Start with the map. Packs 1–4 form the care cycle, Pack 5 scales that cycle across organisations, and Pack 6 keeps every deployment local, plural and ready to end. The map is the framework index: pack stations open their chapters, measure links land on the audit entries below, and handoffs show what each phase gives the next.¶ 1 (permalink to paragraph 1)
Framework map
The care cycle — Packs 1–4
Pack 1 · Attentiveness
Headline measureRepresentation gap
- Also hands Pack 3 · Competencecaution areas become safe-to-fail trials
Pack 1 · Attentiveness hands to Pack 2 · Responsibility: who, what, why + rights flags
Pack 2 · Responsibility
Headline measurePromise fidelity
- Also hands Pack 4 · Responsivenessremedies wired before launch
Pack 2 · Responsibility hands to Pack 3 · Competence: specs, SLAs, brakes
Pack 3 · Competence
Headline measureVerified execution rate
Pack 3 · Competence hands to Pack 4 · Responsiveness: traces, incidents
Pack 4 · Responsiveness
Headline measureTrust-under-loss
Pack 4 · Responsiveness hands to Pack 1 · Attentiveness: new needs restart the cycleback to Pack 1
The field between organisations
Pack 5 · Solidarity
Headline measureUncommon-ground index
- From Pack 1 · Attentivenessbridging map hands over the terrain
- From Pack 2 · Responsibilityportability + exit clauses in every contract
- From Pack 4 · Responsivenessrepair culture builds cross-org trust
Pack 6 · Symbiosis — the boundary in time
Headline measureExit readiness
- From Pack 5 · Solidaritytreaties, IDs, portability make symbiosis feasible
- From Pack 3 · Competencecompetent, responsive systems earn stewardship
- To Pack 1 · Attentivenessretired systems gift maps, evals, records
- Both ways with Pack 2 · Responsibilitybounds and sunsets are contractual
Pack 6 keeps each pack from its negation — surveillance · contractual trap · indispensability · dependency-as-care · uncontrolled infrastructure — and guards itself: symbiosis without service continuity becomes abandonment.
Start with a question you can ask
You do not need a dashboard to begin. Ask your school, care provider or residents’ group: “Who did we miss?”, “Who promised to act?” or “What changed after someone said it went wrong?” Follow the relevant pack below and ask to see the record. A bridging map shows whose account was heard; an obligation ledger names a promise and its owner; a repair log records what happened after an appeal.¶ 2 (permalink to paragraph 2)
Each pack then answers its own public audit question, with one headline public measure and supporting diagnostics.¶ 3 (permalink to paragraph 3)
These six measures are our proposals for public evaluation and have not been validated as measurement standards. Their definitions and thresholds need testing with the people affected.¶ 4 (permalink to paragraph 4)
These metrics are designed for sufficiency, not maximisation. Each deployment context defines a threshold: “good enough” for that community. Crossing it is the goal, and chasing a higher score risks the same metric gaming the 6-Pack warns against. A headline measure is a public test, not a totalising score.¶ 5 (permalink to paragraph 5)
Headline public measures
Representation gap. The Attentiveness measure: Which materially affected groups are still missing or badly under-represented in the record?¶ 6 (permalink to paragraph 6)
- Traditional Mandarin name
- 代表性落差
- Diagnostics
- Coverage of affected people; voice equity between least-heard and most-heard groups.
- Refuses
- Counts only when the least-heard groups gain real standing in the record. Averaging dissenters into the middle narrows the gap on paper and earns no credit.
- Threshold
- Set by each deployment context: “good enough” for that community.
- Read as
- A public test.
Promise fidelity. The Responsibility measure: What share of material obligations are explicitly owned, properly authorised, and kept on their published terms?¶ 7 (permalink to paragraph 7)
- Traditional Mandarin name
- 承諾保真度
- Diagnostics
- Named-owner coverage; authority-match rate; adopt-or-explain rate.
- Refuses
- Rewards promises that are owned and honoured. A promise that is merely well-worded earns no credit, so a high score needs authority and funding that match the duty.
- Threshold
- Set by each deployment context: “good enough” for that community.
- Read as
- A public test.
Verified execution rate. The Competence measure: What share of audited decisions or releases pass guardrails, include a usable trace, and stay inside release bounds?¶ 8 (permalink to paragraph 8)
- Traditional Mandarin name
- 經驗證的執行率
- Diagnostics
- Guardrail integrity; trace completeness; canary health; audit overturn rate.
- Refuses
- Rises only when an independent auditor can reconstruct the traces and the canaries cover the cases they claim to cover. Passing easy cases and calling the ceremony proof earns no credit.
- Threshold
- Set by each deployment context: “good enough” for that community.
- Read as
- A public test.
Trust-under-loss. The Responsiveness measure: After a bad outcome and attempted repair, do affected people report that the system became more trustworthy rather than less?¶ 9 (permalink to paragraph 9)
- Traditional Mandarin name
- 失利下的信任
- Diagnostics
- Appeal closure time; repair completion rate; correction throughput ratio; correction backlog depth and age; harm recurrence within 90 days.
- Refuses
- Counts only when tied to an accountable identity and corroborated by an independent signal: a return to the service, an appeal withdrawn because the repair settled it (not abandoned from exhaustion), or a third party who can attest. A campaign can manufacture sentiment, and the actor who caused the harm can astroturf self-reported trust after a repair, so sentiment alone earns no credit.
- Threshold
- Set by each deployment context: “good enough” for that community.
- Read as
- A public test.
Uncommon-ground index. The Solidarity measure: Are shared decisions showing real cross-group participation and co-endorsement, rather than separate silos?¶ 10 (permalink to paragraph 10)
- Traditional Mandarin name
- 罕見共識指數
- Diagnostics
- Portability success rate; accountable-identity coverage; federation participation, the structural conditions that keep co-action freely given.
- Refuses
- Rewards groups that listen across lines. A group that falls silent earns no credit, and neither does a group curated into agreement. Read the index against Pack 1’s representation gap, because the cheapest way to raise co-endorsement is to exclude whoever would withhold it.
- Threshold
- Set by each deployment context: “good enough” for that community.
- Read as
- A public test.
Exit readiness. The Symbiosis measure: Could this system hand over or shut down on schedule without rights loss, continuity failure, or recentralisation?¶ 11 (permalink to paragraph 11)
- Traditional Mandarin name
- 退出就緒度
- Diagnostics
- Scope compliance; sunset compliance; ecology diversity; handover rehearsal pass rate.
- Refuses
- Counts only when a system can hand off or shut down on schedule without rights loss, continuity failure or recentralisation. An exit that exists only on paper earns no credit.
- Threshold
- Set by each deployment context: “good enough” for that community.
- Read as
- A public test.
Supporting diagnostics
- Pack 1 — Attentiveness. Coverage of affected people; voice equity between least-heard and most-heard groups. Evidence artefact: the bridging map.¶ 12 (permalink to paragraph 12)
- Pack 2 — Responsibility. Named-owner coverage; authority-match rate; adopt-or-explain rate.¶ 13 (permalink to paragraph 13)
- Pack 3 — Competence. Guardrail integrity; trace completeness; canary health; audit overturn rate.¶ 14 (permalink to paragraph 14)
- Pack 4 — Responsiveness. Appeal closure time; repair completion rate; correction throughput ratio; correction backlog depth and age; harm recurrence within 90 days.¶ 15 (permalink to paragraph 15)
- Pack 5 — Solidarity. Portability success rate; accountable-identity coverage; federation participation, the structural conditions that keep co-action freely given.¶ 16 (permalink to paragraph 16)
- Pack 6 — Symbiosis. Scope compliance; sunset compliance; ecology diversity; handover rehearsal pass rate.¶ 17 (permalink to paragraph 17)
Named instruments
The packs describe these tools so you can build them, or ask for them, without having read the book. Each supports a public audit question. A completed document alone does not show that care worked.¶ 18 (permalink to paragraph 18)
- Perspective receipts (Pack 1) — a receipt each contributor can hold, which lets them find and correct how their words were represented in the record. It is separate from the civic receipt a decision trace carries for settlement.¶ 19 (permalink to paragraph 19)
- Engagement contract (Pack 2) — a short, readable public document for every significant Kami deployment. It records what the system is supposed to do, who is answerable for that, what happens when it goes wrong and how the deployment will end. Pack 6 can render the same contract machine-readable so infrastructure can check purpose limits, consent rules, data-retention limits, portability guarantees and shutdown procedures. A machine-readable clause is not yet an enforceable shutdown.¶ 20 (permalink to paragraph 20)
- Adopt-or-explain (Pack 2) — the rule that when an Alignment Assembly produces a recommendation, the team either integrates it into the system’s behaviour or publishes a reasoned explanation of why not, with the remedy offered instead.¶ 21 (permalink to paragraph 21)
- Obligation ledger (Pack 2) — a public, weekly, digitally signed record of what the deployment has committed to deliver and whether it is meeting those commitments, kept by a named Participation Officer.¶ 22 (permalink to paragraph 22)
- Shadow mode (Pack 3) — the first stage of the Apprentice Model: the system sees real inputs and proposes real actions but does not act, and its proposals are compared against the human or prior system so divergences can be studied before trust is granted.¶ 23 (permalink to paragraph 23)
- Canary release (Pack 3) — the second stage: deployment to a small, stratified slice of real cases with automatic rollback triggers if drift exceeds defined bounds. A canary is a probe and cannot stand in for the public.¶ 24 (permalink to paragraph 24)
- (Packs 3–4) — a public, Wikipedia-like registry of versioned, community-authored evaluations. Affected people write the tests for harm and repair (Pack 4), and passing them becomes a release gate (Pack 3). Weval, built by the Collective Intelligence Project, is the working example, and one of us (Audrey) is a senior research fellow there.¶ 25 (permalink to paragraph 25)
- Decision trace (Pack 3) — the per-action record showing which rule fired, which sources were consulted and what uncertainty the system carried. It is what makes the engagement contract live in practice. When a Kami draws on a community’s knowledge, the same trace is designed to double as a civic receipt (Pack 6), settling value back to the custodians through pre-funded escrow. That is a design commitment and no settlement runs yet. Pack 1’s perspective receipts are different: they let contributors see and correct how they were represented.¶ 26 (permalink to paragraph 26)
- The brake (circuit breaker) (Packs 3–4) — a single prominent control, wired up and tested, that stops the system now. A person can reach it as fast as they recognise a problem, so they can halt a wrong action in the moment as well as flag it for later.¶ 27 (permalink to paragraph 27)
- Override ledger (Pack 4) — a plain-text working memory that records every time a person in the group said “no” to the Kami. It is evidence that the governance charter is live and the group still has standing to correct the Kami.¶ 28 (permalink to paragraph 28)
- Correction backlog board (Pack 4) — a public queue of unresolved corrections with age, ownership and escalation status, so correction debt cannot hide and ageing cases get reassigned or escalated.¶ 29 (permalink to paragraph 29)
What each measure refuses to reward
Each headline measure also names what it must refuse, so a good number cannot pass for good care.¶ 30 (permalink to paragraph 30)
- Representation gap (Pack 1). Counts only when the least-heard groups gain real standing in the record. Averaging dissenters into the middle narrows the gap on paper and earns no credit.¶ 31 (permalink to paragraph 31)
- Promise fidelity (Pack 2). Rewards promises that are owned and honoured. A promise that is merely well-worded earns no credit, so a high score needs authority and funding that match the duty.¶ 32 (permalink to paragraph 32)
- Verified execution rate (Pack 3). Rises only when an independent auditor can reconstruct the traces and the canaries cover the cases they claim to cover. Passing easy cases and calling the ceremony proof earns no credit. Operational controls are not a safety proof.¶ 33 (permalink to paragraph 33)
- Trust-under-loss (Pack 4). Counts only when tied to an accountable identity and corroborated by an independent signal: a return to the service, an appeal withdrawn because the repair settled it (not abandoned from exhaustion), or a third party who can attest. A campaign can manufacture sentiment, and the actor who caused the harm can astroturf self-reported trust after a repair, so sentiment alone earns no credit.¶ 34 (permalink to paragraph 34)
- Uncommon-ground index (Pack 5). Rewards groups that listen across lines. A group that falls silent earns no credit, and neither does a group curated into agreement. Read the index against Pack 1’s representation gap, because the cheapest way to raise co-endorsement is to exclude whoever would withhold it. A floor protects standing opposition, so the measure treats a persistent minority as legitimate disagreement.¶ 35 (permalink to paragraph 35)
- Exit readiness (Pack 6). Counts only when a system can hand off or shut down on schedule without rights loss, continuity failure or recentralisation. An exit that exists only on paper earns no credit.¶ 36 (permalink to paragraph 36)
How to read them
- Trust is split into six questions. Pack 1 asks whether people were heard. Pack 2 asks whether promises were real. Pack 3 asks whether delivery held up under inspection. Pack 4 asks whether repair worked after harm. Pack 5 asks whether groups can act together fairly. Pack 6 asks whether stewardship can remain bounded over time.¶ 37 (permalink to paragraph 37)
- Artefacts are governance evidence. Bridging maps, engagement contracts, repair logs and exit drills support the headline measures. None is a score by itself.¶ 38 (permalink to paragraph 38)
- The map and the index measure different things. Pack 1’s bridging map charts where crossings are possible: who agrees, who clashes and why, among those present in the record. Pack 5’s uncommon-ground index counts the crossings actually made: shared decisions co-endorsed across group lines. Terrain first, then traffic. Composition precedes co-action, so a high index over a wide representation gap measures selection, not solidarity.¶ 39 (permalink to paragraph 39)
- Training signals are separate from the public audit. Cross-group endorsement may inform model tuning or routing. Its public role in this framework is the Pack 5 uncommon-ground index. Pack 4 owns whether affected people could correct the system and whether repair restored trust.¶ 40 (permalink to paragraph 40)
- The record is small and named. Most strong cases come from places with unusually strong civic infrastructure. Each community earns its thresholds setting by setting, and none inherits them from Taiwan or California.¶ 41 (permalink to paragraph 41)
Before you call it Civic AI
The measures above audit a running deployment. These six questions gate whether a deployment should carry the name at all:¶ 42 (permalink to paragraph 42)
- Human standing (Packs 1, 5) — can affected people prove standing without surrendering unnecessary identity?¶ 43 (permalink to paragraph 43)
- Adversarial load (Packs 1, 4, 5) — what stops synthetic publics, bots, or paid campaigns from flooding the channel?¶ 44 (permalink to paragraph 44)
- Decision force (Pack 2) — where does the output bind actual decisions, and where is it only advice?¶ 45 (permalink to paragraph 45)
- Repair path (Pack 4) — who can appeal, who pays for harm, and where is the repair logged?¶ 46 (permalink to paragraph 46)
- Exit capacity (Packs 5, 6) — can the community migrate, fork, pause, or shut down without losing service continuity?¶ 47 (permalink to paragraph 47)
- Boundary honesty (Pack 6) — what problem is out of scope, and who is carrying that problem instead?¶ 48 (permalink to paragraph 48)
If the answer to any of these is vague, the deployment is not ready. The aim is to make the people running it answerable.¶ 49 (permalink to paragraph 49)
Where the measures come from
The measures are ours. The ideas under them come from earlier work. “Sufficiency, not maximisation” is Herbert Simon’s satisficing (1955). The warning that a measure which becomes a target stops measuring has three classic statements: Campbell’s law (1979), Goodhart’s law (1975), and Marilyn Strathern’s shorthand for it (1997). “What each measure refuses to reward” is written against all three.¶ 50 (permalink to paragraph 50)
We propose trust-under-loss as a deployment test informed by procedural-justice research, including Tom Tyler’s Why People Obey the Law. That research does not validate this AI measure or guarantee that repair restores trust. The uncommon-ground index audits the cross-group signal that Aviv Ovadya’s bridging-based ranking uses to rank. Exit readiness is Albert Hirschman’s exit turned into a drill. The Sources page carries the full trail.¶ 51 (permalink to paragraph 51)
