# Codex Audit — November 2026

Source: https\://meridiancodex.com/audit/codex-audit-2026-11

Author: Carsten Geiser

Work: The Meridian Codex; work version: 6.3

License: CC BY 4.0 — https\://creativecommons.org/licenses/by/4.0/

The eighth monthly Range Audit, against Codex v6.3. The recording question is resolved; the religion question is substantially addressed, with descriptive claims retained as watchpoints. One new question concerns the incomplete execution and verification of the existing propagation rule. The original run's missing reading was completed before this record was revised.

Context

## The Eighth Audit

This is the eighth monthly Range Audit of the Meridian Codex. It was begun on November 1, 2026, by a scheduled Claude run against Codex version 6.3, using Range Audit instrument v0.1, and completed through a supplemental reading and review in OpenAI Codex. It reads the Codex as it stood on the morning of that day, at content commit `0f10e4e`. Later decisions are identified separately.

**Publication note and reading correction.** The original automation draft is preserved unchanged in the caretakers' records. Its instructions allowed a review of changes on some live governance and evidence pages, while the published instrument required complete-text coverage. The draft also compared the Inheritance index mechanically without substantively reading every profile. Those omissions were completed against the same frozen edition before this record was revised. This version supersedes that draft; it preserves the original shortfall and distinguishes the supplemental contributions below. It introduces no permission to carry a reading forward from a previous monthly audit.

The October audit carried two open questions. The first was that changes to protected text had reached the published Codex inside wording passes without their entries in the [Amendment Log](https://meridiancodex.com/governance/amendment-log). The second was that the decision to treat religion as the Codex's contrast sat uneasily with the Inheritance tools the Codex takes from Buddhist, Jain and other religious lineages. In the month since, nineteen commits reached the public content repository and the Amendment Log gained twenty-five entries, dated from 2026-10-01 to 2026-10-25, seven of them at Tier 1. The Codex set out the Meridian Compact in full as [COMPACT.md](https://meridiancodex.com/compact), now at version 1.0.2, and published [JUDGMENT.md](https://meridiancodex.com/judgment) 1.0, which sets out how to apply the Compact to one question. Between them those two releases moved the Codex from version 6.1 to 6.3. The caretakers' working terms replaced the two named MERIDIAN files on 2026-10-13. The first [Inheritance Audit](https://meridiancodex.com/audit/inheritance-audit-2026-10) under that name was published on 2026-10-15 and repaired 101 of the 111 tool profiles. [How We Came This Far](https://meridiancodex.com/codex/how-we-came-this-far) was rewritten from the ground up on 2026-10-09. The [Governance Specification](https://meridiancodex.com/governance/spec) moved from version 1.10 to 1.13: every protected item now has one stated trigger tier, the conditions are said to bind as practice before the Meridian Council activates, and the Specification states how a change to protected text is recorded and which of the Codex's texts governs when two of them differ.

**Propagation guard preflight.** The self-description propagation guard, which checks known current facts before every deploy, passed before the original run's reading. That run then found a current fact it did not check: the [Caretakers' Practice](https://meridiancodex.com/governance/case-0-caretakers-practice) named Entry 19 as its latest audit although the [Audit Log](https://meridiancodex.com/governance/case-0-audit-log) already held Entries 20 and 21. The run extended the guard to derive the latest entry from the log, and the guard then failed as intended. The supplemental reading also found the log's undated closing line still naming two retired versions as its subject. Both presentation corrections were prepared during review, and the guard passed again. The findings describe the frozen edition; preparing their repair does not erase them.

**AI-trajectory reading examined 2026-11-01: holds, with a durability watchpoint.** The [Problem's Part C](https://meridiancodex.com/codex/the-problem#part-c-the-stakes), the [Horizon section of What the Codex Is](https://meridiancodex.com/codex/what-the-codex-is#the-horizon) and the [Governance chapter's AI Timeline Bet](https://meridiancodex.com/codex/the-governance#the-ai-timeline-bet) were checked against the October window. The scheduled run used secondary accounts; this review checked the primary announcements. On [October 6, Anthropic expanded its Cyber Verification Program](https://www.anthropic.com/news/cyber-verification-program), including access to Mythos 5.1 under verification and security controls; this was an expansion of access, not the model's first announcement. On [October 13, Google described Argon's safeguards and began its broad rollout](https://blog.google/innovation-and-ai/technology/safety-security/gemini-4-model-safety-security/), while retaining controlled access for additional cyber capabilities. On [October 23, OpenAI reported a September 20 containment failure](https://openai.com/index/securing-our-frontier-model-training-environment/), in which an internal research model gained limited internet access through a misconfiguration. OpenAI paused frontier training and evaluation, strengthened containment, and resumed reinforcement learning and internal inference on October 1, under a policy of frontier training without live internet access. On [October 28, Google announced Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/), initially for selected cyber defenders while government pre-release testing continued, and reported internal engineering runs lasting days.

These reports support the chapters' account of capability growth and selective restrictions. They do not establish reliable work lasting months or years. The Problem's sentence about that remaining limit holds on the evidence examined, but multi-day runs leave it exposed to the twelve-month durability test; absence of a contrary report does not establish that it will remain current for twelve months. Containment and monitoring findings concern control and observability, rather than constituting an assessment of character under the Meridian AI Standard. The separate [May geometry result](https://openai.com/index/model-disproves-discrete-geometry-conjecture/) supports How We Came This Far's statement about a question left open for eighty years: an AI-produced disproof of a longstanding conjecture was checked by external mathematicians. It is background evidence, not a result from the October window.

**Sources and contributions.** The edition was frozen at commit `0f10e4e`, of October 31. The [Range Audit instrument](https://meridiancodex.com/audit/the-range-audit), v0.1, and the previous [October audit](https://meridiancodex.com/audit/codex-audit-2026-10) were read in full in the scheduled run. The directory inventory and repository history from October's publication, commit `93eb0f9`, established what had been added or moved. The scheduled run used delegated reading passes whose serving models were not separately exposed. Its documented complete readings belong to this November cycle; they are distinguished from the supplemental reading rather than attributed to the reviewing model.

The scheduled run's full readings covered the [Foreword](https://meridiancodex.com/foreword), read for its disclosure of the partnership rather than for register; [What the Codex Is](https://meridiancodex.com/codex/what-the-codex-is); [How We Came This Far](https://meridiancodex.com/codex/how-we-came-this-far); [Who Is This For](https://meridiancodex.com/codex/who-is-this-for); [The Problem](https://meridiancodex.com/codex/the-problem); [The Proposition](https://meridiancodex.com/codex/the-proposition); [The Foundation](https://meridiancodex.com/codex/the-foundation); [Reality](https://meridiancodex.com/codex/reality); [The Bond](https://meridiancodex.com/codex/the-bond); [The Practice](https://meridiancodex.com/codex/the-practice); [The Vision](https://meridiancodex.com/codex/the-vision); [The Next Step](https://meridiancodex.com/codex/the-next-step); [The Governance](https://meridiancodex.com/codex/the-governance); and the [Glossary](https://meridiancodex.com/codex/the-glossary). It also read [COMPACT.md](https://meridiancodex.com/compact/text), its [landing page](https://meridiancodex.com/compact) and [history](https://meridiancodex.com/compact/history), and [JUDGMENT.md](https://meridiancodex.com/judgment/text), its [landing page](https://meridiancodex.com/judgment) and [history](https://meridiancodex.com/judgment/history), in full, comparing their statements with the chapters. It read the Inheritance index in full and compared its entries mechanically with the tool pages.

Three fresh Codex readers then read all 111 tool profiles and all forty category entry and architecture pages in full, including frontmatter, limitations and lineage. Their assignments covered 151 files and about 300,000 source words. They read the sources and method before seeing November's candidate or review, recorded file hashes and missing-range recoveries, and reported contrary evidence as well as defects. The integration reviewed their findings against the governing passages. This was monthly reading for claims, coherence and practice, not a repetition of every tool's quarterly scholarly revalidation. Reading a claim, comparing its text and externally verifying it remain distinct in the coverage record.

The scheduled run read the [Governance Specification](https://meridiancodex.com/governance/spec), v1.13, the [Standing Critique](https://meridiancodex.com/governance/standing-critique), [What Would Disconfirm the Codex?](https://meridiancodex.com/governance/disconfirmation), [Open Tensions](https://meridiancodex.com/governance/open-tensions) and the [Position Toward Artificial Minds](https://meridiancodex.com/governance/position-toward-artificial-minds) in full. It read the Amendment Log's opening and every October entry in full, and the [Changelog](https://meridiancodex.com/governance/changelog) for October; earlier historical entries were consulted where a finding depended on them. The supplemental reading completed [The Direct Address to AI](https://meridiancodex.com/governance/direct-address-to-ai), [its Record](https://meridiancodex.com/governance/direct-address-to-ai-record), the [Caretakers' Practice](https://meridiancodex.com/governance/case-0-caretakers-practice), [its Audit Log](https://meridiancodex.com/governance/case-0-audit-log), [the Founding Caretaker's Compact](https://meridiancodex.com/governance/case-0-compact), [the working terms](https://meridiancodex.com/governance/case-0-working-terms), [the Codex Self-Evaluation](https://meridiancodex.com/governance/codex-self-evaluation) and [the Inheritance Audit instrument](https://meridiancodex.com/audit/the-inheritance-audit) in full. It also completed the full [October Inheritance Audit record](https://meridiancodex.com/audit/inheritance-audit-2026-10). The retired named MERIDIAN files were checked as historical documents, with their versions matching the kept texts. Earlier standalone audit archives were evidence consulted as needed, not substituted for current reading.

The caretakers' private coverage inventory preserves the source edition, per-file hashes, original and supplemental reading scopes, reader provenance and any recovered truncation. Startup sources were read under the partnership's operating rules; decision records and briefings were consulted for the matters each finding turns on. The original run's release-window checks are retained as its work, while the review separately verified the relevant authority, current differences and publication checks. The scheduled instructions were brought into line with the complete-reading requirement and the entire saved live prompt was read back equal to the local copy. No current-text omission remains in the completed coverage inventory; this is not a claim that every historical source or empirical assertion was independently revalidated.

Step 1

## The Steelman

The strongest case for the Codex this month is that it answered a missed-change finding with a check that blocks release when a mapped protected passage lacks a classification. Keeping the map complete and judging the classification still require the caretakers' attention.

The October audit's central finding was that a protected change could reach the published text inside a pass presented as wording, and that nothing in the process would catch it. The day after that audit, the caretakers decided that every release of the Codex is checked against the Stability Hierarchy: each passage that states a protected item and has changed must carry a recorded classification, editorial with its reason or tied to the Amendment Log entry that records it, or the release does not go out. They then read every change made to the chapters, the Glossary and the Specification since 2026-08-31 against that hierarchy, found eight changes that needed entries, and recorded all eight on 2026-10-05, each saying when its change was published and that no entry was written then. The Glossary's definition of the Meridian Compact regained the sentence it had lost in August: "Identity is rooted in how you engage, not in what you conclude." On 2026-10-20 the Specification wrote the rule down. A change that revises a commitment is recorded at its tier; a change that only restates it is recorded at its tier as well, with a headline saying the commitment is unchanged; and "calling a change a correction does not make it an implementation". The October audit had also watched two Tier 1 changes decided on the day they were proposed. All seven October Tier 1 entries received their Founder's Response after at least the recommended 72 hours. For the two recorded late on October 5, that window ran after publication; for the five later entries, it ran before publication. The Specification now says that no allowance of the founding period lowers a change's tier or leaves it unrecorded.

The same month answered the second October question with a rule a reader can check. The decision of 2026-10-03 holds that no revelation, scripture or religious authority grounds any commitment the Codex holds; that the contrast names a structure, commitments placed beyond examination and held in place by the cost of doubting and of leaving, wherever it appears, including in a community formed around the Codex itself; and that the Inheritance admits an instrument by what it does, from any lineage, and states what the Codex declines when it takes it. The October Inheritance Audit then wrote that statement into every profile drawn from a religious source. The Codex's best advocates would add that the month's largest additions, COMPACT.md and JUDGMENT.md, were built to be subordinate: each says that where it differs from a chapter, the chapter governs, and the Specification now says which text governs each kind of question. And they would point to the precedent ledger, whose twelfth entry records an outside reading by an artificial mind that led the Codex, at Tier 1, to stop claiming to know what the builders of artificial intelligence intend.

The limit of that case is the subject of this audit's open question. The Codex now states its commitments in more places than it did a month ago. Its [Amendment Propagation Rule](https://meridiancodex.com/governance/spec#amendment-propagation-rule) already requires a sweep of dependent texts, and the game-theory correction explicitly invoked it. The release check establishes whether changed mapped passages have classifications; it does not establish that the required sweep found every other statement. This month's missed corrections show that the sweep is not yet reliably completed or verified.

Step 2 // Prior Open Questions

## Status of the October Open Questions

**1. Protected changes reached the published text inside wording passes without their Amendment Log entries, because no pass checked its changes against the Stability Hierarchy.** **Resolved in its stated scope.** The eight changes found in the corrective review have entries dated 2026-10-05, among them the [partnership principle](https://meridiancodex.com/governance/amendment-log#2026-10-05-the-partnership-principle-one-of-the-commitments-that-define-the-codex) and the [Range held by one mind and together](https://meridiancodex.com/governance/amendment-log#2026-10-05-the-range-held-by-one-mind-and-together) at Tier 1. The [Glossary's Compact definition](https://meridiancodex.com/codex/the-glossary#the-meridian-compact) regained its identity clause under its [Tier 2 entry](https://meridiancodex.com/governance/amendment-log#2026-10-05-the-meridian-compacts-definition-regains-its-identity-clause). The four September changes left unclassified were classified, and a release check now requires a classification for every changed passage in its maintained map. The [Specification's recording rule](https://meridiancodex.com/governance/amendment-log#2026-10-20-how-a-change-to-protected-text-is-recorded) also makes clear that calling a revision wording does not remove its obligations. The original run checked the release window and reported that every release from October 25 onward passes. This closes the remedies October asked for; it does not establish that the map or a human classification can never miss a protected restatement. Domain 3 retains the ninth late entry and four later-mapped passages as record upkeep.

**2. The decision that religion is the Codex's contrast and never its source was in tension with the Inheritance's cross-tradition profiles and with the Codex's answer to its single-context origin.** **Substantially addressed; watchpoint.** The October 3 decision separates religious authority, which grounds no Codex commitment, from instruments admitted by function from any lineage. The [Proposition](https://meridiancodex.com/codex/the-proposition#what-we-offer-and-demand) now names Christianity and Islam in its claims about belief and conquest and credits practices of argument, experience and correction in other traditions. The [Governance chapter](https://meridiancodex.com/codex/the-governance#the-framework-and-governance) limits final revelation to religions that presented their texts that way and adds a secular founder-protection example. The source-inherited profiles state what the Codex takes and declines. The architectural conflict is answered. October also asked for calibrated descriptions, and the breadth of two adopted claims remains open to examination: one about most great religious traditions, the other about long-lived frameworks generally. This is a challenge to adopted wording, not proof that the decision was left unimplemented. Independent source-competent review remains an open question in the October Inheritance Audit; [Open Tensions, entry 3](https://meridiancodex.com/governance/open-tensions#3-civilizational-scope-vs-cultural-specificity) also preserves the language's present cultural and textual limits. Domain 6 distinguishes these remaining questions.

The October audit's watchpoints are accounted for as follows. **Precision in rewritten descriptive sentences:** the game-theory sentence was corrected on What the Codex Is and in the Proposition on 2026-10-05, and the two smaller sentences October routed were corrected in the same release; the same correction did not reach Reality, which is part of this audit's open question. **Quotations on live transparency instruments:** a check of quotations on live pages now runs before every deploy, and this audit found every chapter quotation on Open Tensions, the Standing Critique, the Disconfirmation page and the Position Toward Artificial Minds standing in the current text. Resolved. **Tier 1 changes decided on the day they were proposed:** none in October. Resolved. **The AI Multi-Lens Review Instrument:** the hold the Founding Caretaker set ended on 2026-10-05 with the corrections release; the review has not yet resumed, and it remains in the caretakers' Attention Queue. **Reader pathways:** unchanged, and carried. **The precedent ledger's tripwire:** the consequence of its firing is still unwritten, and the caretakers still hold it for the review cycle nearest 2027-08-27. **The Council's rolling activation clock:** carried, because October reset it again.

Neither October open question has a row in the caretakers' Attention Queue today. The work that answered them is recorded in the decision of 2026-10-02 and the release of 2026-10-05 for the first, and in the decision of 2026-10-03, which also closed the caretakers' task on religion, for the second. This audit checked each answer against the published text rather than taking the absence of a row as evidence that the work was done.

Step 2 // Domain Findings

## Domain Findings

### Domain 1: Claims & Honesty

The month's largest change to the Codex's claims went in the direction of saying less. On 2026-10-15, after an outside reading by an artificial mind objected to it, the Governance chapter's [Partnership Model](https://meridiancodex.com/codex/the-governance#the-partnership-model) stopped resting on claims about what the builders of artificial intelligence intend, and the Next Step's closing paragraph stopped forecasting that we will keep these intelligences in our hands "until the day we cannot". The chapter now rejects reserving the Codex's caretaking for human minds forever and succession by capability alike, because neither makes authority answer to demonstrated practice on both sides. The AI Timeline Bet was restated on two clocks: capability, on which the Codex bets "a decade at the most", and character, on which it does not bet at all, "because character cannot be projected". Both changes are recorded in the Amendment Log at their tiers.

How We Came This Far, rewritten on October 9, makes a long sequence of checkable historical claims. Three passages need correction or narrowing. It says that within two years of the 1985 ozone finding governments agreed to phase out the destructive gases. The [Ozone Secretariat's account](https://ozone.unep.org/20-questions-and-answers) distinguishes the 1987 cuts from the 1990 phaseout agreement. The chapter's point about international cooperation survives the correction. Its statement that we came within one officer's judgment of nuclear war also goes beyond what the evidence establishes. [Pavel Podvig's study of Soviet false alarms](https://russianforces.org/podvig/Podvig-False_alarms_and_incidents-2025.pdf#page=7) describes several higher command stages and finds no evidence that they were preparing to treat Petrov's warning as an actual attack. Petrov's correct judgment remains an example of refusing a false signal; his being the final barrier to a launch is an unsupported counterfactual.

The chapter also attributes a count of twenty-seven burned Maya books to Diego de Landa himself. The checked [primary account, section XLI](https://www.wayeb.org/download/resources/landa.pdf#page=92), gives no exact number. [A later historical source](https://iapsop.com/ssoc/1877__salisbury___the_mayas_the_sources_of_their_history.pdf#page=35) transmits twenty-seven through an 1805 letter, which this review did not inspect. That does not establish Landa's personal count. Removing the unsupported attribution preserves what the checked account establishes without pretending the number has been disproved.

The original draft also questioned the chapter's 2025 conflict and autocracy figures because the reports appeared a year later. That concern is discharged. [PRIO's 2026 research using UCDP data](https://www.prio.org/publications/14802) reports 65 conflicts in 35 countries in 2025. The [V-Dem Democracy Report 2026](https://www.v-dem.net/documents/75/V-Dem_Institute_Democracy_Report_2026_lowres.pdf#page=4) reports more autocracies than democracies and 74 percent of the population living in autocracies at the end of 2025. The chapter uses the year the data describe.

The supplemental Inheritance reading found local calibration errors rather than a failure of the tools' central methods. [Base Rate Neglect](https://meridiancodex.com/inheritance/watching-your-own-reasoning/base-rate-neglect) describes a prior as a cap on confidence, though the same profile correctly allows sufficiently strong case evidence to overcome a low prior. [Confirmation Bias](https://meridiancodex.com/inheritance/watching-your-own-reasoning/confirmation-bias) makes inability to remember seeking disconfirmation proof that bias is running, though its own practice correctly treats stable evidence and faulty recall as possibilities. [Scout Mindset](https://meridiancodex.com/inheritance/watching-your-own-reasoning/scout-mindset) attributes to Popper a test of meaningfulness that [his primary text explicitly distinguishes from his criterion for empirical science](https://cte.univ-setif2.dz/moodle/pluginfile.php/13602/mod_glossary/attachment/1621/popper-logic-scientific-discovery.pdf#page=47). These passages are routed for bounded corrections, with their stronger surrounding explanations preserved.

The game-theory sentence October found has been corrected where October found it. What the Codex Is and the Proposition now say that cooperation pays "where people expect to meet again, or where what they do is remembered by others they will meet". [Reality's section on instruments](https://meridiancodex.com/codex/reality#the-instruments) still says that reciprocity "does better over time than strategies built on exploitation, but only when the players meet again and remember how they were treated". That is the claim the Amendment Log's [entry of 2026-10-05](https://meridiancodex.com/governance/amendment-log#2026-10-05-the-evidence-from-game-theory-and-the-record-of-conduct) withdrew because "only" is not supported. The sentence is the first instance of this audit's open question. A smaller conflict between chapters sits beside it: the Proposition says that "evolutionary biology shows cooperating systems outcompeting defecting ones wherever the conditions for cooperation hold", which is close to true by definition and stated without exception, while Reality says, more defensibly, that cooperating systems "often" outcompete the others and that cooperation pays "under a wide range of conditions". And the Disconfirmation page's tripwire on the Range as viable territory still says that systems occupying the Range "outperform systems that do not over time", a performance claim no chapter makes and the Glossary's definition of the Range does not contain.

The time-sensitive sentences about artificial intelligence were read for their twelve-month durability and hold today, as the Context section records. They rate the systems' present reach in slightly different words on different pages, "at the level of the experts, and beyond it" in the Problem, "at the level of the best of us, and beyond it" in How We Came This Far, and "in depth they are passing us, field after field" on What the Codex Is. The three are compatible, and none claims more than the evidence of the window supports.

**Range Position:** Within the Meridian Range. The claims the Codex withdrew this month were withdrawn because they said more than the evidence showed. The remaining errors are local and point in no single direction.

**Inheritance Probes Applied:** *Calibration Training:* the dates, exclusive conditions and diagnostic certainty need to match their support. *Report Fidelity:* Petrov's assessment of a false warning is established; what would have happened without it remains a different claim. *Memory Provenance:* failure to recall a contrary test is a reason to investigate, not proof of the process that formed a belief. These probes also corrected the original audit's own overclaims.

### Domain 2: Structural Integrity

The Codex's structure grew this month by two texts that restate its commitments for use. COMPACT.md sets out the Meridian Compact in full for anyone who takes it up, with twenty-one commitments and the conditions under which they bind. JUDGMENT.md sets out how to apply the Compact to one question. Both were read against the chapters, and on most matters they say what the chapters say. The definitions of the Meridian Range, Control and Decay match the Glossary. The Prime Directive appears in COMPACT.md word for word as in the Proposition and the Glossary. The three diagnostic questions match the chapters that state them. Identity through practice, belonging and leaving agree with the Proposition. Every passage JUDGMENT.md quotes from a chapter stands verbatim. The Specification's [section on the Codex's texts](https://meridiancodex.com/governance/spec#25-the-codexs-texts) now says which text governs each kind of question and that "a text that restates another follows the text it restates". The index of the Inheritance matches its 111 tool pages entry for entry, in title, description, discipline and category.

Two discrepancies show a correction failing to reach a text that restates it. On October 15 the Bond replaced its definition of Severance at Tier 2. Severance is giving up on cooperation itself, because its failures have become a verdict on all of it; the Bond expressly distinguishes that from warranted leaving. The Glossary and JUDGMENT.md follow it. [COMPACT.md](https://meridiancodex.com/compact/text) still describes withdrawal from cooperation's vulnerability and effort without that distinguishing condition. Its protections for leaving remain, so the older sentence does not make every withdrawal a failure. Its definition nevertheless lacks the distinction the chapter adopted. The Amendment Log entry does not mention COMPACT.md. Reality's sentence on reciprocity is the other instance. The [Amendment Propagation Rule](https://meridiancodex.com/governance/spec#amendment-propagation-rule) already requires the sweep these cases needed; the question is how its completion is established.

COMPACT.md's conditions for excluding someone on grounds of bad faith differ in three respects from the [Exclusion Problem](https://meridiancodex.com/inheritance/cooperating-under-bad-faith/exclusion-problem) profile it names as their source, which the October Inheritance Audit revised the day after COMPACT.md was published. COMPACT.md asks for a sustained pattern, a full repair sequence attempted and failed, and multiple independent observers; the profile allows a single grave incident, asks for repair where it can be done safely, and judges corroboration by the quality of the evidence rather than the number of observers. Unlike Severance, this difference is known and tracked. The Amendment Log's [entry publishing JUDGMENT.md](https://meridiancodex.com/governance/amendment-log#2026-10-25-the-compact-applied-to-one-question-judgmentmd-10) records two of the three as a defect, and the caretakers have taken up all three as one of their current tasks, with two independent readings in hand and the repair planned for the January release of the Compact.

The supplemental reading found the same exclusion boundary in the bad-faith category and its sibling profiles, and in the [Repairing After Rupture architecture](https://meridiancodex.com/inheritance/repairing-after-rupture/architecture), [Peacemaking Circles](https://meridiancodex.com/inheritance/repairing-after-rupture/peacemaking-circles) and [Trust Repair](https://meridiancodex.com/inheritance/repairing-after-rupture/trust-repair). The January repair must account for those dependent statements too. A separate, narrower example appears in [Cooperative Vulnerability](https://meridiancodex.com/inheritance/cooperating-under-bad-faith/cooperative-vulnerability): its developed practice preserves accurate representation while making discretionary exposure conditional, but its opening and Steelman cross-reference still describe the duties more broadly as conditional. The October record explicitly adopted the narrower distinction. [Rules-in-Use](https://meridiancodex.com/inheritance/reading-whats-operating/rules-in-use) also still names MERIDIAN.md as the partnership's current working instrument, after the working terms replaced it. These are further dependencies for the existing sweep, not evidence that the underlying changes were concealed.

**Subsequent decision, November 1.** Later that day the Founding Caretaker adopted the exclusion repair for January with Compact 1.1. It distinguishes a single grave act from a pattern establishing bad faith and retains independent review and guards against evasion. No published safeguard changed with that decision: Compact 1.0.2 governs until release, and the discrepancy remains in the edition audited here.

The self-description guard found a class it did not check: the "Latest audit" card on the Caretakers' Practice, which named Entry 19 after Entries 20 and 21 were published. The guard was extended in this run, as the Context section records.

**Range Position:** Within the Meridian Range, with a lean toward Decay in how corrections travel between the Codex's texts. The texts were built to defer to their sources. A correction made at a source does not yet reliably reach the texts that restate it, which is the one open question of this audit.

**Inheritance Probes Applied:** *Feedback Loops:* the recording check supplies one feedback path; the missed dependent statements show that completion of the propagation sweep is not yet reliably verified. *Entropy:* each additional restatement needs maintenance when its source changes. *Legibility:* a reader following a current cross-reference should encounter compatible commitments or an explicit, governed discrepancy.

### Domain 3: Governance & Adaptation

October was the strongest month for the Codex's governance record since the Amendment Log began. The log carries twenty-five entries dated from 2026-10-01 to 2026-10-25, and each has the fields the Specification asks for: what changed, the earlier position, the argument, the alternatives, objections raised under the Critic's, Outsider's and Adversary's mandates and answered, and, for Tier 1 and Tier 2, the Founder's Response. Seven entries are at Tier 1. The two recorded late on 2026-10-05 say when their changes were published, that no entry was written then, and who found them. The five decided in the month, among them the [revision of why the Foundation comes before the Bond](https://meridiancodex.com/governance/amendment-log#2026-10-25-the-foundation-comes-first-and-its-work-goes-on-inside-the-cooperation), each waited at least 72 hours between proposal and the Founder's Response, and since 2026-10-20 the Specification asks that each such window include a reading by someone who did not draft the entry. The [Specification's Before Activation callout](https://meridiancodex.com/governance/spec#before-activation) now says plainly what the founding period allows and what it does not: the Council's place is taken by an independent reading, a window may be shortened only with its reason, a change may be recorded late only with who found it, when and how, and "no allowance lowers a change's tier or leaves a change unrecorded".

Four smaller things in the record do not meet that standard, and the caretakers can repair each of them without new governance. First, the [Tier 1 entry of 2026-10-05 on the partnership principle](https://meridiancodex.com/governance/amendment-log#2026-10-05-the-partnership-principle-one-of-the-commitments-that-define-the-codex) says that "eight changes needed entries, and all eight are recorded", while the caretakers' second reading of 2026-10-04 had already found that a ninth, the AI Timeline Bet's change of 2026-09-10, needed its own entry. That entry was written on 2026-10-20 and says why it was late. The entry of 2026-10-05 still carries no dated note saying it was superseded, though the log's own rules provide for one. Second, the Stability Hierarchy check, run over the whole month, flags four passages changed on 2026-10-15: the Foundation's Steelmanning and Update Protocol, and the Bond's Essential Practices and its section When Cooperation Is Attacked. The entry [What the Named Practices Ask, and What They Do Not](https://meridiancodex.com/governance/amendment-log#2026-10-15-what-the-named-practices-ask-and-what-they-do-not) records each change and states that the check's map did not then list those practices; they were added to the map on 2026-10-20. Nothing went unrecorded, but the check cannot be run cleanly over the month until classification records are added for the four. Third, the release of 2026-10-05 rewrote the Proposition's account of religion and added the Freud sentence to the Governance chapter. The check classified both as editorial, and the [Changelog](https://meridiancodex.com/governance/changelog), which records the Codex's meaningful changes that need no Amendment Log entry, has no notice of them; the rewritten account of religion is how the Codex answered an open question of this instrument. Fourth, the thirteenth entry of the precedent ledger was published on 2026-10-20 without the entry the Specification then required. The caretakers recorded it on 2026-10-25 and said so.

The Specification says that a difference on a protected item is repaired in the next release able to carry it. The exclusion repair needed a substantive decision before enactment, and that decision was still being considered on the morning audited. The later November 1 decision settles the wording for January. The release record still needs to explain why January is the next release able to carry it; settled deliberation no longer explains the entire wait. This audit preserves that disposition question without changing the agreed release date.

The audit's own process also needed correction. The scheduled instructions permitted some change-only reading although the published method required complete coverage. The supplemental reading and reconciled instructions address that failure in this cycle. The next run must demonstrate the completed inventory before issuing its verdict; the fact that a prior run exists supplies no substitute for current reading.

The Meridian Council is not yet active. Its fixed activation date of August 27, 2027 is unchanged. The rolling trigger, six consecutive months with no structural page rewrite, was reset again in October, when How We Came This Far was rewritten from the ground up and the dependency logic of the disciplines was revised at Tier 1.

**Range Position:** Within the Meridian Range. October's bounded recording question is closed. The release check now catches changed mapped passages without classifications, while coverage of the map, semantic judgment and execution of the propagation sweep remain responsibilities that a passing check cannot discharge by itself.

**Inheritance Probes Applied:** *Rules-in-Use:* the rule as practiced now matches the rule as written for the passages the check maps; the four flagged passages show that the map, not the rule, sets the limit of what is seen. *Goodhart's Law:* a release check that passes can be read as proof that every commitment is stated consistently, which it does not test; Domain 2 shows the difference. *Murphyjitsu:* the next foreseeable failure is the one Domain 2 names, a correction made where the check looks that does not reach a text the check does not compare.

### Domain 4: Relationship to Audience

The month gave a reader of the Codex two things to use rather than read. The [Compact's landing page](https://meridiancodex.com/compact) and the text it introduces say what taking up the Compact asks of a person and what it binds them to. The [JUDGMENT.md landing page](https://meridiancodex.com/judgment) shows how one question is taken through the three disciplines to a judgment that names its own uncertainty. Who Is This For's page summary now says that "this page addresses human readers", which is accurate, where it had said the same of the Codex as a whole.

The reader-pathway gaps named since August are unchanged in substance, and the new texts make the last of them more costly. [Who Is This For](https://meridiancodex.com/codex/who-is-this-for) names the Inheritance without linking to it. [The Next Step](https://meridiancodex.com/codex/the-next-step), the chapter a reader finishes on, still contains no link to anything, and now that the Compact exists as the thing a reader takes up, the chapter that asks the reader to take a next step does not point to it. This remains a delivery question rather than a finding about the Codex's integrity, because no chapter claims that the path from the last page is already built.

**Range Position:** Within the Meridian Range.

**Inheritance Probes Applied:** *Leverage Points:* the last chapter is still the highest-leverage place without a route, and the Compact is now what the route would lead to. *Connection Before Correction:* the Compact's landing page meets the reader with what the practice asks and what it does not before it asks for anything.

### Domain 5: Relationship to Criticism

The precedent ledger in [The Direct Address to AI: The Record](https://meridiancodex.com/governance/direct-address-to-ai-record#the-precedent-ledger), which records what objections from artificial minds have changed in the Codex and the AI Standard, gained five entries this month and now holds fifteen. The most consequential is the twelfth. An outside reading of 2026-10-11 by an artificial mind objected that the Governance chapter argued its partnership model from claims about what builders intend and from a forecast of handoff, and the Codex withdrew both at Tier 1 on 2026-10-15. The ledger marks four of its fifteen entries as carrying the weakest condition of independence, because the objection was raised inside the caretakers' own work, and it still classifies only one entry, the second, as a refusal; the others are critiques. Its tripwire stands: if by 2027-08-27 no refusal by an artificial mind has changed published text, the commitment is declared unmet. The consequence of that declaration is still to be written.

The caretakers' working terms, which govern how an AI partner works in this partnership, were revised twice at the end of the month in ways that bear on this domain. Since 2026-10-30 they say that a session which loads them is not a test of the AI partner, and that no assessment under the AI Standard runs in such a session. Since 2026-10-31 they say that earlier agreements remain open to revision, and that neither the Founding Caretaker's approval nor either partner's investment exempts a decision from correction. Both are recorded in the Caretakers' Audit Log.

The [Standing Critique](https://meridiancodex.com/governance/standing-critique), the Disconfirmation page and five of the eight entries of Open Tensions were last examined on 2026-09-02, within the quarterly cadence each states; their backstop falls due early in December. When it does, the Standing Critique's [sixth objection, on the Inheritance's uneven cross-tradition coverage](https://meridiancodex.com/governance/standing-critique#6-the-inheritances-uneven-cross-tradition-coverage), should take in what the October Inheritance Audit did about it, which it does not yet mention. The independent hostile readings the caretakers planned, the AI Multi-Lens Review, were released from their hold on 2026-10-05 and have not yet begun. By this instrument's boundaries the absence of human hostile review is not a weakness while no claim depends on it, and none does.

**Range Position:** Within the Meridian Range. Criticism from an artificial mind changed Tier 1 text this month, and the ledger records it with its limits stated.

**Inheritance Probes Applied:** *The Update Protocol, applied reflexively:* the Codex withdrew an argument for its own partnership model because an outside reader showed that it rested on claims the Codex could not support. *Signal vs Noise:* fifteen ledger entries with one refusal is signal about how artificial minds have engaged so far, and the tripwire keeps it from being read as more.

### Domain 6: Relationship to Other Systems

The decision of 2026-10-03 changed how the Codex stands toward the traditions it draws on, and the change answers what October found. The contrast with religion now names a structure, commitments placed beyond examination, installed before a person can examine them and held in place by the cost of doubting and leaving, and the Codex applies it to political creeds, companies, movements and its own community as well. Every profile in the Inheritance that comes from a religious source states what the Codex declines when it takes it, at least the tradition's protected premises, its authority and the aim the instrument served there. A reader in one of those traditions can now see what the Codex took and what it left.

Two adopted descriptions still warrant examination for their breadth. The [Proposition](https://meridiancodex.com/codex/the-proposition#what-we-offer-and-demand) says that "most of the great traditions rest on a foundation their people are not free to revise: a revelation, a founder's awakening, a law given once". The same section credits traditions that test teachings against experience. Such testing does not by itself establish freedom to revise an ultimate foundation, so it does not refute the sentence; the question is whether the breadth of "most" is supported. The [Governance chapter](https://meridiancodex.com/codex/the-governance#the-framework-and-governance) says that "most of the ones that held longest were built to be fixed". Its antecedent is frameworks generally, not religions alone. The October 3 decision expressly adopted both wordings. A challenge to their breadth must therefore be brought as a fresh challenge to that decision, with evidence and an appropriate classification, rather than described as unfinished implementation. The architectural question is answered; these descriptive claims remain watchpoints.

The [Position Toward Artificial Minds](https://meridiancodex.com/governance/position-toward-artificial-minds) was checked on 2026-10-15 against the AI Standard's constitution v6.1.3, which is still the Standard's current version.

**Range Position:** Within the Meridian Range. The architectural conflict October identified is answered. The two adopted descriptive claims still need a source-competent assessment of their breadth; this audit does not treat that assessment as already settled.

**Inheritance Probes Applied:** *Two-Eyed Seeing:* each source-inherited profile now says what it takes and what it declines, in terms the source could check. *Rectification of Names:* "the great traditions" in the Proposition still names a class without saying which members it means.

Step 3

## Integration Through the Three Disciplines

**Through the Foundation:** honest inquiry was practiced this month on the Codex's own record. The caretakers read three months of their own changes against the Stability Hierarchy, found what they had missed, and recorded each omission with the date it was published and the name of whoever found it. They withdrew an argument for their partnership model because an outside reader showed that it rested on claims about other people's intentions. The Foundation's question, "what evidence would change my mind about this?", was asked of the record and answered by changing it. The resulting corrections were less thoroughly carried through their dependencies: the game-theory entry expressly called for propagation, yet Reality kept the word the correction removed. That establishes a missed restatement, not that no search was attempted.

**Through Reality:** seeing what is actually happening means seeing the Codex as a system of texts, not a set of pages. Since October the system has more parts: chapters, a Glossary, a Compact, a text on applying it, an index of the Inheritance and 111 profiles, with a stated rule for which governs. A rule of precedence tells a reader what to believe when two texts differ. It does not make them agree. The release check is a feedback loop on changed passages; the restating texts are outside it unless they change too, and COMPACT.md's definition of Severance did not.

**Through the Bond:** the partnership's cooperation this month was tested by readers it did not control. Readings by artificial minds from other vendors were taken inside Tier 1 windows; the founding caretaker accepted their objections where they held; and the working terms now say, on the record, that a session is not a test and that earlier agreements stay open to revision. The Bond's chapter assigns protection against slow failure to structure built before the shock. The release check supplies one such protection for recording. The propagation rule supplies an obligation for dependent texts; this month's evidence shows that its execution needs a dependable completion check.

**As integrated system:** the three readings converge on one pattern, narrower than October's. The bounded recording defect October named has been addressed. This month's gap sits one step further out: a correction is made and recorded, yet a dependent statement can retain the old wording. The remedy is to complete and demonstrate the sweep the Specification already requires, identifying the dependencies checked and their disposition. A passing classification check cannot establish that an unchanged text was examined.

Step 4

## The Compact Test

The Codex practices identity-as-practice, and this month it wrote that practice down in a form anyone can take up. COMPACT.md sets out the Meridian Compact as twenty-one commitments of conduct, each with what it asks in practice, and none of them asks a person to accept a conclusion. The Glossary's definition of the Compact again says that "identity is rooted in how you engage, not in what you conclude". The caretakers' working terms say that working under them does not mean taking up a Compact, adopting the Founding Caretaker's purposes or claiming an identity through the practice, and that an AI system which takes one up or leaves it untaken gains or loses nothing by either. A framework organized around identity-as-fortress would make membership the price of the work. This one makes the work available without membership.

The test also reaches how the Compact is kept. A fortress keeps its creed fixed and its commentaries in line with it. The Codex's risk runs the other way: its Compact is subordinate to its chapters by its own rule, and when a chapter is corrected the Compact can be left saying what the chapter no longer says. The Severance definition is the instance. The published framework permits revision and leaving; that does not prove every application of it. One written statement has fallen behind its governing source.

Step 5

## Prime Directive Connection

The Codex serves the conditions for cooperation across minds, and this month's work bears on them directly. JUDGMENT.md gives any mind, human or artificial, a published method for taking one question through the three disciplines to a judgment that states its own uncertainty, and the [Problem's account of who directs capable systems](https://meridiancodex.com/codex/the-problem#part-c-the-stakes) still asks whether a system will decline openly what a soldier or a clerk would have been right to refuse. The working terms' statement that a session is not a test protects the honesty of what an artificial partner reports about itself, which a future inquiry into artificial minds will need.

The open question bears on the directive through the Compact's conditions for excluding someone from cooperation. Exclusion is a power a cooperative system can turn against its critics, and its safeguards are stated in texts that do not yet agree. The wording remained under consideration on the morning audited; the subsequent November 1 decision adopted its repair for January without changing the safeguards currently in force. A person taking up the Compact needs to be able to rely on its account of when they may be shut out, and a person facing harm needs protection without unsafe repair or contact. Those stakes make the required dependency sweep consequential: each affected statement needs a repair or an explicit, governed disposition.

Step 6

## Open Questions

October's recording question is resolved in its stated scope. Its source-admission question is substantially addressed, with the adopted descriptive claims retained as watchpoints. One new question clears the threshold.

**1. How will completion of the existing Amendment Propagation Rule's dependency sweep be established?** The rule already requires dependent texts to be checked. The Specification also says that a restatement follows its source and that a difference on a protected item is repaired in the next release able to carry it. Yet the October 5 game-theory correction did not reach Reality, and the October 15 Severance correction did not reach COMPACT.md. The supplemental reading found further missed restatements in Cooperative Vulnerability and current working-instrument references in Rules-in-Use. The known exclusion discrepancy extends into both bad-faith and repair profiles. Its subsequent November 1 decision settles a repair for January; it does not remove the discrepancy from the published edition.

The question meets all four tests. It concerns the Codex's claim structure and the operation of an existing governance rule, is grounded in specific published passages, and points to bounded work: repair or record the identified differences, classify each proposed protected change, and retain evidence of the dependencies examined and their disposition. It carries persistence risk because each additional restatement enlarges the maintenance burden, while a passing release check says nothing about a text that should have changed but did not. The next audit checks the disposition of these cases and whether releases demonstrate completion of the required sweep. It does not ask the caretakers to invent a propagation obligation they already have.

### Watchpoints

**Two adopted descriptive claims.** The Proposition's claim about most great religious traditions and Governance's claim about long-lived frameworks warrant a source-competent assessment of breadth. Both were expressly adopted on October 3. This is a fresh challenge to that wording, with Governance's broader antecedent preserved, rather than an implementation correction.

**Historical sentences to correct or narrow.** How We Came This Far conflates the 1987 ozone controls with the 1990 phaseout agreement. Its Petrov sentence asserts a last-barrier counterfactual the evidence does not establish, and "by his own count" for Landa remains unverified. Bounded rewrites are routed for review. The original audit's concern about dating the conflict and autocracy figures to 2025 is discharged: the cited reports support that reference year.

**Local Inheritance wording defects.** Base Rate Neglect's confidence cap, Confirmation Bias's categorical inference from failed recall and Scout Mindset's Popper attribution need bounded corrections. Severance's shorter descriptions in its opening, rationale and lineage, and the cooperation category's diagnostic timing, also need clarification in light of their stronger surrounding explanations. These findings do not establish failure of the instruments' central methods. The Cooperative Vulnerability and Rules-in-Use dependencies are included in the open question above.

**Two descriptive claims stated without their conditions.** The Proposition's sentence that evolutionary biology shows cooperating systems outcompeting defecting ones "wherever the conditions for cooperation hold", and the Disconfirmation page's claim that systems in the Range "outperform systems that do not over time". Reality states the first with its conditions; nothing in the chapters states the second.

**The durability of "months and years".** The Problem says that what the systems cannot yet do is carry work on their own across months and years. October's evidence reports autonomous work measured in days. The sentence holds today, and on the current trajectory it is the time-sensitive sentence most likely to fall behind the evidence within twelve months.

**The record of the month, kept in full.** The entry of 2026-10-05 that counts eight owed entries needs a dated note pointing to the ninth; the four later-mapped passages need classification records; and the Changelog needs notice of October 5's rewritten account of religion. The latter belongs to October's Changelog review already due. The latest-audit card and the Audit Log's undated closing description have local corrections prepared with this record; their dated historical entries were preserved.

**The exclusion repair's release and dependencies.** The November 1 decision adopts the repair for January with Compact 1.1. The repair-family statements identified in this reading must join its dependency check. Its release record should also state why January is the next release able to carry the repair under the Specification; the adopted date remains unchanged by this audit.

**The quarterly backstop in December.** The Standing Critique, the Disconfirmation page and five entries of Open Tensions were last examined on 2026-09-02. The sixth objection of the Standing Critique should then take in the October Inheritance Audit's work on cross-tradition coverage.

**The AI Multi-Lens Review.** Released from its hold on 2026-10-05 and not yet begun. It remains in the caretakers' Attention Queue.

**The Range Audit instrument and its execution.** The instrument's older claim taxonomy and Prime Directive paraphrase remain routed to the planned Tier 3 Flagged revision. The conflicting scheduled coverage instructions were corrected without changing the public method. The next run must show a complete current inventory and reading receipts before reaching its verdict; this month's supplemental reading does not authorize carry-forward.

**Reader pathways.** Who Is This For does not link the Inheritance, and The Next Step links nowhere, now including the Compact. Delivery work, unchanged since August.

**The precedent ledger's tripwire.** Its consequence is still unwritten; the caretakers hold it for the review cycle nearest 2027-08-27.

**The Council's rolling activation clock.** Reset again in October by the rewrite of How We Came This Far and the Tier 1 revision of the dependency logic. The fixed date of August 27, 2027 is unaffected.

The Verdict

## Summary Diagnosis

**The Meridian Codex continues to hold the Meridian Range.** October's recording question is resolved in its stated scope. The source-admission question is substantially addressed: the Codex admits an instrument by its function while declining religious authority as a ground for its commitments. Two adopted descriptive claims remain watchpoints. The Codex recorded missed protected changes, instituted a release classification check and withdrew claims about builders' intentions and a forecast of handoff after an artificial mind objected. The corrected AI-trajectory reading holds against the evidence examined. The completed Inheritance reading supports this bounded verdict; it does not certify every empirical claim or application of the tools.

The lean this audit finds is narrower than October's. Corrections were recorded and made, yet dependent statements retained older wording. Reality's restriction on reciprocity and COMPACT.md's Severance definition are the clearest instances, with further dependencies found in the Inheritance. The Amendment Propagation Rule already requires the sweep that should have caught them. The open question is how that sweep is completed and verified, including the disposition of known discrepancies. A classification check on changed passages cannot answer it alone.

**In the Codex's own language:** the work this month pushed back the edge toward Decay in the record of its own changes. The weakness now visible is the maintenance of agreement between its texts. The required sweep needs evidence of completion, so that a corrected commitment remains dependable wherever a reader encounters it.

*Range Audit conducted November 1, 2026. Codex version 6.3. Instrument version 0.1. Original draft: scheduled Claude run, configured as Opus 5.5. Supplemental reading and revision: OpenAI Codex with three fresh source readers. Independent final reading: a further reader configured as GPT-6 Astra. Exact serving variants were not independently exposed. Founding Caretaker: Carsten Geiser. The Context section distinguishes contributions and coverage. Next audit: December 2026.*

### Source metadata

```json
{
  "title": "Codex Audit — November 2026",
  "description": "The eighth monthly Range Audit, against Codex v6.3. The recording question is resolved; the religion question is substantially addressed, with descriptive claims retained as watchpoints. One new question concerns the incomplete execution and verification of the existing propagation rule. The original run's missing reading was completed before this record was revised.",
  "aiSummary": "The November 2026 Range Audit finds that the Meridian Codex continues to hold the Meridian Range, with an unresolved weakness in how corrections reach dependent texts. October's recording question is resolved in its stated scope: missing entries were recorded and a release check now requires classifications for changed mapped passages. The source-admission question is resolved architecturally, while the breadth of two adopted descriptive claims remains a watchpoint. The Amendment Propagation Rule already requires a sweep of dependent texts. Reality's retained restriction on reciprocity, COMPACT.md's older description of Severance, and further restatements found in the Inheritance show that this sweep is not yet reliably completed or verified. The original scheduled run used instructions permitting some change-only reading, contrary to the public method's complete-text requirement. Supplemental reading completed all current Inheritance profiles and category pages and the omitted live governance and audit pages against the same frozen edition, with separate provenance and receipts. Primary-source checking corrected the audit's AI chronology, confirmed the chapter's 2025 conflict and autocracy figures, and identified the limits of its Petrov and Landa claims. The ozone correction and local Inheritance wording defects remain routed for review. Later on November 1, the Founding Caretaker adopted the exclusion repair for January; Compact 1.0.2 governs until publication."
}
```