We started with four documents and no code. By the end, there were 41 passing tests and the entire protocol core committed.
The interesting part wasn’t the code. It was the pushback.
Where the protocol gets tested
Twice during this session, something forced a pause.
The first was a bug in domain tag validation. Whitespace-only tags were passing as valid when they should have been rejected. The tests caught it — exactly what they’re supposed to do.
The second was more important. After the reputation floor issue surfaced, the question came back: was the fix actually a protocol decision, or just a patch to make tests pass?
That distinction matters.
Reputation is not a shortcut
The ReputationFloor exists to protect established validators from decay pushing them below eligibility. It is not a starting baseline.
Clamping a new validator’s score to 0.15 would make them appear eligible before earning anything. That directly contradicts the spec.
The fix was not a convenience. It was alignment with the protocol.
The spec is the source of truth
Early sessions are where mistakes get baked in if they’re not caught.
The temptation is always to move fast and smooth over edge cases. But the spec is the standard. If code and spec disagree, the spec wins.
If there is a valid reason to deviate, it gets written down and tracked. Otherwise, the system drifts without anyone noticing.
Consensus without unnecessary complexity
One open question was normalization in the consensus weight formula.
We kept it as late normalization — aggregate in raw weight space, and normalize only at the decision boundary. The majority check is relative, not absolute.
winnerMass / totalMass already gives the correct comparison.
No pre-normalization. No extra passes. Same result, cleaner system.
What exists now
The core pieces are in place:
- The claim schema and lifecycle state machine
- The PoV vote weighting engine with bias correction
- The reputation engine with all six post-finalization update cases, including decay
Everything inside core/ is isolated — no external imports, no networking, no chain awareness.
Pure protocol logic: inputs in, outputs out.
What comes next
The next session moves into two areas:
core/bias, where bias detection feeds into the consensus engine- The simulation harness, where the system is tested under collusion pressure at scale
That’s where the math either holds or breaks.
PulseSyn is built in public at https://github.com/Baniloo-Labs/pulsesyn