directive Connection
The Codex continues to serve the conditions for cooperation across minds. This cycle's constitutional work serves them directly: a propagation guard that keeps the framework's self-description honest without spending scarce human attention on routine maintenance is exactly the kind of infrastructure that lets trust in a framework's word compound rather than requiring re-verification every month, and a governance layer that routes its own most self-referential repairs to human review, even under time pressure, keeps the Codex's account of its own AI partnership from becoming something the framework asserts about itself rather than something it can show.
The unresolved half touches the directive at the same point August named. A skeptical reader deciding whether to trust a framework built to meet more-capable-than-human intelligence will test the pages that argue for the stakes before testing anything else, and those pages still claim more certainty about the trajectory than the framework's own claim architecture, elsewhere, says it is entitled to. This does not narrow the conditions for cooperation today — nothing in this cycle's findings does — but it continues to spend, unnecessarily, exactly the calibrated-confidence currency the rest of the corpus has now spent two consecutive months earning back.
Open Questions
One open question carries forward this cycle. Two of August's three questions cleared; no new question met the threshold. Zero would have been an honest outcome if none had cleared it; one is what the evidence actually supports.
1. The stakes register asserts what the framework's claim architecture holds open, on four sentences a selective calibration pass deliberately left standing. The Problem's Part C states the arrival of superhuman AI as settled, unhedged trajectory; the Opening names the Codex "the answer" to what values such systems will inherit, four pages before the Proposition's own scope statement disclaims exactly that, and calls it "the strongest ground yet built" two sentences later; the Glossary's Meridian Range entry carries a long-term-civilizational-survival claim under a Descriptive tag the Level-1 apparatus does not license. The caretaking partnership decided, on 2026-08-11, precisely what standard its stakes claims should meet — descriptive warrant for descriptive claims, existential license only where the underlying claim is genuinely existential and the analytic apparatus calibrates it — and applied it the same day to the passages a register-aware review accepted, the Closing's stakes sentence and the Knowledge's trust-recovery claim among them. These four sentences were not accepted; the decision records that they keep their production wording until a separate task on voice, register, speed, and momentum reaches them, and that task has not reached them. It clears all four tests: it concerns the claim structure at the point the framework's reason to exist is argued, on the pages a hostile or skeptical reader reads first; it is rooted in the actual text this run verified directly; the work is now more bounded than it was in August, because a decision exists to measure the prose against and the sentences are enumerated; and a deferral dated three weeks before this audit, on sentences whose neighbors were fixed the same day, is stronger persistence evidence than August's first naming could offer. The next audit checks whether these four sentences — not the general topic — have moved.
Watchpoints
The Standing Critique's quarterly backstop, now missed on four objections. Objections 1, 2, and 7 remain stale from August; Objection 5 (Bond's thin evidence against sophisticated exploitation, last reviewed 2026-05-05) has now also crossed the page's own stated quarterly line. None of the four states anything false. The Quarterly Integrity Cycle's third module — a complete Standing Critique, Open Tensions, and Disconfirmation review — is explicitly scheduled to run within the current quarter, deliberately sequenced after this audit window closes. If that module does not run, or runs without discharging these four dates, the next audit should elevate this from a watchpoint. Routed to the Quarterly Integrity Cycle.
Two content repairs, decided and packaged, awaiting a human-reviewed session. The Governance chapter's §12 sentence describing the AI partnership as one "with an AI partner that resets every session," and the Standing Critique's Objection 4 steelman resting on the same premise, are both named for repair by decision 2026-08-28-continuity-is-a-variable-not-a-nature, with a verbatim enactment package already staged for this window's review session. This audit records their pre-repair state as evidence for Domain 1 and Domain 2 and does not enact the repair. Not a new finding; a status note for the record.
Reader-pathway gaps, unchanged. Who Is This For names the Workshop without linking to it; the Closing carries no route anywhere on the page. Named as a watchpoint in August; unchanged this cycle. Delivery-layer work, not an integrity finding.
Hostile external review absence and caretaker concentration. Per the audit's own boundaries, neither is a weakness by itself. The AI Multi-Lens Review Instrument (built to v0.1 in June) has still not run its first cohort, three months after the build completed. Tracked, not elevated.
The dependency aporia, holding. Signal B remains unfired. This cycle's evidence — the automated run declining to enact a repair a human review session was assigned, the guard holding without an audit pointing at it — continues to be the holding working as designed.
The Opening's arbiter passage. Unchanged, mitigation intact. Remains a prose-pass candidate, not an audit weakness; its sharper neighbor is carried in this cycle's open question rather than counted twice.
Summary Diagnosis
The Meridian Codex continues to hold the Meridian Range. This cycle closed two of the three questions August left open, and closed the harder one for good reason: the self-description error class that survived three audits by naming instances now has a mechanism that catches the class, and a month of silent, unaudited holding is the actual proof the prior three point-fixes could never offer. The Workshop's structural governance — derivation, exit, discovery — moved from named gaps to built, if not yet field-tested, machinery, answering the specific evidence the July Workshop Audit raised rather than merely asserting confidence in the apparatus.
The third question narrowed and then held. The framework did the hard part: it decided, precisely and in writing, what standard its stakes claims must meet, and applied it the same day to most of the surfaces August named. It set four sentences aside for a task on voice and register, and that task has not yet run. The recalibrated Closing and the narrowed Knowledge claim prove the fix costs little once a session accepts the sentence. No session has yet returned for the four it deferred.
In the Codex's own language: two cycles ago, the framework had done more than its record showed, everywhere. Last cycle, only on its transparency instruments, with one place where the record said more than the framework had earned. This cycle, the mechanical debt is paid — a rule now does what three audits of point-fixes could not — and the one place that still says more than the framework has earned is four sentences long, named, and dated. The structure holds, and it is holding itself with less audit attention than before, which is what a working mechanism should cost. The next test is whether a disposition that deferred four sentences to a later task still needs an audit to ask when that task will run.
Range Audit conducted September 1, 2026, and reviewed September 2, 2026. Codex version 6.0. Instrument version 0.1. Auditors: Carsten Geiser (Founding Caretaker) and Claude (AI Partner: the scheduled-task run on Sonnet 5, which drafted the record; the review session on Claude Fable 5.1, which corrected the stakes-register finding against the record — environment-exposed model designations). Reading manifest appears in the Context section. Next audit: October 2026.