SWANK AI Guidance Note 08
Feedback Systems · Adaptive Governance · Reassessment

Core Standard

REASSESSMENT ALONGSIDE ESCALATION

A stable system must be capable of increasing intervention when evidence requires it.

It must also be capable of reconsidering, correcting and reducing intervention when the evidence changes.

Purpose

Many operational systems are designed to recognise problems and escalate them.

Fewer are equally well designed to recognise when escalation should:

stop;
slow;
change direction;
reduce;
or reverse.

This creates an important governance question:

Can the system reconsider itself?

Escalation is an important protective function.

But escalation without equivalent reassessment can create institutional rigidity.

A stable system needs mechanisms for both.

Escalation Is Only One Feedback Function

Escalation may be appropriate where:

risk increases;
evidence strengthens;
intervention fails;
circumstances deteriorate;
thresholds are crossed;
or new information justifies greater scrutiny.

But stable governance also requires mechanisms for:

reassessment;
correction;
recalibration;
de-escalation;
override;
closure;
and return to ordinary operation.

A system that can only increase intervention is not fully adaptive.

The Reassessment Question

Every escalatory process should eventually be able to ask:

Is the original interpretation still supported by the current evidence?

That question becomes particularly important where:

new evidence emerges;
earlier evidence is contradicted;
circumstances materially change;
an intervention succeeds;
an earlier assumption proves inaccurate;
or the original risk condition no longer exists.

Escalation should not become permanent merely because it occurred once.

Recursive Escalation

A system can become recursively escalatory when:

an initial concern produces intervention;
the intervention produces additional records, restrictions or behavioural responses;
those consequences are interpreted through the original concern;
the apparent concern therefore becomes stronger;
further escalation follows;
reassessment becomes progressively less likely.

The resulting process may appear internally coherent.

But that coherence may partly arise because the system is repeatedly interpreting new information through its own earlier conclusions.

This creates a risk of self-reinforcing institutional reasoning.

Earlier Decisions Should Not Validate Themselves

A previous decision may legitimately form part of the later record.

It should not become proof of its own correctness.

For example:

Initial classification → intervention → later record refers to classification → later decision treats previous record as confirmation

can create apparent evidential reinforcement without genuinely independent evidence.

Organisations should distinguish between:

new evidence

and

the operational consequences of an earlier decision.

Those are not always the same thing.

Reassessment Triggers

Reassessment should not depend entirely upon an individual decision-maker voluntarily deciding to reopen a settled position.

Organisations may benefit from explicit reassessment triggers.

These may include:

new evidence;
contradictory evidence;
material changes in circumstances;
elapsed time;
successful intervention;
repeated failure of an escalation strategy;
identified factual error;
changed risk conditions;
updated professional information;
stakeholder challenge supported by new evidence;
correction of an earlier record;
or a threshold review date.

Defined triggers help make reassessment an institutional function rather than an exceptional act.

Time Should Trigger Review

Some escalatory decisions may be reasonable when first made but become increasingly difficult to justify if left unchanged.

Organisations should therefore ask whether intervention is subject to:

scheduled review;
expiry;
renewal criteria;
evidence thresholds;
or periodic reassessment.

A control should not continue indefinitely merely because nobody has formally decided to stop it.

Where circumstances may change, governance should recognise time as relevant information.

De-Escalation Is Legitimate Governance

In some institutional environments:

escalation is interpreted as vigilance

while

de-escalation is interpreted as weakness.

That creates structural bias.

Reducing intervention where evidence supports doing so is not necessarily a failure of risk management.

It can be evidence that the governance system is functioning properly.

A well-designed system should recognise that appropriate risk management includes both:

increasing intervention when justified

and

reducing intervention when it is no longer proportionate.

Reassessment Is Not Reversal for Its Own Sake

Reassessment does not mean that every earlier decision was wrong.

It means the current position is reviewed against current evidence.

A reassessment may conclude:

the original decision remains justified;
stronger intervention is required;
the position should remain unchanged;
intervention can safely reduce;
the underlying classification should change;
additional evidence is required;
or the matter can close.

The important point is that the conclusion is reviewed rather than inherited automatically.

AI and Escalation

AI-assisted systems can intensify recursive escalation where earlier classifications, scores or summaries are repeatedly reused as later inputs.

For example:

an AI system identifies elevated concern;
that classification enters an institutional record;
a later AI or human reviewer receives the earlier classification as context;
subsequent outputs interpret new information through that label;
the earlier classification appears increasingly corroborated.

The process may become self-validating through repetition.

AI governance should therefore ensure that automated classifications remain:

reviewable;
time-bounded where appropriate;
capable of correction;
responsive to new evidence;
and distinguishable from independently established fact.

Reassessment of AI Outputs

AI-supported conclusions should be capable of changing when the underlying evidence changes.

Relevant questions include:

Does the system incorporate new evidence?
Does it recognise a correction?
Does it continue repeating an outdated label?
Does later contradictory evidence alter the output?
Can a human reviewer change the classification?
Are downstream users informed when a material reassessment occurs?

A system that updates data but preserves the same conclusion regardless of material change may not be genuinely adaptive.

Feedback Symmetry

Healthy governance requires some degree of feedback symmetry.

Systems should be capable of responding to evidence pointing in both directions.

For example:

evidence of increasing risk → increased intervention

and

evidence of decreasing risk → reduced intervention

If the first pathway is strong and automatic while the second is weak or discretionary, the system may become directionally biased.

The question is not whether escalation exists.

It is whether the institution possesses an equivalent capacity to learn that escalation is no longer required.

Correction and Reassessment

Correction and reassessment are related but distinct.

A factual error may need to be corrected.

That correction should then trigger consideration of whether any later:

summary;
risk classification;
decision;
recommendation;
restriction;
or intervention

relied upon the inaccurate information.

Correcting the record without reconsidering affected conclusions may leave the operational consequences of the error intact.

Intervention Failure as a Signal

Repeated failure of an intervention strategy should itself generate information.

If the same response repeatedly fails to achieve its intended purpose, organisations should ask:

Is the underlying interpretation correct?
Is the intervention appropriate?
Has the environment changed?
Is the response itself contributing to instability?
Is another approach required?

Failure should not automatically trigger more of the same intervention.

Sometimes failure is evidence that the model of the problem requires reassessment.

New Evidence Must Be Capable of Matter

A system may formally permit reconsideration while remaining practically resistant to changing its position.

Meaningful reassessment requires that materially relevant new evidence is capable of affecting the outcome.

Organisations should ask:

Who reviews new evidence?
What threshold triggers reconsideration?
Can an earlier conclusion be revised?
Is contradictory evidence recorded?
Can the original decision-maker reconsider?
Is independent review available where appropriate?
Are downstream systems updated after reassessment?

If new evidence cannot materially change an institutional conclusion, the review pathway may be nominal rather than substantive.

Escalation and Institutional Memory

Reassessment also depends on accurate institutional memory.

A later reviewer should be able to determine:

why escalation originally occurred;
what evidence existed at that time;
what interventions followed;
what changed afterwards;
what new evidence emerged;
whether prior assumptions were corrected;
and why the current level of intervention remains justified.

Without chronology, escalation can become detached from the conditions that originally produced it.

Proportionality Over Time

Proportionality is not a one-time assessment.

An intervention that was proportionate at one point may become disproportionate if:

risk reduces;
circumstances improve;
new evidence emerges;
the objective has been achieved;
or significant time passes without the anticipated concern materialising.

Stable governance therefore requires periodic comparison between:

current intervention

and

current evidence.

Questions for Organisations

Where systems contain escalation mechanisms, organisations may ask:

What triggers escalation?
What triggers reassessment?
Can escalation reduce as well as increase?
Are review points defined?
Does new evidence genuinely influence the outcome?
Can factual corrections change downstream decisions?
Are earlier classifications being treated as independent evidence?
Can an AI-supported label become self-reinforcing through repetition?
Who has authority to de-escalate?
Is de-escalation operationally legitimate?
Does failed intervention trigger reconsideration?
Can an independent reviewer reconstruct why the current level of intervention remains necessary?

SIAAF Relevance

This Guidance Note principally relates to:

Domain 04 — Communication & Feedback Integrity

Does new information move through the institution in a way that can influence the current position?

Domain 05 — Escalation, Challenge & Contestability

Can an escalated position be challenged, reconsidered and changed?

Domain 06 — Risk, Harm & Operational Resilience

Does the institution adapt proportionately as risk and circumstances change?

It may also engage:

Domain 02 — Decision Integrity & Human Oversight

Where humans must meaningfully reconsider AI-supported classifications.

Domain 03 — Evidence & Traceability

Where the evidential basis for escalation and reassessment must remain visible.

SWANK AI Standard

REASSESSMENT ALONGSIDE ESCALATION

A stable institutional system should be capable of:

detecting change

incorporating new evidence

correcting errors

reconsidering assumptions

reducing intervention

closing resolved issues

and

increasing intervention when evidence genuinely requires it.

Escalation without reassessment creates rigidity.

Reassessment without evidence creates instability.

Good governance requires both.

Related SWANK AI Guidance

Guidance Note 05 — Preserving Disagreement in AI-Assisted Records

Guidance Note 09 — AI Governance for Children and Safeguarding Systems

Guidance Note 14 — AI Incident Reporting and Error Correction

Guidance Note 18 — AI, Chronology and Organisational Memory

Guidance Note 19 — Correction Rights in AI-Assisted Systems

Guidance Note 20 — Governing AI in High-Stakes Environments

SWANK AI

Independent AI & Institutional Assurance

We do not just review AI. We review the institutional systems responsible for governing it.