Detect the risk
Mirage Core inspects evidence for faces, likely minors, embedded location data, consent state, and the signals needed for verification.
The trust architecture behind the story
Protect first.Publish second.
The first ATP brief showed what verified impact can do for ministry and donors. This companion explains the engineering discipline that makes that promise safe for children, field teams, stewards, and the people ATP serves.
July + August hardening thesisMirage did not simply add privacy features. It rebuilt the path from field capture to public delivery so a processed claim is not enough. The safe artifact must be the artifact that is actually served.
01 · Executive frame
Trust is a system behavior, not a checkbox.
ATP's work creates an unusually high bar: field evidence may contain children, trauma survivors, at-risk adults, and exact coordinates. July and August focused on making privacy structural across the entire delivery chain, then applying that foundation to reporting, compliance, and partner workflows.
Create the safe artifact
The pipeline produces a blurred or metadata-scrubbed derivative rather than merely recording that processing occurred.
Control every exit
Public and donor surfaces select the sanitized copy. For minor-face safety, uncertain or incomplete processing withholds media instead of exposing the original.
02 · July + August hardening
What the development cycle actually changed.
The work began with a security assessment, then followed defects all the way through storage, callbacks, serving paths, human review, and repeatable regression tests.
July 9-15
Security assessment becomes engineering work
Authorization isolation, browser security headers, rate limits, error handling, and GPS/EXIF controls were reviewed. An automated two-organization test found and closed a cross-organization project-access gap before scale.
July 15-16
Minor protection fixed end to end
The team discovered that blur computation alone did not guarantee the blurred bytes reached users. Mirage was changed to persist the sanitized image, authenticate the callback, route every public photo path through one serving policy, and fail closed when minor detection is uncertain.
July 17-21
Stewards gain "Blur all faces"
For hostile environments and at-risk adults, stewards can require every face to be obscured. The photo stays off public pages while the stronger redaction is produced. Blur-all overrides a narrower face-permission decision.
July 21-22
Location scrubbing becomes server-side
Mirage Core began producing an EXIF/GPS-stripped copy for every photo whose source carried metadata, including clean photos without minors. Serving paths use the stripped derivative while extracted coordinates can remain internal for project-distance validation.
July 21-Aug 4
Privacy becomes continuously testable
Synthetic image batteries now exercise the real detect → blur → strip → persist → serve pipeline. The admin self-test reached 16/16 scenarios in production; the deterministic 4/4 suite runs in the shared CI gate.
Aug 4-17
ATP-specific reporting moves forward
The first PCITF press-release ingest path was delivered, independent public evidence was connected to the steward workflow, and the ECFA roadmap translated Standards 4, 5, and 7 into practical evidence and reporting steps.
Aug 25
A year-end runway replaces a premature commitment
ATP keeps access without a financial obligation while Mirage proves that it can reduce work for Malia and ATP's project partners. The goal is a clear November-December decision and, if the fit is proven, a January 1 launch.
03 · Location-data scrubbing
The image can prove the place without exposing it.
A phone photo may carry latitude, longitude, device, timestamp, and camera metadata that never appears on screen. That hidden payload is exactly what can endanger a field team when media reaches donors or the public.
Server-side public-media path
01Ingest
Reporter upload today, cloud library import tomorrow.
02Inspect
Read faces, consent signals, EXIF, and GPS.
03Separate
Keep useful coordinates as structured internal data when needed.
04Sanitize
Create a stripped derivative; apply the required face redaction.
05Serve
Donor and public pages receive the sanitized artifact.
Why this mattered before Dropbox
Browser uploads already removed much metadata client-side, but that is not a sufficient security boundary. Cloud imports bypass the reporter's browser, so protection had to live in Mirage Core before ATP's existing media library could safely enter the same pipeline.
The careful split
Coordinates may help verify that evidence was captured near a project. Mirage can preserve that analytical value internally while ensuring the donor-facing file does not carry the coordinates that produced the insight.
04 · Multimodal intelligence
"Understand" sits at the center of all four.
Language is not a fifth file type. It is the intelligence layer that crosses photo, video, audio, and text. Every submission is read for what it says and for what it quietly carries, so a reporter under pressure does not have to explain the context in writing.
Any language · Any place
Understand
Mirage transcribes speech, translates meaning, and identifies the language and location behind every submission - including the technical details a photo or video carries on its own. Context reaches your steward with the evidence, even when no one had time to type it.
PHOTO
See
Frames become observations, privacy decisions, and metric evidence.
VIDEO
Watch + hear
Visual analysis pairs with the spoken track and its transcript.
AUDIO
Listen
Testimony becomes searchable, reviewable text without losing the voice behind it.
TEXT
Read
Written reports become structured claims and proposed metrics.
Why this matters in the field: a reporter can send a photo, a voice note, or a few words in their heart language and still produce a complete record. Mirage supplies the language and place context, then ATP's steward decides what is accurate and what is safe to publish.
05 · Human control
Automation proposes. The steward governs.
The architecture is intentionally asymmetric: AI can accelerate review, but the safest instruction wins and public delivery remains a human-governed outcome.
The steward's authority
Show it. Blur it. Withhold it.
Reporters capture consent at the source. ATP stewards can confirm or correct permission, keep an asset internal, require all faces to be blurred, reject extracted metrics, and decide what becomes donor-facing proof.
- Minor detectedWithhold until an approved sanitized copy exists.
- Blur all facesHide immediately; republish only after full redaction completes.
- Internal onlyKeep the evidence inside the steward workflow.
- Video safetyFlagged video is withheld from public delivery. Face blurring for video is on the backlog and may be released in the future.
06 · Proof that the protection works
No real children required to test the system.
The test harness uses documented, synthetic people and known face scenarios. It exercises the real pipeline, gives administrators before-and-after evidence, and turns past privacy failures into regressions that future code must defeat.
16/16
Live scenarios passed
Children, adults, mixed groups, distant faces, ambiguous ages, blur-all cases, false-positive controls, and a GPS-bearing image.
4/4
Deterministic tests on main
Blur applied, control regions untouched, EXIF/GPS removed, and clean-photo stripping protected in CI without using real ATP field content.
Control
Technical behavior
Executive meaning
Fail-closed minor policy
Bad AI output, an unauthenticated callback, or an incomplete minor result cannot become "clean."
Uncertainty produces a blank space, not an exposed child.
One serving chokepoint
Donor reports and other public paths share the same media-resolution policy.
A new page cannot casually bypass the privacy rule.
Synthetic regression battery
Known scenarios run on demand in Admin and automatically in CI.
ATP does not become the test-data supplier for its own safety controls.
07 · Optional field incentives
A provision, not an ATP promise.
Mirage can support a bounty attached to a particular project or reporting task. That is a configurable participation mechanism, not a default fee, an ATP charge, or a commitment that Mirage will fund reporters.
Where incentives fit
Use them when participation is outside the job.
An employee or established associate is typically an affiliated reporter contributing through the organization's normal workflow. A community witness, partner, or independent participant may be unaffiliated, making a project-funded incentive useful for a specific evidence request.
- Organization controlledThe project decides whether an incentive exists and how it is funded.
- Task specificThe provision attaches to a defined reporting need, not every submission.
- Nothing presumedThe amount and the funding source are ATP's call. Neither is assumed or billed by default.
- Separate from verificationPayment motivates participation; steward review and evidence quality determine what counts.
08 · Road to January 1
Earn the business by reducing the burden.
ATP should not have to staff a new program before Mirage proves it can help. The runway through year-end is designed to use work ATP and its partners already perform, give Malia useful leverage when capacity allows, and make the January decision evidence-based.
Keep the tools available
ATP retains access without a subscription obligation. When useful, Matt works alongside Malia on one priority project using already approved content, rather than asking ATP to launch a broad reporting program.
PROOF: USEFUL OUTPUT, LOW ATP LIFT
Reuse what already exists
With ATP's permission, connect approved Dropbox media and test how ECFA, Broken Shepherds, and public press-release data can reduce duplicate entry and content production.
PROOF: HOURS SAVED + ONE EVIDENCE CHAIN
Review the fit with Chaz
Once ATP's direction and staffing picture are clearer, review what Mirage produced, what Malia actually had to do, and whether a sustainable internal owner and project cadence now exist.
PROOF: VALUE, CAPACITY, OWNERSHIP
Prepare a clean launch
If ATP is ready, confirm the $500/month Standalone plan, name a steward and backup, select the first partner-supported projects, and set a realistic January reporting cadence.
TARGET: JANUARY 1, 2027
One ATP portfolio
Many contributing ministries.
ATP remains the steward of each project while approved collaborators such as Shadow Warriors Project, Sterling Advisory Group, Safe Path Group, PCITF, and in-country operators contribute the evidence closest to the work. Malia and ATP's designated stewards decide what counts, what stays internal, and what reaches donors. Mirage earns the partnership by making that network easier to coordinate without blurring who owns the mission.
09 · What this means for ATP
Six things worth remembering.
July and August built trust infrastructure.
The team followed safety from analysis through storage and all the way to the public bytes a donor actually receives, then extended that foundation into ATP-specific reporting workflows.
Location scrubbing protects both people and programs.
Mirage can use location internally as a verification signal without publishing hidden coordinates to the world.
The design anticipates ATP's archive.
Server-side protection was necessary before Dropbox or other cloud imports, because imported media never passes through the reporter's browser.
Transcription and translation are both core.
Audio and video are not attachments to store. Their spoken content becomes part of the verified, multilingual record, along with the language and place it came from.
Your stewards keep the final word.
Mirage accelerates analysis while ATP controls consent, redaction, internal-only handling, metric approval, and publication.
Incentives are available, never assumed.
A project can attach a bounty when it depends on unaffiliated reporters. ATP decides whether one exists and how it is funded; nothing is charged or presumed.
Video note: face blurring for video may be released in the future. Until then it remains on the backlog.