Status: DRAFT

Concept note based on reading passes completed 2026-08-08 over a public discussion of decision governance, set against the Charter's existing correction duties. The proposals in the final section are not adopted Charter text; re-verify time-sensitive source context before relying on it.

Representation error and the correction route¶

The question. The Charter's Rely gate lets an affected person challenge a determination and have reliance on it reopened, corrected, or withdrawn. But what happens when the determination was reached correctly and what is wrong is the picture the rules encode — the categories, the process rules, the operating envelope the mandate was written against? A publicly published account of exactly this failure appeared in late July 2026. This note checks it against what the Charter already says and keeps only what is new.

Short answer. Most of it the Charter already covers, and in one place more completely than the source does. Five things it does not name: representation error as a failure class, the case that fits no category, the second reading of an exception pattern, assumption-linked expiry of standing authority (from the completed thread), and entrenchment through correction (from the republished thread).

Source¶

Wojciech Zygmunt Kaleta (linkedin.com/in/wkaleta) calls his work Decision Engineering and the failure False Situational Awareness: confidence remains high while the situational picture loses accuracy. Sources were read on 2026-08-06/08. The first post's date is approximate, derived from a relative timestamp read on 2026-08-06; exact times below came from publication metadata or the comment feed.

Post	Read as	Content used here
Public Field Case (10-page attachment)	≈2026-07-30 ("1w", edited)	model/representation distinction; the correction chain; the category and exception questions
A decision can remain in force…	2026-08-03, 08:14 UTC (exact)	inherited authority; revalidation triggers
Public Field Case, republished (9-page attachment)	2026-08-07 (attribution comment 06:07 UTC, exact)	the 2026-08-07/08 comment thread: contraction on contradiction, compellable reassessment, escalation flattening, entrenchment through correction

He states the failure as a chain that must remain unbroken: "contradiction → preserved unresolved case → evidence about the representation → institutional standing → competent authority → binding reassessment → changed representation, authority or execution." Two of his framings are load-bearing for what follows: a complaint channel is "not necessarily a correction mechanism", and an exception log is "not necessarily a reopening mechanism".

He labels the artifact "a conceptual stress-test, not as a validated governance framework", credits thirteen named contributors on it, and states that attribution there recognises contribution without implying co-authorship or endorsement. That caveat travels with the material: nothing below should be read as a validated result of his, and no affiliation, endorsement, or joint work is claimed by this repository.

The Charter's own question in reply — who performs the reassessment when the body that would revalidate the authority is the body relying on it — was put publicly under the second post on 2026-08-06. Kaleta answered the same day with a distinction this note adopts: institutional standing is not epistemic standing. An organisation may hold the formal authority to reassess its own picture and still lack independence from the assumptions, incentives, and evidence chain that produced it. His test is therefore not who is authorised to review, but what gives a review "the capacity to genuinely disconfirm" the picture the institution still relies on; a reassessment that cannot reduce, suspend, or reopen the prior decision's authority is, in his phrase, "procedurally valid while remaining epistemically captive".

In that exchange, the Charter proposed the three structural conditions recorded below. Kaleta accepted that disconfirmation "cannot depend on individual courage. It has to be protected structurally": reviewers must be able to "access withheld evidence, alter the permission state and remain economically independent of the reviewed party". The 2026-08-07 republication has a boundary list that includes epistemic independence and credits Robert Schaub among eight contributors. The same attribution caveat applies: contribution acknowledged, no endorsement or joint work implied.

One further post is flagged here and deliberately not treated: The Autobiographer (≈2026-07-29), on organisational memory as the substrate AI inherits — selection, not falsification, is enough to corrupt it. Its thread adds two boundaries that belong to the evidence layer rather than the correction route: provenance is not epistemic completeness — a fully traceable archive can still exclude every warning and rejected alternative — and record generation is asymmetrical: approval produces an artifact by default, dissent and overrides often produce none. Both land where split-custody is already load-bearing — custody proves the integrity of what entered the record, not the completeness of what was allowed to enter — and are to be taken up there when that note is next worked, not here.

What the Charter already covers¶

Recorded so the overlapping parts are not re-imported as if new.

Rule-level correction and lineage. Charter Commitments, duty 8, already reaches past the individual decision: where a material change to an authority source, policy, rule set, evidence corpus, or verification logic could alter outcomes, its effects are assessed before deployment against prior cases; and where evidence, a source, or a rule is corrected or withdrawn, lineage identifies the affected decisions so reliance can be reopened, corrected, or withdrawn rather than left standing. This is a superset of the chain's final step, and it adds a lineage stage the chain does not have.
Authority that expires. Consequential-action baseline 3: the action mandate is specific, purpose-bound, time-limited, and revocable.
Self-approval barred at the model layer. Consequential-action baseline 4: the admissibility check runs in a component that does not rely on the acting model to approve its own request.
Exceptions that do not harden. Runtime gates POC §5: an exception is a visible state transition scoped to one action, tool, or turn, expiring back to baseline — never silent standing authority.
Pattern as a trigger. Same section: per-mandate cumulative counters and an escalation-pattern counter; crossing a ceiling escalates even where each action was individually admissible, and a recurring escalation or override pattern narrows the operating envelope pending re-authorization.
Triggers that change state, not alerts. Kaleta's in-thread test for any revalidation trigger — whether it "merely creates an alert, or actually changes the decision state by requiring reopening, narrowing permission or suspending continued reliance" — is one the Charter's existing trigger machinery passes where it has triggers at all: baseline 5's declared patterns trigger re-scoping, re-authorization, constraint, suspension, or withdrawal rather than a counter; duty 8's lineage reopens reliance rather than filing a report; duty 3 attaches material-change triggers and a validity period to the release claim itself.

What it does not name¶

1. Representation error as a failure class. The Charter Commitments do not name this failure class. Duty 8 treats rules as things that get changed and then requires the consequences to be traced; it never names the state in which the system is performing exactly as specified and the specification itself has stopped describing the world. The distinction is between a model that is wrong and an organisational picture — data, categories, process rules — that has gone stale while every component still passes its own checks. Nothing in the Charter is inconsistent with this, which is the problem: it is implied everywhere and owned nowhere, so no duty attaches to detecting it.

2. The case that fits no category. The Charter's runtime verdicts are allow, deny, or escalate, and its risk register carries unresolved findings. Neither is the same as preserving a case because no available category represents it adequately, and treating that preservation as evidence about the category set. Escalate routes an individual case to a human, who resolves it and closes it; the anomaly is consumed rather than retained. A category set that fails repeatedly generates a stream of individually-resolved escalations and no signal at all.

3. The second reading of an exception pattern. The POC's escalation-pattern counter reads a run of overrides one way — the actor is pressing against its envelope — and responds by narrowing the envelope pending re-authorization. The same run supports a second reading: the envelope is drawn in the wrong place, and the overrides are evidence about the rule rather than about the actor. The Charter currently implements only the first. Where the second reading is correct, narrowing the envelope is the exactly wrong response, and it will look like successful governance while making the mismatch worse.

4. Assumption-linked expiry of standing authority. (added 2026-08-08, from the completed thread) The thread's central device attaches revalidation to the grant itself, not the calendar: a mandate names the conditions it was granted under — his trigger list runs changed scope, new dependencies, altered permissions, supplier or model substitution, repeated exceptions, evidence outside the original operating envelope — and a material change to one reopens the grant, because governance needs "a mechanism that distinguishes preserved approval from current warrant", "continuity needs an expiry condition, not only a review ritual", and otherwise "absence of challenge becomes an unintended approval mechanism". The Charter has every layer around this but not this one: duty 3 puts material-change triggers on the release claim, duty 8 on the rule set and evidence corpus, baseline 5 on intervention patterns, and baseline 6 blocks a request that is itself defective — while a standing mandate whose world has changed still reads as valid at every gate between grant and expiry. The one existing instance is the POC mandate's pinned approved-model set, which treats one such assumption's violation as an authority defect; the general mechanism is unnamed.

5. Entrenchment through correction. (added 2026-08-08, from the republished thread) A correction that succeeds can install the next failure: the replacement representation acquires authority and, once in service, can become insulated from the evidence that would invalidate it — so each successful correction risks being the last one possible on that rule. The requirement is an anti-entrenchment condition familiar from constitutional practice, restated for runtime governance: a corrected rule set re-enters service under the same challenge, reopening, and expiry machinery that displaced its predecessor, and no correction may narrow the correction pathway itself. The test is whether the machinery that displaced the old representation can still displace the new one. Nothing in the Charter contradicts this, and nothing states it: duty 8 traces the consequences of a rule change but never says the changed rule set remains as reopenable as the one it replaced. For a charter whose distinctive claim is capture-resistance, that is the missing meta-commitment — entrenchment through correction is a capture route the current text does not bar.

What would follow¶

Not adopted text. Recorded as candidate work, in ascending order of cost.

Name the failure class. Add representation error to the Charter's vocabulary as a distinct thing from model error, and state that duty 8's re-check triggers include evidence that the rule set has stopped describing the assessed reality — not only that it changed.
Give the no-fit case somewhere to go. A verdict or record state for no category applies, distinct from escalate, that is retained and countable rather than closed on resolution. The cheapest version is a required field on an escalation disposition: whether the case was resolved within the category set or despite it.
Split the pattern counter. Where a pattern trigger fires, record which reading it supports before acting on it. Narrowing the envelope and reopening the rule are different responses to the same signal, and the POC currently cannot express the second. Whichever reading is recorded, the escalation carries at least one preserved case alongside the count: contradiction that travels upward only as a category, a metric, or an exception count arrives with the evidence flattened out of it, and the aggregation that makes a pattern visible is the same operation that destroys the context a reassessment needs.
Declare what a mandate assumes. Generalize the POC's approved-model set: a mandate carries the conditions it remains valid under, and a material change to a declared condition is an authority defect at the next gate check — suspension pending re-authorization by the accountable owner, as a state change, not an alert (the test above). Renewal asks whether the evidence is current, the constraints still hold, and "the named owner still accepts the consequence under today's conditions" — silence renews nothing. Two conditions bound the design. Automation removes the incidental human pauses where staleness used to surface, so reopening cannot be left to someone noticing by chance — the friction is rebuilt deliberately at the one point where it is still needed. And under a material contradiction the burden inverts: continued authority, not the challenge, is what has to be justified, and it contracts until revalidated so that execution cannot outrun evidence and review — the suspensive effect administrative procedure has long attached to serious challenges.
Close the propagation window. The same machinery needs one invariant. A revalidation trigger that has fired but not yet reached the execution layer leaves the original authority technically active on a stale basis: the decision was valid at T1, a material condition changed, and execution is still pending at T2. The minimum is that a material change can halt or narrow execution pending revalidation; for the POC that is a consistency requirement, not new machinery — the suspension or narrowing takes effect in the authority state the gates read before the next gate check, so a fired trigger and a passing gate can never both be true of the same action.
Keep the correction pathway corrigible. The counterpart of gap 5: one normative clause stating that a corrected rule set re-enters service under the same challenge, reopening, and expiry conditions as the one it replaced, and that no correction may reduce the reachability of the correction pathway itself. Cheap to state, expensive to mean — it binds the rulemaker role, which is exactly where entrenchment would otherwise occur.
Answer the standing question. All of the above still routes correction through the institution operating the system. The Charter's distinctive claim is that this is insufficient — Runtime AI Governance Gets When Right. The Harder Question Is Who Gets to Check? (published 2026-07-23) — and the POC records the same gap honestly: independent reviewer and remedy decider are the two roles it cannot simulate, and its challenge beat populates a routing obligation rather than a remedy. Naming representation error sharpens that gap rather than closing it, because the party best placed to notice a stale representation is often the party relying on it.

A criterion for epistemic standing. Capacity to disconfirm has to be structural, or it reduces to the reviewer's virtue. Three conditions, proposed in the exchange above and consistent with the Charter's existing independence rule: the reviewer can obtain evidence the institution would not volunteer; the finding changes the permission state rather than producing a report; and the reviewer does not depend on the reviewed party for appointment, funding, or renewal. The third is usually decisive — a reviewer who needs the relationship to continue can afford every finding except the disconfirming one. The Charter already states the economic half of this: the party that writes the method does not evaluate anyone, the assessor is paid a flat fee regardless of outcome, and the method-steward takes no per-product royalties from the systems it judges. What it does not yet state is the access half — what a reviewer may reach without the reviewed party's cooperation — which is where the split-custody design becomes load-bearing rather than optional. Under the republished case, answering a different commenter, Kaleta extended the same boundary a day later: evidence with standing still fails if the organisation has no "legitimate mechanism that can compel reassessment" once standing is established — "otherwise contradiction reaches authority, but authority can simply absorb it" — which matches the independent-reviewer role the routing section below already assigns: able to force reassessment without owning the outcome. The evidence-layer consequence: where contradiction is logged but cannot alter the governing state, the audit trail proves awareness, not control — so an assurance regime that accepts logs as proof of governance is certifying the wrong artifact.

After the verdict: what the Charter does not route¶

Three questions raised 2026-08-07, checked against the existing documents. One pattern underlies them: the Charter has a verdict vocabulary and no onward routing. Allow, deny, escalate, and fail-closed all describe what happens to the action. None describes what happens afterwards to the case, the rule, or the person.

This section reads the gaps against the normative Charter Commitments. The same three questions read against the two published runtime articles resolve differently, because the articles contain less than the Commitments do.

1. What follows "no basis"¶

Not stated anywhere. The consequential-action baseline ends at ambiguity or missing authority fails closed, and the POC invariant stops an approval from manufacturing the missing basis. Both are about stopping. Neither says what happens next — so from outside, "no basis" is indistinguishable from being ignored.

Four things should follow. Only the first exists today.

The action stops. Fail closed. (exists)
The case is preserved, not closed. The absence of a basis is recorded as a fact about the mandate and rule set, not as a failed request that disappears with the session — the same mechanism as the no-category state above.
It routes to whoever can create the missing basis. (partly exists, pattern-only) Baseline 5 already routes to re-scoping and re-authorization — but only on "declared patterns or thresholds", never on a single unresolved case. One no-basis stop is recorded and goes nowhere.
The affected person gets an answer and a clock. (absent) Baseline 7 requires a receipt saying what the record evidences, what it does not, and whether it was independently confirmed — all about the evidence. Nothing requires telling the person what stopped, what would have to change, who can change it, and when they will hear.

Items 3 and 4 also need the two roles the POC does not have (§9).

2. Escalation graded by consequence¶

Correction (2026-08-07). An earlier version of this section claimed the escalation verdict was undifferentiated and that four parameters needed adding. That was wrong, and it was wrong because the section was written from search results rather than from the text. Consequential-action baseline 5 already specifies all four: the human-intervention contract states the authorized role, competence and independence requirements, and an authorized substitute or no-substitute rule; the response bound and safe no-response default; the permitted dispositions; and that only a declared reversible fallback already within existing authority may proceed. It also states, verbatim, that "Routing follows the authority required, not mere human availability" — which is already the Charter's answer to the fixed-role-versus-most-senior question.

The real gap is one link, not four parameters. Baseline 2 records reversibility; baseline 5 never reads it. The intervention contract is defined per escalation path by whoever writes it, with no required severity taxonomy — so two paths with identical consequence classes may carry different response bounds and defaults, and nothing detects that. The three-way taxonomy that would key it (reversible / compensable / irreversible, with the rule that the later the category the stronger the gate) exists only in ambient agentic AI control, a working note, while the Charter Commitments link that note's technical control plane without binding the intervention parameters to its reversibility taxonomy.

Closing it needs no new machinery: require baseline 5's parameters to be set per reversibility class as recorded under baseline 2, and lift the taxonomy into the definitions block of Charter Commitments alongside the existing definition of responsibility map.

Severity selects the body, not the rank. The recipient is not one named reviewer but one of the separated roles, and which one depends on what would have to change. An operational error the accountable owner can fix stays inside the institution. A contested determination needs an independent reviewer — outside the deciding line, able to force reassessment without owning the outcome. A missing or expired basis is a rulemaker matter, because the mandate itself has to change. A consequence the institution cannot undo needs a remedy decider, because no amount of internal reassessment gives the affected person anything back. These are different kinds of authority, not rungs of one ladder: a rulemaker cannot grant a remedy, a remedy decider cannot amend a mandate, and an independent reviewer can do neither but can compel both to be reconsidered. The routing question is therefore not how serious is this but which kind of change would answer it — with severity deciding how far outside the institution that change has to be sourced. One consequence is worth stating as a rule: evidence can earn the standing to reopen a determination or a rule, but contradiction alone never authors the replacement — review annuls, and only the competent rulemaking authority re-decides. The role split enforces this by construction: the reviewer who compels reconsideration is not the rulemaker who rewrites.

3. Who checks the checkers¶

Already answered, for certification. The certification model sets out standard-setter → certifier → accreditor → peer review, states the chain in as many words ("who checks the checkers?" → the accreditor; "who checks them?" → peer review), and proposes riding ISO/IEC 17065, national accreditation bodies, and the IAF against ISO/IEC 17011 rather than reinventing it. The same chain appears in three published articles. No work needed here.

The gap is the runtime roles. That chain polices conformity assessors. It does not say who accredits an independent reviewer or a remedy decider inside an operating system — network-overview carries one conditional sentence and nothing further. They may not ride the same chain: a remedy decider exercises authority over an individual case, which resembles ombuds or dispute-resolution accreditation more than conformity assessment. Which family of precedent applies has not been checked.

Open questions¶

Is representation error a distinct duty, or a trigger condition inside duty 8? Adding a duty is expensive and the Charter's duty set is deliberately short.
Can a no category applies state be made countable without becoming a dumping ground for hard cases?
Who has standing to submit evidence about a representation, as opposed to evidence about their own case? The affected person usually sees one instance; the pattern is visible only to the operator, which is the party the evidence is about.
Does any existing regime already carry a representation-error duty under another name — regulatory model validation, safety-case revalidation, clinical-guideline review? Not checked.
Which assumptions can a mandate declare without turning every grant into a full risk assessment? The POC pins one (the model set); scope, supplier, and operating envelope are nameable in advance; data quality and operating conditions drift toward the representation-error problem above, where the party best placed to notice the change is the party relying on it.
The addendum's requirements have older homes — anti-entrenchment doctrine, the suspensive effect of challenges, uncertainty absorption in reporting hierarchies, detective-versus-corrective controls. Which canonical statements to cite when any of them becomes normative text has not been checked.