Use Case Paper — the reflexive instance (a monitor over A–D)
Use Case Paper is the reflexive squeeze loop (instance-paper, kind: monitor,
terrains A/B/C): the paper is produced under the strategy it describes, and the same
loop monitors the four object instances A–D. Its SL-1.0 self-description is
src/paper/instance-paper.sl.json; the companion
write-up with the diagram is
src/paper/docs/use-case-paper.md.
The dominant failure it guards
Coherent-and-wrong here = a manuscript whose prose reads as faithful to the literature while a reported number does not recompute or overclaims a result — including a claim about one of the instances A–D. Fluent, citable, and wrong.
Two roles in one loop
- The deliverable. The manuscript (
tex/paper.tex) + the claim ledger, written by the writer agent. - The monitor over A–D. The verifier re-derives each ledgered claim — including every
claimed result of instances A–D — from the hard ground truth and the instances'
returned results, never from the writer's prose. The four
sub_loopbridge rows link down toinstance-a..dby their returned results (the mediated channel), never their internal sources — so the monitor never bounds itself on an instance-internal artifact (the cross-loop disjointness rule).
The bounds
- Upper bound
U— the read literature (every cite has an archived read record) + the paper-spec (the authored normative ceiling). - Lower bound
L— the recompute + build harness (every number recomputes from a generator, the manuscript compiles, every cite resolves) over the raw repository ground truth and the instances' returned results.
The cast (ten actors)
| Actor | Role | Builds | Must not see |
|---|---|---|---|
| Main-thread coordinator (the only judge) | coordinator | approvals, gate verdicts, ledger reconciliation, the .sl.json |
— |
| Literature agent | property author | the paper-spec + the read-before-cite guard | the manuscript, the ledger |
| Writer agent | implementer | the manuscript + the claim ledger | the verifier's re-derivation |
| Verifier agent | exerciser | the independent re-derivation of every ledgered claim | the manuscript, the ledger |
| Bench agent | probe | the recompute/build harness that pins every number to its generator | the manuscript, the ledger |
| Bridges → instance-a..d | sub_loop | each object instance collapsed to one row; returns its machine-verdicted results | — |
| External / human reviewer | human terminus | the terminal editorial judgment closing the reflexive soft side | — |
The barrier and the terminus
The verifier (and bench) never see the writer's manuscript/ledger and re-derive independently; the monitor reads only the instances' returned results via the bridges, never their internals. The coordinator's editorial judgment — is the paper a faithful reading of the literature? — is the one soft-side hole, and it is closed at the human/external reviewer terminus, by disjoint authority, not mechanically. A claim that the automation alone closed the soft side would itself be the coherent-and-wrong this loop exists to catch.
The gates
- Gate A (editorial). The coordinator judges each bridge / item before accepting it.
- Gate B (machine). Every number recomputes from its generator; the manuscript builds; every cite resolves to bib + an archived read record.
- Gate C (coverage / no-blend). The claim ledger reconciles with the reflexive macros (no ledgered claim silently skipped); prose never stands in for a recomputed number.
Disjointness
The reflexive cast holds disjoint bases: lit holds the literature (never the manuscript); writer holds the spec + literature + the build (never the verifier's evidence); verifier re-derives from the hard truth and the instances' returned results (never the writer's prose); bench pins the hard truth from raw repo data (never the writer's claims). The four bridges link to A–D by returned results, not internals. The coordinator's editorial judgment is closed at the human reviewer terminus.
python config/skills/sl-internal/scripts/sl_disjointness_check.py src/paper/instance-paper.sl.json
(0 FAIL / 31 checks.)