v0.11.0 — the thesis phases are done, and the harness now catches itself
Phases P1–P5 of the breakthrough roadmap are COMPLETE (31/56 items, 55%). That is
the whole thesis spine:
P1 Honest observations 5/5 ✅
P2 The verified loop 3/3 ✅ "done" is a fact, not a claim
P3 Measurement 4/4 ✅ bench proves harness-lift per lever
P4 Context integrity & memory 8/8 ✅
P5 Small-model amplification 8/8 ✅ the thesis, bench-validated
P6 Loop architecture 3/8
P7–P10 0/20
The roadmap's own one-sentence breakthrough was: bench measures harness-lift per
lever per model (P3), the verified loop makes "done" a fact instead of a claim
(P2), and the swarm compiles goals into verified commits from laptop-sized models
(P9). Two of those three legs now stand. P9 — the moat — does not.
Headline feature: the claim guard (P5.9, #42). A model that claimed done, was
rejected, and claimed the identical thing again would do so until the step limit —
qwen2.5-coder:7b did it 33 times, burning 63s and the whole budget to re-send one
sentence. loop_detect watches ACTIONS; nothing watched the completion path. The
cause was starvation, not stubbornness: the bounce named the failing command while
the evidence contract kept only a digest of its output, so the model was told "it
failed" 33 times and never what failed. Now: claim → show the real failure → stop.
3 claims, 18s, an honest ending. It does not make models smarter; it converts a
confused failure into a fast truthful one — and it never opens the gate, because
repetition must not become an escape from the done-gate (H05).
P5.9 is the first roadmap item NOT produced by the 31-agent audit. Thirty-one
agents read every line and could not have found it: it only exists when a real
model runs a real task. A static audit finds what the code says; running it finds
what the code does.
Also in this line since v0.10.0: a real first-run experience, TTY-gated output,
one canonical install URL (topk1.com/forge/install.sh), a mobile-usable landing
site, grouped forge --help, and CLAUDE.md.
723 tests, green on 3.10–3.13.