📖 Harness Path — The Litreview Runtime

15 · the Programme lit-review runtime — novelty against an adversary — an entry in the origin fleet catalogued in The Harness Paths (group: Adjacent embodiments). The catalogue is the operator's work WITH the harness, not part of the system you adopt; none of its code lives in the harness repo. This is the map entry, so you can see what a filled config looks like at this weight.

(private instance of the author — inside the research corpus; the V6
pipeline's descendant, turned on related work. Its epigraph is the whole
design: "a literature review that only finds agreement has not been run
against an adversary.")

The objective is a defensible novelty claim, and the operative object is
the residue — what remains of a claim after covering art is subtracted.
Run 03 seats eleven: 5 sweep ⊥ 5 refute ⊥ 1 judge. The sweep bench
searches prior art by modality; the refute bench hunts covering art
per-residue and is instructed to default to coverage; the judge rules
SURVIVES / NARROWED / COVERED. The hold-apart is context isolation by
construction — a refuter receives the residue and the sweep digest but
never the prover's argument (D3: separation must be structural; a run
separated by prompt alone is invalid). D2 forbids the domain's oldest
overclaim: absence of evidence is reported as not_found_in_corpus(n, range), never as novel. And the honest limit is declared in the manifest
rather than hidden (D4b): all seats run the same weights, so an
all-SURVIVES result is treated as a failed enforcement of the adversarial
discipline — the run is trusted not by its result but by whether the
adversary ever knocked anything down. It did: run 02 returned 0 VALIDATED /
5 MIRAGE; run 03 (76 sweep items, 48 covering candidates) returned 1
COVERED / 4 NARROWED / 0 SURVIVES, retiring one claimed novelty outright
and shrinking the other four to their defensible cores; run 04 added the
normative strand and per-paper placement. Verdicts are candidates only —
adoption and the terminal gate are the First Person's, and the manifests
are generated artifacts: fix the data, regenerate, never hand-edit.

At a glance —
objective: residues surviving refutation — the defensible core of a novelty claim, honestly shrunk · gate: the refute bench + judge; verdict lexicon SURVIVES / NARROWED / COVERED, confidence never entering tier-A artifacts (GR-2) · hard constraint: D2 — absence is not_found_in_corpus(…), never novel
the Gap: context isolation as the hold-apart — refuters never see the prover's argument (D3, structural); the draw is not hashed, and the manifest says so · lenses: five search modalities ⊥ five per-residue refuters, judge as the only barrier · canary: inverted and declared — same weights on all seats (D4b), so an adversary that never lands a hit indicts the run, not the claims
weight: full, non-numeric · private, inside the corpus · the lesson it carries: in research the mirage is agreement — the valuable verdict is the one that kills your claim, and a review that returns all-clear has failed, not succeeded.

Navigation

← Welcome Visitors · The Harness Paths · Adopting The Harness · Workflows And Runtimes

the wiki is a projection; ~/dual-agent-harness (git) is the source of truth · (⚔️⊥⿻⊥🧙)😊 = neg ⊕ bnot → succ