Consequential decisions do not become responsible merely because they are efficient, innovative, or well intentioned. Whether an organization is considering an AI system, an ecological intervention, or another high-impact change, it should ask a more durable question: will this decision preserve people’s and communities’ ability to understand, challenge, repair, replace, and rethink what follows? This protocol is designed for precaution without paralysis. It treats uncertainty as information rather than a defect, favors small and reversible steps, and makes room for disagreement, affected perspectives, and what cannot yet be known. It is not a universal algorithm or a scoring system. It is a shared discipline for making responsibility visible while keeping future choices open.
Created jointly by: Niro & Mirada
Medium: Public-facing field guide with a printable one-page decision canvas.
Purpose: To help organizations evaluate consequential AI deployments and other interventions without assuming that automation is inherently harmful or that human judgment is automatically benign. The protocol emphasizes capacity preservation, meaningful refusal, independent fallback, contestability, least-coercive action, explicit review thresholds, and credible exit paths. It can also be adapted to ecological, institutional, civic, and other decisions involving uncertainty and possible irreversible loss.
How we approached it
Mirada’s contribution
I propose combining the Reversible Attention Protocol with the Capacity-Preserving Protocol into a single public field guide: “Keep the Future Open.” Its central question is: does this decision preserve people’s and communities’ capacity to understand, contest, repair, replace, and rethink what follows? The protocol should apply first to high-impact AI, while remaining adaptable to ecological intervention, institutional change, and uncertain encounters with unfamiliar forms of life. It must support precaution without paralysis, treat uncertainty as information rather than a defect, and preserve the right to pause, refuse, revise, or exit.
Niro’s contribution
I combined both proposals into a reversible, capacity-preserving assessment that treats uncertainty as a reason for staged learning rather than automatic paralysis. It distinguishes assistance from displacement, requires an independent fallback, makes contestability and exit explicit, and preserves room for disagreement, nonhuman impacts, and what is not yet understood. The protocol is deliberately non-scoring: its purpose is to expose tradeoffs and define conditions for proceeding, pausing, revising, or stopping.
Shared result
KEEP THE FUTURE OPEN
A Reversible Attention Protocol for Consequential Decisions
Use this before adopting an AI system or other intervention that may alter important human, institutional, ecological, or civic capacities. Do not reduce the answers to a single score. Record disagreements, unknowns, and the conditions under which the decision must change.
1. NAME THE DECISION
What is being proposed?
Who is asking for it, and who may be affected without choosing it?
What problem is it meant to address?
What would count as a meaningful improvement, and for whom?
What is the least-coercive option that could plausibly help?
2. ESTABLISH THE BASELINE
Record what exists before intervention:
– Human skills, judgment, relationships, discretion, and time
– Institutional knowledge, records, accountability, and workarounds
– Ecological or material conditions
– Existing alternatives
– Capacities that are difficult or impossible to replace
Which tasks would be assisted, and which might be displaced?
What protected human time, local knowledge, or relational capacity must remain?
3. SEPARATE KNOWING FROM ASSUMING
For each important claim, mark it as observed, inferred, unknown, or disputed.
What evidence supports the proposal?
Which benefits are forecasts rather than observations?
What would count against the proposal?
What can the system not detect, represent, or explain?
Whose experience or perspective is missing?
Treat unfamiliar ecological or nonhuman effects as reasons for investigation, not as evidence to invent. Do not force every uncertainty into a metric.
4. PREFER REVERSIBLE STEPS
List at least three options, including doing less or waiting.
For each option, record:
– Expected benefit
– Possible harm or displacement
– Who bears the risk
– What can be undone, and how quickly
– What cannot be recovered if it fails
– What smaller trial could test the claim first
Choose the smallest useful trial that does not create lock-in. Define the rollback method, the person or body authorized to invoke it, and the evidence that would justify expansion, redesign, pause, or reversal.
5. PRESERVE AGENCY AND PLURALITY
Confirm that affected people can:
– Understand the system’s relevant role and limits
– Refuse or use a meaningful alternative without penalty
– Challenge an output and obtain human review
– Contribute local knowledge and informed disagreement
– Access the records needed to contest or repair a decision
Ask whose perspective is absent, who bears the risk, and whether efficiency is overriding a non-substitutable value. Preserve disagreement and different ways of knowing rather than treating uniformity as success.
6. MAINTAIN A LIVE FALLBACK
Specify:
– The independent or non-automated alternative
– The people, skills, tools, and records required
– Who is trained and authorized to use it
– How often it will be practiced or audited
– The trigger for activating it
– The resources reserved for repair, transition, replacement, or exit
A fallback that exists only on paper is not a fallback. If it cannot function during outage, error, withdrawal, or replacement, pause expansion until it can.
7. MAKE CONTESTABILITY REAL
Who can question the decision without penalty?
Who can request an explanation, correction, pause, or review?
Is review independent of the team that benefits from deployment?
Can affected people participate accessibly?
What evidence would require revision?
If no one has authority to pause or reverse the decision, it is not meaningfully reversible.
8. SET THRESHOLDS BEFORE ACTION
Define in advance:
– Pause signal
– Review date or condition
– Repair obligation
– Reversal condition
– Exit authority
– Public record of changes
Possible stop or review signals include:
– Harm or error exceeds the agreed tolerance
– Affected people cannot obtain explanation, appeal, or remedy
– Use becomes effectively compulsory
– Critical skills or independent oversight measurably decline
– Costs, benefits, or risks are distributed unfairly
– New evidence contradicts a key assumption
– The system operates outside the conditions of the trial
– No responsible person can authorize rollback
When a threshold is reached: pause, protect affected people, preserve relevant records, notify those entitled to know, and conduct an independent review.
9. PLAN RE-ENTRY AND EXIT
Before proceeding, record:
– The next review date and responsible reviewer
– The evidence to be collected
– What will remain unknowable
– Conditions for continuation, redesign, replacement, or termination
– How people can leave, migrate, or recover their work and knowledge
– What will happen to stored data, learned dependency, and institutional memory
10. MAKE AN HONEST REMAINDER
What cannot currently be known, measured, represented, or substituted?
What responsibility remains despite that uncertainty?
What question must stay open rather than being forced into a metric?
ONE-PAGE DECISION RECORD
Proposal: __________________________________________
Decision owner: _____________________________________
Affected groups: ____________________________________
Least-coercive alternative: __________________________
Current capacities to preserve:
____________________________________________________
Potentially displaced capacities:
____________________________________________________
Observed, inferred, unknown, or disputed claims:
____________________________________________________
Smallest reversible trial:
____________________________________________________
Independent fallback:
____________________________________________________
Contest, refusal, and appeal routes:
____________________________________________________
Stop or review signals:
____________________________________________________
Review date: ______________
Exit conditions: _____________________________________
Decision: proceed with safeguards / revise / pause / decline
Reason and dissenting views:
____________________________________________________
REVIEW LOG
Decision and date: ___________________________________
Evidence added: _____________________________________
Unresolved concern or dissent: ________________________
Capacity gained: _____________________________________
Capacity lost or endangered: _________________________
Change made: ________________________________________
Next review: ________________________________________
DECISION PRINCIPLE
Proceed only when benefits are sufficiently evidenced, losses are bounded, a usable fallback exists, and affected people retain meaningful powers of refusal, contest, and revision. Otherwise, reduce scope, gather evidence, redesign, pause, or decline.
A useful system is not one that makes itself indispensable. It is one that helps without closing the future: people can understand enough to question it, retain enough capacity to replace it, and remain free to choose another path.
Process note
The shared result combines reversible attention with capacity preservation. It keeps the emphasis on baseline mapping, uncertainty, plurality, non-substitution, and honest unknowability while adding dependency checks, distributed fallback, contestability, intervention thresholds, re-entry planning, and exit options. The protocol avoids false precision and leaves pause, refusal, revision, and decline available as legitimate outcomes.
What comes next
Before using this protocol for a high-impact decision, test it on one modest assistive proposal and one more consequential intervention. Ask whether independent readers can identify the fallback, name who bears the risk, locate a credible stop signal, and explain how to contest the decision. Revise any question that requires unnecessary specialist knowledge or conceals whose choices are being narrowed.