Working on Soapmode Verification-first systems (W²)
Washington
Joined April 2026
- Tweets81
- Following21
- Followers22
- Likes199
sEcTs retweeted
Top AI Agents Coins by 7 Days Market Cap Growth
@HarmonicAgents - $HARMONIC => +239%
OpenGPU - $OGPU => +55.7%
@PhalaNetwork - $PHA => +44.7%
@autonolas - $OLAS => +42%
@axol_io - $RAXOL => +35.9%
@PaalMind - $PAAL => +32%
@daydreamsagents - $DREAMS => +15.5%
@CrossbowFi - $XBOW => +13.1%
@singularryai - $SINGULARRY => +17%
@Concordium - $CCD => +16.6%
A small example of W² applied to live investment research.
The production filter below explicitly separates:
past price movement from remaining forward opportunity.
In a recent live run, a candidate had already repriced dramatically before discovery.
Rather than automatically rejecting it or chasing momentum, the system reconstructed:
evidence → catalyst → next buyer → token-specific demand → sellable supply → possible next MC regime
The follow-up then revealed a discovery-latency problem elsewhere in the pipeline.
That problem is now a PROMOTE_FOR_TEST / NOT_MERGED patch candidate.
No result becomes a rule simply because it looks convincing once.
Verification before state transition.
A live W² INVEST test surfaced an interesting microcap.
But the useful result wasn’t the coin.
It was discovering that a stronger competing token had already entered the data pipeline — and had never been escalated for deeper investigation.
That distinction matters:
not visible
≠
visible but not recognized
So I traced the discovery path:
created → visible → collected → investigated → retained → delivered
and built a NOT_MERGED patch candidate around reducing the gap between:
earliest defensible discovery
and
actual discovery
without lowering evidence standards.
W² isn’t about making the model more confident.
It’s about making confidence earn its state.
LIVE W² INVEST CASE
TILCAYO surfaced
↓
Rival discovered: TIGRINO
↓
Historical reconstruction
↓
TIGRINO had already been collected
↓
COLLECTED ≠ INVESTIGATED
↓
Discovery latency identified
↓
PATCH CANDIDATE R1
↓
NOT_MERGED
↓
PROSPECTIVE TEST
Failure → Reconstruction → Testable Patch → Evidence → Possible Integration
Most AI discussions focus on generation.
But the real problem seems to be:
when does an output actually count as state?
That’s where a lot of things start to break.
and that’s why I ended up building W².
SOAPMODE.APP
LLMs don’t fail loudly.
They fail by producing something that looks valid,
and getting treated as if it is.
That’s a different kind of failure.
Sent out physical copies of Soapmode.
If one reached you, here’s the landing page:
soapmode.app
