A 9.5 GB local quant (Qwen3.6-27B IQ2_XXS) works through a real annual-report analysis
(FinQA, Entergy 2008 — mechanically selected, criteria in _data.js) end-to-end, through the real
MCP surface and the real engine:
- Act 1 — piece-by-piece: a typed plan; each step served with only its bounded context.
- Act 2 — repair + external think mode: the certified method stock (
finqa-stock-q6.sgc, 23 KB, sha256-dossiered) orients the low-quant's output; an out-of-stock trap is refused with the reason and the admissible options — never silently guessed; a forced write is recordeduntrusted, never admitted. - Act 3 — memory that reopens: an erratum retracts its consequences in cascade and re-derives selectively at 0 model calls; a withdrawn value reopens its dependent tasks with the reason (JTMS).
- Act 4 — crash / replay: bit-identical replay at 0 calls; a corrupted checkpoint is rejected fail-closed.
# Replay the recorded film — deterministic, NO model, NO GPU, self-contained:
node examples/integrated-demo/run.js --replay
# Live run (records a fresh transcript) — needs a local GGUF:
# put Qwen3.6-27B-UD-IQ2_XXS.gguf under ./models (or DEMO_MODEL=<path>), then:
node examples/integrated-demo/run.js # add --quick for 4 steps instead of 8The replay verifies 7 checks (4 acts · zero ungated results in the synthesis · trap never
admitted · 0-call selective re-derivation · REOPEN with reason · bit-identical replay · corrupted
checkpoint rejected) and prints the verdict table with per-step provenance. Every check is
derived from a run signal that can actually fail — none is a constant of the script — and
tests/unit/integrated-demo-checks.test.js pins a negative control per check (each bar shown red
on its failure shape).
mission-data.json is the serialized mechanical selection, so the 75 MB FinQA dataset is NOT
needed; to regenerate it from source, set FINQA_DATA=<dir containing train.json>.