Plan, apply and roll back changes to a production ontology, with a blast radius report and a proof an auditor can re-check.
Rust ai-nativedescription-logicsformal-verificationinfrastructure-as-codeisabelle-languageknowledge-graphleanlean4mcpmcp-serverontologyowlowl2reasoningrustshaclsparqltheorem-proving
fix(demo): make demo-verify runnable on a fresh clone with no model
FABIOTESS10bdb994578c
diff --git a/Makefile b/Makefileindex 610d35d8..62bd4fdd 100644--- a/Makefile+++ b/Makefile@@ -1,4 +1,4 @@-.PHONY: build test lint audit bench bench-pizza bench-ontoaxiom bench-mushroom bench-vision bench-reasoner bench-oaei docker clean demo demo-verify+.PHONY: build test lint audit bench bench-pizza bench-ontoaxiom bench-mushroom bench-vision bench-reasoner bench-oaei docker clean demo demo-verify demo-verify-pipeline# ─── Development ─────────────────────────────────────────────────────────────@@ -90,11 +90,31 @@ demo:python3 demo/bundle_fixtures.py --in demo/precomputed --out demo/precomputed/bundle.jsoncd demo/precomputed && shasum -a 256 *.json | sort > MANIFEST.sha256+# demo-verify is the command a sceptic runs: it must work on a fresh clone,+# with no model server, no engine process, and no prior `make demo` run.+# demo/verify.py needs demo/derived/_ontology.ttl (written by `make demo`,+# which calls a local model) and a running engine over MCP, neither of which+# a fresh clone has, so it is NOT part of this target -- see+# demo-verify-pipeline below. Everything demo-verify does runs the+# conformance script and its tests first (dcat_conformance.py has no network+# access and no model call of its own), so the published numbers in+# demo/precomputed/findings.json are re-derived, not merely checked as bytes+# that have not moved; the checksum check after it additionally guards every+# other precomputed artifact against silent drift.demo-verify:- python3 demo/verify.py- cd demo/precomputed && shasum -a 256 -c MANIFEST.sha256python3 demo/dcat_conformance.pypython3 -m pytest demo/tests/test_dcat_conformance.py -q+ cd demo/precomputed && shasum -a 256 -c MANIFEST.sha256++# demo-verify-pipeline re-checks the document-to-ontology pipeline itself+# (parse, stats, lint, enforce, vocab, reason) against whatever is currently+# under demo/derived/. Unlike demo-verify, this needs state that only `make+# demo` produces (demo/derived/_ontology.ttl) and a running engine reachable+# over MCP -- ontology_from_docs.py's call_model() also expects a local model+# server at ONTO_LLM_BASE_URL (default http://localhost:8081/v1). It is a+# separate target precisely so a fresh clone can still run demo-verify.+demo-verify-pipeline:+ python3 demo/verify.py# ─── Cleanup ─────────────────────────────────────────────────────────────────diff --git a/demo/README.md b/demo/README.mdindex 0ad86683..c233346a 100644--- a/demo/README.md+++ b/demo/README.md@@ -90,8 +90,26 @@ make demo`make demo` runs the pipeline end to end against the corpus in`demo/corpus/dcat-us`: read, tokenise, derive, merge, reconcile, populate,-verify, and scan for contradictions. Each stage can also be run on its own,-for example:+verify. It does not scan for contradictions as part of this target;+`contradiction_scan.py` is a separate script, targets a different bundle+layout than this corpus uses, and is not wired into `make demo` (see the+Makefile's own comment on the `demo` target). What `make demo` does call --+`ontology_from_docs.py`'s own Stage 6 and `verify.py`'s Check 6 -- both search+for disjointness contradictions and find none on this corpus by construction;+see "The conformance finding" below for what this demonstration actually+reports instead.++`make demo` calls a local model server for the DERIVE, POPULATE and REFINE+stages: `demo/ontology_from_docs.py` talks to `ONTO_LLM_BASE_URL` (default+`http://localhost:8081/v1`) and to the engine binary over MCP at `ONTO_ENGINE`+(default `http://127.0.0.1:8137`). Neither is started for you; without both+reachable, `make demo` fails partway through STAGE 3/4/5. `make demo-verify`+needs neither: it re-derives the published conformance figures from+committed inputs only, so it is the command to run on a fresh clone.+`make demo-verify-pipeline` re-checks the pipeline's own parse/stats/lint/+enforce/vocab/reason state under `demo/derived/`, and does need both a prior+`make demo` run and the model server above; it is a separate target for+exactly that reason. Each stage can also be run on its own, for example:```bashpython3 demo/ontology_from_docs.py --corpus demo/corpus/dcat-us@@ -113,13 +131,14 @@ pipeline's contradiction scanner detects provenance-split typing conflictsdisagreement is not that shape: it is a README's conformance claimcontradicted by the artifacts published alongside it. That finding isestablished by a validator, not the model, and lives in-`demo/dcat_conformance.py`. `make demo-verify` runs it (and its tests) as-part of the verify target, so the command a sceptic runs re-derives the-figures below rather than only checking that committed bytes have not moved.+`demo/dcat_conformance.py`. `make demo-verify` runs it (and its tests) FIRST,+before checking any committed bytes, so the command a sceptic runs on a fresh+clone re-derives the figures below rather than only checking that they have+not moved -- and it needs no model and no prior state to do so.The script reads only files committed under `demo/corpus/dcat-us/`: the fullGSA/dcat-us `jsonschema/definitions/` and `jsonschema/examples/` tree-(`demo/corpus/dcat-us/jsonschema/`, vendored separately from the seven+(`demo/corpus/dcat-us/jsonschema/`, vendored separately from the eightdocuments the pipeline above reads), `recovered-shapes.ttl`, and`recovered-context.jsonld`. Pull request 120 deleted two files, not one:the SHACL shapes and the profile's only published JSON-LD `@context`@@ -165,3 +184,15 @@ which alone moves the real-context count by several more violations ifcorrected. `demo/tests/test_dcat_conformance.py::test_shacl_violation_counts_disagree_across_methods` pins the disagreement,not a winner.++An earlier, already-public figure for the same deleted-shapes-plus-binding+measurement is 287 (`case-studies/dcat-us-binding/README.md`, a separate,+standalone repository's pySHACL run against its own generated binding). That+binding's triple/predicate/DCAT/empty-file counts match this script's+`observed` binding exactly (1,510 / 123 / 228 / 10), but its violation count+does not: 287 there against 272 here, on the same shapes and the same bound+corpus. The gap is investigated, not adjusted: it is not severity filtering+and the two runs reach the identical 228 focus nodes, so the cause is still+open. Combined with this script's own three internal methods (178 / 272 /+147), that is four legitimate-looking counts for what sounds like one+question, and none of the four should be quoted as settled.diff --git a/demo/verify.py b/demo/verify.pyindex dbeb9bd3..250416eb 100644--- a/demo/verify.py+++ b/demo/verify.py@@ -277,12 +277,12 @@ def main() -> int:f"{len(contradictions)} contradictions across documents{OFF}")print(f"{DIM}CHECK 6 only re-derives disjointness contradictions in the pipeline ontology "f"above; it is not wired to this script's exit code, and finds none here by "- f"construction (see docs/superpowers/plans/2026-08-24-studio-public-port.md, "- f"'PIVOT, 25 August 2026'). The findings this demonstration actually reports live "- f"in demo/precomputed/findings.json, computed independently by "- f"demo/dcat_conformance.py and re-derived every time 'make demo-verify' runs -- "- f"see its own output above, or run 'python3 -m pytest demo/tests/"- f"test_dcat_conformance.py' directly.{OFF}\n")+ f"construction (see .superpowers/sdd/progress.md, 'PIVOT, 25 August 2026'). The "+ f"findings this demonstration actually reports live in "+ f"demo/precomputed/findings.json, computed independently by "+ f"demo/dcat_conformance.py. This script (verify.py) does not run that check --"+ f"run 'make demo-verify' (no model, no engine, works on a fresh clone) or "+ f"'python3 -m pytest demo/tests/test_dcat_conformance.py' directly to see it.{OFF}\n")return failures