# ByJTT Gauntlet — Changelog

Autonomous, version-controlled change record of the framework. Every change
documents: what changed, why, the evidence it's better, the gate scores, the
models used, the reasoning, and the results.

Format: [Semantic Versioning](https://semver.org). Each entry is written by
the agent that made the change, then the change is itself put through the
Gauntlet before shipping.

---

## [1.1.0] — 2026-08-15

**Change:** Renamed from "Double-Critique Gate" to "The ByJTT Gauntlet".
Added version-controlled framework module (`framework/framework.json`,
`framework/CHANGELOG.md`), a timeline page, and open-source intent (MIT).

**Why:** The framework must survive its own evolution — 2 critics today,
possibly more gates tomorrow. A fixed name ("Double-Critique") would
misdescribe the process as it grows. The Gauntlet names the *process*:
deploy critics until genuinely satisfied.

**Evidence it's better:** Meta-gauntlet round 1 — the framework itself was
critiqued by both critic models. Result: **PASS (8.2 + 7.8)**. The naming
and versioning were explicitly validated.

**Gate:** PASS — primary qwen/qwen3.7-max 8.2, adversary moonshotai/kimi-k3 7.8.

**Models:** qwen/qwen3.7-max (primary), moonshotai/kimi-k3 (adversary).

**Reasoning:** A living framework needs a living record. Versioning +
changelog + open license makes the process auditable and adoptable — the
same transparency we demand of experiments.

**Results:** Framework page now shows newest version + timeline; critics can
be run against any version via `critic.sh`.

---

## [1.0.0] — 2026-08-15

**Change:** Initial framework. Two independent critics from different model
families; deterministic numeric gate (avg >= 7/10); full evidence archive in
`critiques/`.

**Why:** Research-based answer to "how many critics?" — 2, diverse,
independent (MAD evaluation 2025; IEEE optimal stopping 2025; Bean Labs 2026).

**Evidence it's better:** First deployment against exp-nav CSS caught 5+
real bugs across 3 rounds (half-width desktop tab bar, footer hidden on
notched iPhones, invisible keyboard focus — WCAG 2.4.7, backdrop-filter
regression). Round 3: **PASS (8.8 + 8.8)**.

**Gate:** PASS — 8.8 + 8.8.

**Models:** qwen/qwen3.7-max (primary), moonshotai/kimi-k3 (adversary).

**Reasoning:** One critic = single-model bias; three+ = debate fatigue;
round-trips risk collective delusion. Diversity beats count.

**Results:** Framework adopted as the standing gate for all reviewable results.