🚀 Building MyZubster: What if AI preserved uncertainty instead of inventing missing data?
I'm building MyZubster, and I'm using a surprisingly simple real-world experiment to explore a much bigger software problem: milk kefir and community knowledge.
The question behind the project is:
How can we transform personal experience into structured community knowledge without turning experience into “verified truth”?
The problem
Imagine a user submits:
“I leave my kefir in the refrigerator for a few days.”
An AI system could easily transform that into:
duration: 72 hours
temperature: 4°C
It looks structured.
It also contains information the user never provided.
For MyZubster, that is a data-integrity problem.
Instead, the system should preserve:
duration:
reported: "a few days"
exact_value: UNKNOWN
temperature:
environment: refrigerator
measured_value: UNKNOWN
UNKNOWN is not an error. UNKNOWN is data.
It tells us what should be measured during the next real experiment.
Evidence-aware Knowledge Cards
I'm building the first version as a Git-based Community Knowledge Library.
Every practice receives a stable ID:
KF-001
KF-002
KF-003
...
The current repository contains cards for:
- base milk-kefir practice
- strained kefir
- whey reuse experiments
- pizza, focaccia and baking experiments
- dehydrated kefir mixtures
- a real community pilot
- a reusable community contribution template
But these aren't treated as automatically verified recipes.
Each card carries an evidence state:
PERSONAL_PRACTICE
TRADITIONAL_PRACTICE
OBSERVATION
EXTERNAL_SOURCE
VERIFIED_GUIDANCE
A contributor saying:
“I did this.”
is useful information.
But it is not equivalent to:
“This has been scientifically verified.”
That distinction is part of the data model.
AI should normalize — not manufacture certainty
The future pipeline looks like this:
community submission
↓
preserve raw contribution
↓
AI-assisted normalization
↓
mark missing fields UNKNOWN
↓
validate against invented data
↓
assign KF-###
↓
generate Knowledge Card
↓
community reproduction
↓
observations
↓
external evidence
The AI contract is intentionally constrained:
DO:
extract
structure
connect
identify missing information
suggest what to measure next
DON'T:
invent
erase provenance
silently upgrade evidence
confuse inference with observation
One of the validation mechanisms I'm exploring is detecting transformations such as:
SOURCE:
"a few days"
GENERATED:
"72 hours"
→ UNSUPPORTED_PRECISION
or:
SOURCE:
"kept in refrigerator"
GENERATED:
"4°C"
→ UNSUPPORTED_PRECISION
Other potential validation states:
UNSUPPORTED_CAUSATION
UNSUPPORTED_SAFETY_CLAIM
MISSING_PROVENANCE
MISSING_CONSENT
EVIDENCE_STATE_ESCALATION
SOURCE_NOT_TRACEABLE
Why Git first?
Right now, the knowledge layer is intentionally simple: Markdown + Git.
That immediately gives the project:
human-readable data
version history
diffs
stable files
commit history
auditability
portability
If somebody changes:
temperature: UNKNOWN
to a measured value after a real experiment, the history remains visible.
Git isn't necessarily the final database.
It's a very useful first knowledge ledger.
From recipes to a knowledge graph
The interesting part is that these objects aren't isolated.
For example:
milk + culture
↓
kefir
↓
draining
↙ ↘
thick whey
fraction ↓
↓ reuse
spread ↓
↓ experiment
└──────┬──────┘
↓
observation
↓
community knowledge
Eventually this can become a graph:
Contributor
│
│ contributed
▼
KF-002
│
│ produces
▼
Whey
│
│ reused in
▼
KF-003
│
│ reproduced by
▼
Participant B
And evidence can eventually move from card-level evidence to claim-level evidence.
The bigger idea
Kefir is only the pilot.
The same architecture could potentially represent community knowledge around:
fermentation
gardening
seed saving
cooking
repair
craft
maker projects
local observations
The underlying problem remains the same:
How do we preserve useful human knowledge without erasing where it came from or pretending it has stronger evidence than it actually does?
My current rule for MyZubster is:
Don't manufacture certainty. Preserve provenance, record observations, expose unknowns, and let evidence accumulate.
The experiment started with a jar of kefir.
Now I'm using it to explore what an evidence-aware, AI-assisted community knowledge system could look like.
🔗 Project:
https://github.com/DanielIoni-creator/Myzubster-fermentation-kefir
I'd especially like feedback from developers working on AI data pipelines, knowledge graphs, provenance, community platforms, or evidence-aware systems.
How would you design the boundary between AI-assisted normalization and unsupported inference?
