On this page
Safety → Sandbox → Skills → Solutions (SSSS) is not a motivational frame. It is operational infrastructure: the minimum structure under which AI-assisted work can scale without trading away governance, evidence, judgment, or durable workflows.
This document serves three roles at once:
- Canonical article — what each stage is, what it produces, and what “done” means in the field.
- Discovery bridge — how organizations find use cases (they are not delivered by vendors).
- Assessment backbone — a structured item bank, scoring logic, stage-integrity rules, and output contracts that product, workshops, and onboarding can implement without reinterpretation.
The load-bearing claim is unchanged from the canon: the order is the framework. Later stages borrow trust from earlier ones. Skip a tread, and you pay inversion tax—in policy drafted under fire, in generic donor voice, in tools that outlive the judgment meant to steer them.
How to read this document
| If you are… | Start here |
|---|---|
| Executive / board | Part 4 (problematic realities) + Part 3 output template |
| Program lead | Sandbox checklist + How use cases are found |
| Comms / development | Eight patterns + weekly mapping exercise |
| Product / data | Appendix A (item bank) + Appendix B (scoring + illusions) |
Part 1 — Stage checklists: what must be true
A stage is not done because a deck exists. It is done when the exit tests hold: you could defend the stage to a new ED, a regulator, or a long-time donor without improvisation.
Safety — governance + conviction + boundaries
Safety is the organization’s published ability to say yes and no before pressure arrives. Three layers, held together:
Governance
- Decision rights are named by risk tier, channel, and audience (who may ship what, who reviews, who escalates).
- Escalation paths exist for incidents, near-misses, and gray-area requests.
- One plain-language source of truth lives where staff actually work—not buried in a compliance drive only lawyers open.
- Review is calendarized for tool, vendor, and legal change—not “we’ll revisit.”
- Procurement and IT can translate Safety into purchasing criteria and environment rules.
Conviction lines (theology, ethics, or moral anthropology—use your org’s vocabulary)
- Principals can finish: “For us, that crosses a line because…” across truth, personhood, care, speech, and power—not only when the case is easy.
- Those lines can refuse plausible shortcuts, not only absurd abuse.
- Board or trustees have aligned, not only staff.
Boundaries
- Data sensitivity tiers are defined and wired to where data may go.
- Categories of work are explicit: never automate / only with human review / sandbox-only until evidence says otherwise—for your real surfaces (donor, pastoral, child safety, HR, board, liturgy, etc.).
- A new hire can be oriented in plain language to forbidden, review-required, and explore-first work.
Exit tests
- Executives state boundaries without notes in a hallway conversation.
- Legal and compliance are woven into governance, not stapled on after tools ship.
- Ambiguity has dropped: people report less guessing, not zero risk.
Sandbox — evidence creation under Safety
Sandbox is bounded experimentation that produces organizational memory. It is not shadow adoption, not “everyone try ChatGPT,” and not a pilot that becomes production because the button was convenient.
Structure
- Tooling and environments match Safety configuration (approved accounts, data tiers, logging expectations).
- Each run has hypothesis, scope, owner, duration, review rhythm.
- Inputs respect tiers (synthetic, public, approved subsets—not whatever is easiest).
- Nothing external ships without a named gate.
Artifacts
- Shared dated log: what was tried, settings, surprises, failures, kills, and holds.
- Failures recorded with the same dignity as wins—no success theater.
- Prioritized use-case portfolio, every candidate screened upstream by Safety (tier, boundary, owner).
- Top candidates have briefs + measurement notes, not only leadership anecdotes.
Exit tests
- Leadership points to dated evidence for “almost good enough” in your voice—not vendor demos.
- Vendor claims were tested on your work, not debated as ideology.
- Use cases are ready to graduate to formation and workflow design—or killed with reasons on paper.
How use cases are found (not given)
Vendors sell answers. Organizations need candidates. Use cases emerge when you map real weekly work to patterns where model-shaped help can matter—then filter ruthlessly through Safety before anything touches production.
These eight use-case detection patterns align with Movemental’s methodology catalog (same names, same traps). They are the bridge from “we should do AI” to “here is what we will test, under what hypothesis, with what risk flag.”
| Pattern | What it is | Examples (nonprofit / church / institution) | Value produced | Risk profile |
|---|---|---|---|---|
| 1. Repetition | Tasks done over and over, similar each time, costing real hours. | Donor thank-you sequences; weekly meeting recaps; volunteer confirmation mail. | Time (throughput); protects relational signal only if you define what must stay human in the touch. | Medium: speed without signal erodes warmth—recipients feel the template. |
| 2. Translation | Same substance, new audience or format. | Sermon → small-group guide; board memo → donor one-pager; policy → staff FAQ in another register. | Scale of appropriate communication; quality when nuance survives the move. | Medium–high: lowest-common-denominator “translation” flattens voice and doctrine-shaped care. |
| 3. Synthesis | Many sources → one clear read. | Strategic inputs → exec brief; interview transcripts → theme map; years of minutes → onboarding digest. | Cognitive load reduced; faster alignment. | High: false coherence—disagreement smoothed into one confident story. |
| 4. Generation | Blank-page work—something must exist that does not yet. | Grant outline; curriculum scaffold; first-pass job description; campaign skeleton before copywriting. | Cycle time to first artifact; momentum. | Medium: first draft mistaken for final; generic “competent” prose ships under fatigue. |
| 5. Transformation | Existing material improved—tone, clarity, length, structure-preserving edits. | Clarity pass on the right idea said badly; compress a sound long report; align voice to audience. | Quality; confidence in outward pieces. | Medium: edges that carried identity get ironed out—“smooth but not us.” |
| 6. Structuring | Messy thinking → legible structure. | Retreat whiteboard → decision memo; spoken discernment → options doc; ramble → decision log. | Coherence; decision hygiene. | High: bullets and matrices before thinking is real—the frame does the leading. |
| 7. Decision support | Options, trade-offs, scenarios around a human-owned choice. | Program shapes under constraints; partnership risk read; staffing scenarios under budget pressure. | Reasoning surface; fewer blind spots in framing. | High: outsourcing the call—the tool’s frame replaces the leader’s judgment. |
| 8. Personalization | One intent, many tailored variants tied to individual context. | Donor follow-ups referencing history; member outreach tuned to situation; coaching nudges at scale. | Relational depth at volume (when real care backs the touch). | Highest: the feeling of care without the fact of care; ethical and relational flag—senior review and explicit safeguards before sandbox entry. |
Curriculum exercise (60–90 minutes, cross-functional)
- List your weekly work — each person writes 15–25 bullets of real tasks (not job titles): emails sent, meetings run, docs produced, reports filed.
- Map tasks to patterns — label each bullet with one primary pattern (1–8). Debate only where it changes the test plan.
- Mark friction — circle high repetition, high latency, high error cost, or high coordination tax.
- Generate candidates — for circled items, write one line each: If we assisted this, hypothesis = …, success signal = …, kill signal = …, data tier = ….
- Safety screen — drop or redesign anything that violates tiers or boundaries; flag pattern 8 for explicit governance and pastoral/comms review before it enters any experiment queue.
That queue—not the vendor roadmap—is what Sandbox prioritizes.
Skills — formation, not training
Skills mean formed judgment: staff who can tell plausible from faithful, hold authorship without self-deception, and refuse shortcuts the letter of policy would allow but the spirit of mission forbids.
Formation
- Practice distinguishes plausible from good in your voice—not generic “quality.”
- Verification habits exist for facts, citations, tone, and mission fit.
- People say “smooth but not us” and know the next corrective move.
- Critiquing a draft is integrity, not obstruction.
Organizational habits
- Managers reward course-correction and named uncertainty—not only velocity.
- Rubrics cite Sandbox artifacts (real near-misses, real wins), not internet templates.
- Formation is cross-role—not owned by the most confident experimenter.
Exit tests
- Mid-level staff describe “good” with AI in the room without reading policy aloud.
- Self-correction in the wild without executive rescue.
- New scenarios not in the handbook still land in-bounds more often than six months ago.
Solutions — workflows, not tools
Solutions are workflows with instruments inside them: named inputs, outputs, owners, quality gates, failure modes, and measurement tied to outcomes—not licenses purchased.
Infrastructure
- Deployed work maps to graduated sandbox use cases and inherits Safety’s legal and tier rules (DPAs, BAAs, sub-processors, etc., as applicable).
- Ownership survives turnover—documentation lives with the workflow.
Portfolio discipline
- Augmentation is the default; automation only where judgment is explicit; composition rare until governance and Skills hold it.
- Tool swap does not erase the practice.
Exit tests
- Vendor conversations shorten: serves a graduated workflow and meets constraints—or not yet.
- Incidents yield proportionate adjustment, not panic-freeze or blanket bans—because intent was clear before scale.
Part 2 — Naming: decision stack (reduced drift)
Long lists of clever names rot in execution. Below: nine vetted options and an explicit stack—pick one public metaphor, one internal spine, one shortform.
Nine names (when you need alternates)
| Name | Best use |
|---|---|
| The Trust Staircase | External narrative; implies no skipping. |
| Safety → Sandbox → Skills → Solutions | Internal policy, board packets, procurement—always spell once. |
| SSSS | Shorthand only after the sequence is taught once. |
| The Integrity Sequence | Donor- or trustee-facing moral seriousness. |
| Govern → Learn → Form → Build | Verb stack for workshop agendas and scorecards. |
| Four Before Scale | Executive headline against premature rollout. |
| Evidence Before Expansion | Sandbox-forward discipline for skeptical practitioners. |
| Judgment Before Automation | Technical and finance audiences; curbs overshoot. |
| Staircase, Not Menu | Anti-workshop-catalog; use with explanation. |
Recommended stack (default)
| Layer | Use this |
|---|---|
| External (site, talks, book jacket) | The Trust Staircase — subtitle: Safety, Sandbox, Skills, Solutions. |
| Internal (ops, HR, legal, IT) | Spelled-out stage names every time; link to one canonical policy page. |
| Shortform (rubrics, Slack, engineering) | SSSS — never redefine the fourth S as “strategy” or “scale.” |
Do not lead with “AI 4S Roadmap” unless you define all four S words in the same breath. Otherwise it reads like a SKU and trains adults to treat the path as four disconnected workshops.
Part 3 — Assessment: product-ready backbone
What follows is not a vibe check. It is an item bank each question can be imported into a form builder, LMS, or database as a row. Scales are 1 = strongly disagree through 5 = strongly agree unless noted.
A. Misdiagnosis risks (read before scoring)
Organizations routinely think they are farther along than they are:
| Illusion | Feels like | Actually is |
|---|---|---|
| “We have Safety” | A PDF exists; counsel “looked at it.” | Principals cannot cite boundaries under fatigue—Safety is archival, not operational. |
| “We have a sandbox” | Many people tried many tools. | No dated log, no tier discipline—shadow adoption with a label. |
| “We did Skills” | Everyone attended a webinar. | No live critique of real work—training theater, not formation. |
| “We’re deploying Solutions” | Licenses and pilots everywhere. | No workflow map, no graduated use cases—shopping, not infrastructure. |
| “We’re advanced” | High tool use, confident champions. | Solutions score > Sandbox + Skills in the integrity profile—classic inversion. |
Scoring exists to surface illusions, not to flatter.
Appendix A — Assessment item bank (schema)
Each item: id, stage, category, weight (default 1), prompt.
| id | stage | category | weight | prompt |
|---|---|---|---|---|
| Q01 | Safety | boundaries_and_authority | 1 | We can state, without notes, what is forbidden in external-facing work and what requires human review. |
| Q02 | Safety | governance_artifact | 1 | We have a published map of decision rights—not only informal habit. |
| Q03 | Safety | conviction_lines | 1 | Our deepest convictions (theological or ethical) are explicit enough to say no to plausible shortcuts. |
| Q04 | Safety | operational_spread | 1 | Data sensitivity tiers and escalation paths are understood across departments, not only legal/IT. |
| Q05 | Sandbox | learning_artifact | 1 | We keep a shared dated log of experiments, surprises, and failures—not only success stories. |
| Q06 | Sandbox | environment_compliance | 1 | Our experiments run in environments that comply with Safety, not primarily as shadow individual use. |
| Q07 | Sandbox | portfolio_discipline | 1 | We have a prioritized use-case portfolio screened by governance constraints. |
| Q08 | Sandbox | evidence_quality | 1 | We can point to evidence of what “good” and “not us” look like for our voice—not only opinions. |
| Q09 | Skills | distributed_judgment | 1 | Mid-level staff can describe good AI-assisted work without reading policy verbatim. |
| Q10 | Skills | culture_of_correction | 1 | We see public self-correction when outputs drift (voice, facts, ethics). |
| Q11 | Skills | verification_norms | 1 | Verification habits are social norm, not heroics by one reviewer. |
| Q12 | Skills | formation_vs_training | 1 | Training time is spent on judgment, not only buttonology. |
| Q13 | Solutions | workflow_infrastructure | 1 | We deploy workflows with clear owners, gates, and failure modes—not tool brands as substitutes for design. |
| Q14 | Solutions | procurement_gates | 1 | Procurement conversations are shortened by pre-baked constraints and graduated use cases. |
| Q15 | Solutions | measurement_legibility | 1 | We measure workflow outcomes, not only licenses activated. |
| Q16 | Solutions | tool_independence | 1 | We could swap tools without losing the practice (documentation + skill). |
| Q17 | Cross | honest_location | 2 | We know where we are in the sequence—and where we skipped—without self-deception. |
| Q18 | Cross | incident_posture | 1 | When something goes wrong, we adjust proportionately rather than panic-freeze or ban everything. |
Weighted total (default weights):
S = sum(score_i × weight_i) over items, max = 5 × sum(weights) = 5 × 19 = 95 (because Q17 weight is 2).
Normalized overall: S_norm = S / max_S → interpret as 0–100% maturity signal, not moral worth.
Appendix B — Stage integrity score and hidden inversion
Stage subscore (equal weight within stage unless you add item weights later)
For stage st with item set I_st:
Subscore_st = (sum of scores_i for i in I_st) / (5 × |I_st|) → 0–1 (or multiply by 100 for percent).
| Stage | Items |
|---|---|
| Safety | Q01–Q04 |
| Sandbox | Q05–Q08 |
| Skills | Q09–Q12 |
| Solutions | Q13–Q16 |
| Cross | Q17–Q18 (optional separate “meta” band) |
Weakest dimension (within stage)
Group items by category; average scores per category; minimum category average is the weakest dimension for that stage—your first surgical fix.
Hidden inversion risk (rules of thumb)
Apply these after subscores are computed:
- Solutions-before-evidence: If
Subscore_Solutions − Subscore_Sandbox ≥ 0.15and Q07 or Q08 < 4 → likely illusion: deployed breadth without graduated use-case discipline. - Skills theater: If
Subscore_Skills ≥ 0.75and Q05 < 3 → likely illusion: confident individuals, no organizational memory. - Safety on paper: If Q02 ≥ 4 and Q01 < 3 → likely illusion: document exists, principals do not carry it.
- Honesty gap: If normalized overall ≥ 0.72 and Q17 ≤ 2 → likely illusion: high scores except honest location—treat overall as inflated.
These flags are heuristics for coaching and product UI, not legal determinations.
Appendix C — Band interpretation (overall normalized score)
Use S_norm from Appendix A. These bands pair likely footing with primary misdiagnosis risk.
| S_norm | Likely footing | Primary misdiagnosis risk |
|---|---|---|
| 0.00–0.42 | Early or Safety collapsed | “We’re being thoughtful” while Solutions-first churn continues in shadow. |
| 0.43–0.55 | Safety partial / policy theater | “Legal signed off” mistaken for executives carrying boundaries. |
| 0.56–0.68 | Safety real; Sandbox immature | “We’re experimenting” mistaken for sandbox with memory. |
| 0.69–0.78 | Sandbox producing evidence; Skills uneven | “We sound fine” while voice drifts toward genre-default donor prose. |
| 0.79–0.87 | Skills strong; Solutions selective | Automation/composition overshoot under vendor pressure. |
| 0.88–1.00 | Solutions as infrastructure | Complacency as models and vendors shift underfoot. |
Lowest-three-items rule (non-negotiable)
Regardless of band: the three lowest item scores (raw, after weighting if you sort by contribution gap) define the next 90 days. Averages lie; minimums tell the truth.
Appendix D — Required assessment output (contract)
Any tool implementing this bank should emit at minimum:
- Stage distribution — for each of Safety, Sandbox, Skills, Solutions:
Subscore_stas percentage (optional: simple bar representation). - Top 3 weaknesses — the three lowest raw item scores, with
id,stage,category, and prompt text echoed. - Weakest dimension per stage — category label + average for that category within the stage.
- Next 90-day focus — one paragraph generated from: lowest three items + weakest dimensions + one inversion flag if fired.
- Likely illusion — single string chosen from:
none|safety_paper|shadow_sandbox|skills_theater|solutions_without_evidence|honesty_gap|inversion_profile(use rules in Appendix B;inversion_profileif multiple flags).
Example output skeleton (copy/paste for workshops)
SSSS Assessment Summary
-----------------------
Overall normalized score (S_norm): __%
Stage integrity (0–100%):
Safety: __% | weakest dimension: _____________
Sandbox: __% | weakest dimension: _____________
Skills: __% | weakest dimension: _____________
Solutions: __% | weakest dimension: _____________
Cross: __% (Q17–Q18)
Top 3 weaknesses (item id — one-line fix mandate):
1. ___
2. ___
3. ___
Inversion / illusion flags: ___
Likely illusion (one): ___
Next 90 days (one focus, one “stop doing”):
Focus: ___
Stop: ___
Appendix E — Remediation map (by dominant gap)
| Lowest stage (by Subscore) | First moves |
|---|---|
| Safety | Pause net-new external AI-assisted channels; executive alignment session; one-page governance + boundaries; procurement freeze until tiers bind. |
| Sandbox | Charter bounded runs; assign log owner; run eight-pattern exercise; graduate or kill candidates with written reasons. |
| Skills | Replace webinars with live artifact critique; rubrics from Sandbox logs; peer review loops tied to real donor/program surfaces. |
| Solutions | Workflow mapping workshop; one workflow end-to-end with metrics; retire redundant tools; composition only with named architect + audit path. |
| Cross (Q17–Q18 low) | Location ritual: principals answer exit tests aloud; rehearse one incident without blame; reread Why Order Matters. |
Part 4 — Problematic realities (what it feels like inside)
Each failure mode below is one felt reality inside the building—plus one concrete scenario you can recognize without theory.
Nothing real (rhetoric only or pure drift)
Inside the org: Speed feels like virtue. No one can say what is forbidden. The real curriculum is whoever is boldest in Slack.
Scenario: Three mid-year donor letters go out—smooth, grateful, mission-flavored—and a longtime donor replies to the ED: “These felt… generic. Did something change?” No one can answer whether AI was involved, what was reviewed, or what “us” means anymore.
Stopped after Safety (fence without field)
Inside the org: Compliance is calmer, but innovation is either smuggled or frozen. Staff still lack shared evidence of what works in this ministry.
Scenario: Policy is published; no sandbox log exists. Program staff quietly use personal tools for grant language because “official channels take too long.” Leadership believes the house is in order; shadow learning widens the gap between paper and practice.
Stopped after Sandbox (museum without judgment or rails)
Inside the org: The org has stories but not distributed judgment. A hero holds the prompts; convenience pushes pilots toward undeclared production.
Scenario: The comms director’s “sandbox” draft goes straight to a appeal segment because the deadline moved up. Two weeks later, two appeals share a phrase cluster with another org’s campaign—no one logged the experiment, so no one can audit what happened.
Stopped after Skills (craft without workshop)
Inside the org: People know what good looks like, but every week still feels artisanal. Wins do not compound into owned workflows; vendor count creeps.
Scenario: Staff nail tone in a workshop critique, then Monday reverts to five tools and no single workflow owner—the ED still gets pulled in to “sense-check” everything because infrastructure was never named.
Why the whole journey
Safety names the creature you refuse to become. Sandbox replaces opinion with dated evidence. Skills distributes judgment policy cannot script. Solutions makes the work boring in the right way—workflows and audits survive tool churn.
When it holds, you are not “doing AI.” You are operating with instruments inside workflows the mission already owns.
From article to system
This document is intentionally multi-headed. It should feed:
| Surface | What to import |
|---|---|
| Assessment product | Appendix A (rows), B (flags), C (bands), D (output contract). |
| Workshop curriculum | Part 1 checklists + eight-pattern exercise + lowest-three-items ritual. |
| Discovery Lab / sandbox engagements | “How use cases are found” + portfolio screening rules. |
| Platform onboarding | Stage exit tests as gates; naming stack for consistent UI copy; illusion strings for coaching tips. |
It is not “just an article.” It is a core system artifact: the same sentences can appear in canon prose, in a facilitator guide, and in a database seed file—without drift—as long as the appendices stay the single source of truth for item text and weights.
When product or curriculum diverges, reconcile here first, then propagate.