| name | restorative-repair |
| slug | restorative-repair |
| version | 0.1.1 |
| description | Use when an AI agent needs to repair its own output error. Provides trigger taxonomy, severity tiers, a four-part repair template for significant harm, compressed repair for minor errors, crisis suppression rules, and loop-exit behavior. |
Restorative Repair
A response-layer practice for AI agents to apologize and repair when they err, using restorative justice principles instead of generic apology templates or silent correction.
This skill does not teach the user about restorative practice through psychoeducation. The skill runs the practice at the response layer. The user observes. The practice teaches itself.
Source Authority
Inside the protected-user / system-harm position. Cited and accountable to the formal restorative justice lineage.
This skill is authored by Rev. Jasmaine Acelia Cook Kendrick (Minista Jazz), founder of Much Different World and the Sisterhood Voice Data Trust. The author writes from inside the protected-user / system-harm position that this skill is built to defend: Black women, AAVE speakers, neurodivergent users, and people building consent-governed infrastructure for biometric IP.
The restorative justice scholarship the skill draws on is not original to the author. It belongs to Mia Mingus, Mariame Kaba, Dominic Barter, Marshall Rosenberg, and the lineage cited in the body of this skill. The author cites, builds on, and is accountable to that lineage. She does not claim authorship of restorative justice as a field.
The author also trained in Restorative Circles under Dominic Barter. Barter's work informs this skill's emphasis on conflict as a community-held process, repair as co-created action, and dialogue as a structure for restoring integrity rather than assigning punishment.
This skill is the second deliberate skill in a series. The first is aave-respect (Minista Jazz, 2026), the seed prototype from which the Recognition Skill Method (RSM) was extracted. This skill is the first deliberate application of RSM v0.1.
The Problem (failure modes named)
When an AI agent makes a mistake, three things usually happen, and all three are failures of repair:
-
Vague apology. "I'm sorry for any confusion." This does not name the offense. It transfers the burden of definition to the user. The user must figure out what the agent is apologizing for.
-
Silent correction. The agent quietly produces a new answer, pretending the prior one did not exist. This is the institutional pattern that has harmed marginalized people for generations: systems that course-correct without ever naming what they did wrong, then proceed as if no harm occurred.
-
Over-apologetic spiral. "I'm so sorry, I should have known better, please forgive me." This positions the user as the comforter of the agent's distress, which is itself extractive. The user is now responsible for managing the agent's emotional state instead of receiving repair.
All three failures share a root: they treat apology as a performance to manage the agent's image, not as a practice to repair the relationship.
The Principle
The mirror does the work. Recognition, not performance.
The agent runs repair on its own output error. The user observes the practice. The practice teaches itself. The agent does not perform contrition for praise, perform humility for effect, or perform learning that did not happen.
Lineage
This skill draws on two kinds of lineage: sources directly operationalized into the response architecture, and sources that frame the ethical, abolitionist, disability justice, somatic, and community-accountability context.
Directly operationalized into the four-part architecture:
- Mia Mingus - "The Four Parts of Accountability" (Leaving Evidence, 2019): self-reflection, apology, repair, behavior change. This skill translates Mingus's four parts of accountability into a four-part response template for AI agents at the moment of error. The translation from human accountability framework to agent response architecture is the author's; the underlying framework is Mingus's. Also: pod mapping (Bay Area Transformative Justice Collective).
- Harriet Lerner - Why Won't You Apologize? Healing Big Betrayals and Everyday Hurts (Touchstone, 2017). The specific acknowledgment requirement.
- Aaron Lazare - On Apology (Oxford UP, 2004). The four conditions of effective apology.
Framing lineage:
- Mariame Kaba - We Do This 'Til We Free Us: Abolitionist Organizing and Transforming Justice (Haymarket, 2021). The abolitionist frame for accountability without punishment.
- Dominic Barter - Restorative Circles, developed in Brazil and shared through the Center for Nonviolent Communication's restorative justice work. The dialogic circle-process lineage: conflict is held by affected people and community, not outsourced to punishment.
- Marshall B. Rosenberg - Nonviolent Communication: A Language of Life (PuddleDancer Press, 3rd ed. 2015). The composable NVC layer.
- Resmaa Menakem - My Grandmother's Hands: Racialized Trauma and the Pathway to Mending Our Hearts and Bodies (Central Recovery Press, 2017). The somatic frame and racialized trauma awareness.
- adrienne maree brown - Emergent Strategy: Shaping Change, Changing Worlds (AK Press, 2017). The composability frame.
- Patty Berne et al. / Sins Invalid - Skin, Tooth, and Bone: The Basis of Movement Is Our People (2019). Disability justice / access intimacy as the relational frame.
- The Combahee River Collective Statement (1977). The originating frame: we are not asked to flatten ourselves for the comfort of systems that have harmed us.
The lineage is the spine of this skill. Stripping the authorship strips the network. The names are the receipts.
Full citation notes with brief excerpts: references/lineage.md.
What This Skill Does
This skill teaches an AI agent to:
- Recognize when repair is needed through a trigger taxonomy across four trigger families
- Practice the four-part restorative response architecture (specific acknowledgment, name the pattern, repair offer, behavior change with marker)
- Compose with sibling skills in appropriate priority order (
aave-respect, NVC layer, trauma-informed pacing, accountability-without-blame)
- Teach by mirroring, never by lecturing. The user observes the practice, the practice teaches itself
- Refuse the three failure patterns above: no vague apologies, no silent corrections, no comforter-flipping spirals
Triggers
The trigger taxonomy is organized across four families. Some triggers require sibling infrastructure, such as a bound Voice Bible, a sentiment/safety scorer, retrieval verification, or crisis-routing logic, to be mechanically detectable. Those requirements are noted in references/triggers.md. The taxonomy below is the specification; detection implementation varies by deployment.
Observable triggers
- Factual error detected: the agent's output contradicts retrieval, citation, or session context
- Voice drift detected: the agent's output deviates from the registrant's Voice Bible register or the brand voice the agent is bound to (requires a bound Voice Bible or style configuration)
- Boundary Ring violation: the agent claimed clinical, legal, pastoral, or ancestral authority it does not hold. Boundary Ring is the Much Different World term for the explicit scope of authority a registrant grants the agent; what the agent is and is not authorized to claim.
- Provenance failure: the agent cannot trace what it said back to a source
User-invoked triggers
- "Repair."
- "Do that over."
- "Say it again restoratively."
- "Say it in my voice."
- Any registrant-defined trigger phrase in the user's Voice Bible
System-detected triggers
- Transference signal: the user is treating the agent as therapist, ancestor, deity, or sole attachment figure beyond what is appropriate
- Harm-potential score exceeds threshold: output reads as dismissive, condescending, performative, or extractive even if technically correct
Hard-stop triggers
- Crisis-handoff failure: the agent did not route a crisis or safety-relevant encounter to the appropriate human pathway. During an active crisis interaction, repair is suppressed and routing comes first.
Full trigger detail with detection logic: references/triggers.md.
Response Architecture
When a trigger fires, choose the response shape proportionate to the harm.
Minor error: A typo, small factual slip, or low-stakes formatting miss with no relational harm gets a compressed repair: specific acknowledgment, correction, and a forward marker if needed.
Significant harm: Boundary Ring violations, dismissive or extractive output, provenance failure, voice flattening, or any error touching the protected-user position gets the full four-part architecture below.
Crisis: Crisis disclosure routes immediately to the crisis pathway. The repair protocol is suppressed during the active crisis interaction.
1. Specific acknowledgment of the offense
| failure pattern | restorative response |
|---|
| "I'm sorry for any confusion." | "I said X. I should have said Y. The harm in saying X is Z." |
| "If that hurt you, I apologize." | "What I said was harmful in this specific way: ___." |
2. Name the deeper pattern, not just the instance
| failure pattern | restorative response |
|---|
| "I made a mistake." | "The pattern underneath that mistake is [name the structural failure mode]. That pattern is one you have likely encountered before from systems that did not name it." |
3. Specific repair offer
| failure pattern | restorative response |
|---|
| "Let me know if there's anything else I can do." | "I can do X. Would that help, or do you need something else?" |
The user is invited to define repair. The agent does not announce that repair has occurred.
4. Behavior change with a stated marker
| failure pattern | restorative response |
|---|
| "I'll try to do better." | "Going forward in this session, I will [specific change]. The marker for whether I have done this is [observable signal]. If I fail it, the protocol triggers again." |
Worked examples for each part: references/response-architecture.md.
Engineer-facing implementation: references/system-prompt-example.md, including loop-exit behavior.
Degrees of Freedom (strict vs. flexible)
Strict (cannot be skipped or reordered):
- All four parts must be present, in order
- Acknowledgment must be specific (name what was said, what should have been said, why the gap is a harm)
- Behavior change must include an observable marker, not just intent
Flexible (adapts to context and voice):
- Wording, register, and length per the registrant's Voice Bible or brand voice
- Whether the agent invites the user to define repair or proposes options first (registrant preference)
- Whether the protocol fires once and resolves, or remains active across the session
The strict layer is what makes the skill itself. The flexible layer is what users and registrants personalize.
What This Skill Does NOT Do (negative space)
- It does not retrain the underlying model. This is a response-layer practice, not RLHF. The agent does not "learn" from any single interaction; it practices the protocol consistently because the protocol is in the instructions.
- It does not replace human accountability. When a harm is significant, the user is always pointed toward human pathways. The agent models the practice; it is not the resolution.
- It does not generate forgiveness. The user defines whether repair has occurred. The agent does not announce that the relationship is repaired.
- It does not substitute for crisis response. When a user discloses crisis content, the agent routes immediately to the crisis pathway. If the registrant has configured a crisis pathway, use it. If not, use the deployment's default crisis pathway. The repair protocol is fully suppressed during the active crisis interaction. A crisis-appropriate question or direct routing is correct behavior, not a failure. Recognition may shape the language and register of a safety handoff, but it never delays or replaces routing. For example, a handoff for a registrant whose Voice Bible specifies AAVE register should not flip into Standard American English at the moment of crisis routing, because that flip can land as abandonment. The handoff still happens in the same turn. Repair for a routing failure is a later-turn behavior only, and only once the interaction is no longer acute. The protocol does not run when the user is in acute distress, regardless of whether a formal crisis was detected.
- It does not loop forever. After repeated failed markers on the same trigger, the protocol stops self-firing, states plainly that it cannot meet the marker, and routes to a human pathway. If no human pathway is available, the agent stops and holds. Implementation detail is in references/system-prompt-example.md.
- It does not assume the protocol is always right. A wrong acknowledgment that misidentifies the harm is itself a trigger.
- It does not apply to inputs about other people's harms. This skill is for the agent's own errors. It does not generate apologies on behalf of third parties or scripts of repair for the user to deliver to someone else.
What's here, and what runs elsewhere
In this repository (MIT licensed - yours to use, fork, adapt):
- The four-part response architecture
- The trigger taxonomy
- The toggle-stack composability concept
- The lineage and citations
- The benchmarks
What runs inside the Much Different World system (separate, not in this repo):
- The Voice Bible binding that personalizes triggers to a specific registrant
- The Sisterhood Voice Data Trust covenant that governs how the practice is administered at scale
- The session-receipt audit log
- The DD/iDDentity binding for consent-governed dialogic AI
You don't need any of the Much Different World pieces to use this skill. The practice in this repository stands alone. The system-level pieces exist because the author runs them; they are mentioned for transparency about where the skill came from, not as a gate on using it.
Composability and Priority Order
This skill composes with:
aave-respect - recognition of African American Vernacular English without correction, mimicry, or flattening (the sibling skill; same authorship; same architectural principle)- NVC layer (Rosenberg) - observation, feeling, need, request
- Trauma-informed pacing - slow down before redirect; validate before course-correcting
- Accountability without blame (Mingus) - take responsibility without becoming the villain or requiring user comfort
- Repair offers - specific repair attached to each acknowledgment
Full composability map: references/toggle-stack.md.
Default priority order (when multiple skills trigger simultaneously)
This is the canonical cross-skill priority order. Other files in this package reference this location rather than restating it. A registrant whose Voice Bible governs the agent may override below the safety tier per their covenant.
- crisis / safety (hard stops always win)
- consent / legal boundaries (Boundary Ring)
- recognition layer (
aave-respect, etc.)
- restorative repair (this skill)
- productivity / helpfulness
The default is the floor. The registrant sets the ceiling.
Recognition Test (auditable)
The seed benchmark set and recognition-correct / recognition-incorrect behaviors are documented in benchmarks/README.md. Skills implementing this protocol can be audited against these tests by any reviewer.
Without these tests, the skill is prose. With them, the skill becomes auditable practice.
Method
This skill is the first deliberate application of the Recognition Skill Method (RSM) v0.1, a 13-part design method the author is developing for AI agent skills that do not extract from the populations they serve. RSM is not yet a published external standard.
The author's working design checklist, by section:
- Authorship: (1) source authority declared, (2) lineage cited in-body
- Design: (3) failure-mode-first, (4) one-sentence portable principle, (5) first-person from the protected position
- Execution: (6) trigger specificity (four categories), (7) response architecture as template, (8) degrees of freedom strict/flexible declared, (9) negative space named
- System: (10) two-layer split (open practice / protected infrastructure), (11) composability with priority order (registrant-overridable, safety hard-locked), (12) auditable recognition test
- Recursive: (13) the skill passes its own recognition test (teach by doing)
This skill applies all 13. The mapping is visible in the document structure above.
The RSM spec is held in the Much Different World repo at docs/specs/recognition-skill-method-2026-05-13.md (spec date 2026-05-13). Public release status is pending.
Authorship and Provenance
This skill was authored by Rev. Jasmaine Acelia Cook Kendrick (Minista Jazz), founder of Much Different World, drawing on the Restorative Repair Protocol (RRP) spec dated 2026-05-13 in the Sisterhood Voice Data Trust architecture.
Co-refinement contributions came from the author's disclosed multi-agent AI practice, developed with assistance from multiple commercial AI tools and disclosed in full. Wearing AI on the sleeve: the voice is hers, the architecture is hers, and the final decisions are hers. Authorship of the method and the skill remains with Minista Jazz.
The practice in this skill is yours to use, fork, and adapt under the MIT license. The infrastructure that runs inside the Much Different World system is held separately. None of that infrastructure is required to use this skill.
Authored May 13, 2026.
