Season 1 · Episode 4

What the Texts Actually Say

Privacy-safe SMS sentiment analysis over time.

This lesson drops soon — check back here for the video and voiceover.

Years of your own messages are a dataset about your relationships — and a minefield if you handle them carelessly. This lesson shows a privacy-first way to analyze message history: consent and scope first, a strict schema, a cleaned archive, and measurements of only what is observable. The analysis reveals patterns; it never diagnoses a person.

The problem this lesson solves

T

w

o

f

a

i

l

u

r

e

m

o

d

e

s

h

a

u

n

t

p

e

r

s

o

n

a

l

-

d

a

t

a

p

r

o

j

e

c

t

s

.

T

h

e

f

i

r

s

t

i

s

p

r

i

v

a

c

y

:

r

e

a

l

n

a

m

e

s

,

n

u

m

b

e

r

s

,

a

n

d

i

n

t

i

m

a

t

e

c

o

n

t

e

n

t

f

l

a

s

h

e

d

o

n

s

c

r

e

e

n

f

o

r

a

t

u

t

o

r

i

a

l

.

T

h

e

s

e

c

o

n

d

i

s

o

v

e

r

r

e

a

c

h

:

a

n

a

s

s

i

s

t

a

n

t

t

h

a

t

g

u

e

s

s

e

s

m

e

s

s

a

g

e

d

i

r

e

c

t

i

o

n

f

r

o

m

w

o

r

d

i

n

g

,

o

r

r

e

a

d

s

“

r

e

s

p

o

n

s

e

t

i

m

e

i

n

c

r

e

a

s

e

d

”

a

s

“

t

h

e

y

n

o

l

o

n

g

e

r

c

a

r

e

.

”

T

h

e

p

r

o

b

l

e

m

t

h

i

s

l

e

s

s

o

n

s

o

l

v

e

s

i

s

d

o

i

n

g

t

h

e

a

n

a

l

y

s

i

s

r

i

g

o

r

o

u

s

l

y

w

i

t

h

o

u

t

e

i

t

h

e

r

s

i

n

—

c

a

r

e

f

u

l

w

i

t

h

p

e

o

p

l

e

’

s

d

a

t

a

,

h

u

m

b

l

e

a

b

o

u

t

w

h

a

t

t

h

e

d

a

t

a

c

a

n

s

a

y

.

The lesson, step by step

Establish consent and scope — then redact

Analyze only messages you lawfully control. Before anything goes on screen: redact names, phone numbers, intimate content, and third-party details. The channel’s privacy-safe demo recipe applies: copy only the fields needed to teach the method, replace identifiers, coarsen dates when exact timing isn’t essential, preserve the structural issue being demonstrated, label the demo as recreated, and secure the working copy afterward.

Define the schema — direction is a hard field

Timestamp, contact, direction, text, medium, thread. Direction — who sent what — is a hard data field from the message record, not a guess from wording. Getting this wrong flips the entire analysis, so it is validated, never inferred.

Clean the archive without corrupting it

Normalize dates, preserve empty multimedia messages as events (a photo sent is still a message), remove duplicates carefully, and keep original text immutable. Cleaning makes the data analyzable; immutability keeps it honest.

Measure only the observable

Reply latency, initiation balance, question rate, future-planning language, warmth markers, conflict markers, conversation gaps. These are counts and timings — things the archive actually contains. Segment by week or relationship phase, and inspect real examples behind every trend before believing it.

Separate measurement from interpretation

“Response time increased” is an observation. “They no longer care” is an unsupported inference. The lesson draws this line in red: the assistant reports the measurement and stops. The meaning is a human question, not a model output.

Ask for counterexamples and uncertainty

Sentiment classifiers miss sarcasm, inside jokes, neurodivergent tone, and everything outside the archive. Demand the counterexamples: the warm message during a “cold” week, the slow reply during travel. Uncertainty is part of the result, not a footnote.

Close with human judgment

The analysis can reveal patterns; it cannot diagnose a person or replace a conversation. The lesson ends where it must: the data is a mirror with smudges, and the human decides what to do with the reflection.

Worked example

The archive shows reply latency doubling across March. The observable trend is real — the chart doesn’t lie. But the examples behind it tell the story: a work trip, a family emergency, a dead phone. Three different causes, one identical chart shape. The analysis surfaced the question; only the human with context outside the archive could answer it. That’s the lesson’s boundary, demonstrated.

Your takeaway

• Consent, scope, and redaction come before any analysis — and before anything goes on screen.
• Direction is a data field, never a guess; original text stays immutable.
• Measure the observable: latency, initiation, questions, planning language, gaps.
• An observation is not an interpretation — don’t let the model cross that line.
• Patterns inform; only humans conclude. The analysis can’t diagnose a person.

The BRAVE lens

Every episode runs through the BRAVE method: brief the outcome, reveal relevant context, authorize the next action, verify the evidence, and evolve the system.