A clinic posts hours, yet patients show up to a locked door. A responsive clinic apologises, posts why it happened, updates hours, and texts people next time. The fix becomes part of the system.
Competent action creates new information. Refusing to hear it is the fastest path to failure.
Tronto defines responsiveness as "observing that response and making judgements about it — whether the care given was sufficient, successful, or complete." Crucially, "the person cared for need not be the one who responds" — others in the care setting can assess. And "new needs will undoubtedly arise," restarting the cycle back to attentiveness.
The illustration's framing: We check the results — not metrics that don't measure what we value, but metrics designed by and for the people affected.
Core ideas
- People closest to harm define harm. They get the author pen for evaluations.
- Right to reply is a right to improvement. A reply that cannot cause change is theatre.
- Shared memory. Post-incident learnings become tests; tests prevent repeats.
- Time as a service. A fast wrong-then-right beats a slow maybe. Use reversible defaults.
- Care labour is labour. The people closest to harm who define harm also bear the labour cost of defining it — content moderation, data labelling, eval authoring, and assembly facilitation are care-giving work. That work must be compensated and visible, not treated as free QA.
What good responsiveness looks like
- Community-designed reward. Train agents with Reinforcement Learning from Community Feedback (RLCF): optimise for cross-group endorsement and trust-under-loss, not raw engagement. The community defines what "good" means — and the definition evolves, but within guardrails: justice, equality, and freedom.
- Weval-style registries. Employ a "Wikipedia for evals": Anyone can draft an eval; civil society partners peer-review; labs adopt or explain. These are not lab-designed benchmarks vendors can teach to the test — these are living, community-maintained test suites.
- Clear appeals. Answer urgent cases in 48 hours, standard in 7 days, complex in 30 days. Remedies include correction, rollback, or compensation.
- Incident run-books. Severity-class playbooks with on-call roles and communication templates.
- Public repair log. Give each incident a page: what happened, whom the incident affected, fixes, dates, and the test that now guards against future instances.
From ideas to practice
- Expose the appeal button. Show that a one-click appeal exists with a clock everywhere a decision is shown. The decision trace is issued to the user, who chooses whether to unseal it for a public challenge.
- Accept harm drafts. Let people submit proposed evals in plain words; convert to tests with partners.
- Triage by severity. Task POs with classifying by the community's severity scale; the highest classes trigger immediate pause or reversible defaults.
- Fix or explain. Publish the remedy or the reason with next steps, on the clock.
- Memorialize. Turn the incident into a test; add to the eval registry; and link from the contract changelog.
- Check back. Close the loop with those who appealed; measure trust-under-loss.
Tools (buildable today)
- Appeal API (with timers, statuses, escalation).
- RLCF pipeline. Community feedback to reward shaping; bridge scores as reward signals.
- Eval editor (plain-language → test harness).
- Incident tracker (severity, owners, deadlines, public notes).
- Repair log template (root cause, remedy, test added).
Flood-bot story: Part IV
- Appeals surge. A language community flags mistranslations in proof rules.
- Local eval. Community partners submit a translation-fidelity eval; the group is compensated from the project's escrow fund, because local cultural knowledge is labour, not free QA. The bot fails the eval; pause triggers; reversible defaults apply.
- RLCF. The payout policy is retrained to reward on-time delivery without spiking appeals in any cluster. The community's eval becomes a permanent reward signal.
- Fix. Bilingual reviewers update rules; new test guards future changes.
- Close the loop. Elena gets a text: "We fixed the error; here is your new decision; and here's how to see what changed." Trust-under-loss ticks up.
What could go wrong
- Appeal maze. Too many steps. Fix: Single button; auto-escalation if SLA breach.
- Eval spam. Low-quality tests flood the system. Fix: Partner moderation; reputation for contributors; merge/duplicate tools.
- Blame storms. People, not processes, get blamed. Fix: Blameless post-mortems; focus on mechanism design.
- Weaponised appeals. Adversaries flood appeals or strategically trigger pauses to disrupt service. Fix: Require authenticated standing (not public identity) for pause triggers; rate-limit by community; preserve priority access for those directly affected.
Interfaces with other packs
- From Responsibility (Pack 2): who acts is clear; remedies are wired.
- From Competence (Pack 3): observability and guardrails feed responsiveness; incident loops start here.
- To Attentiveness (Pack 1): new needs discovered through response reshape what we notice — the cycle restarts.
- To Solidarity (Pack 5): public repair culture builds cross-group trust.
- To Symbiosis (Pack 6): responsive agents earn the right to stay local.
A closing image: the workshop wall of retired broken parts
Imagine a workshop with a wall full of retired "broken parts," each tagged with the story of how it broke, how to avoid future damage, and who fixed it. The wall is not a wall of shame — it is a wall of learning. The shop that hides its breaks will repeat them.