One real compile, receipts included.

Everything on this page is the actual output of bioprism context compile on the shipped reference world (radiogenomic-integrity-demo-v1, 761 facts) — the compile envelope, the Context Certificate, the Decision Section, and the verify result, embedded verbatim in this page's source. Not a mockup: the certificate digest below is the same c0da17ff… digest pinned byte-for-byte by the three-implementation parity tests.

Every number and field rendered here comes from that embedded data block. The page works without JavaScript; with it, each card also shows the raw JSON it was rendered from.

1 · The query

Compilation starts from a typed decision query, not a prose prompt. This one asks a single question of the world — split_integrity_status — under a declared role, policy, decision time, and a mandatory set of protected tags.

query_id: audit-split-integrity-v1 · schema fiber-query/0.1
role: research-auditor
targets: split_integrity_status
decision_time: 2025-01-01T00:00:00Z — facts sourced after this cut are inaccessible
policy: research-only
distortion_tolerance: 0.0
budgets: max_facts 64 · max_tokens 6000
protected_tags (10): identity split site scanner time specimen preprocessing policy negative_evidence protected — facts carrying these tags must survive selection; a budget smaller than their closure fails rather than truncates

Raw query JSON

(JavaScript renders the raw query here. Without it, the same JSON sits verbatim in this page's embedded data block — view source.)

2 · The selection — 761 in, 11 out

The compiler selected 11 of 761 facts (1.45%) and 6 of 756 factors (0.79%). The other 750 facts were omitted — and accounted for, in the receipt below.

world: 761 factsselected: 11 (1.45%)

What this world does and does not show. On this exact reference world, a 5-hop graph walk and a BM25 retriever at k=11 select the identical eleven facts — the headline benchmark is a deliberate tie. This page is about what ships with the selection (the receipts), not about beating baselines; the benchmarks page carries that story.

The eleven facts

fact.cohort identity

provides cohort_id = RG-DEMO-001
provenance manifest/cohort.json

fact.subject_aliases identity protected

provides subject_aliases: S001 → S001, ALT-77 · S002 → S002 · S003 → S003, ALT-77
provenance manifest/subjects.csv

fact.split split protected

provides split_assignment: S001 train · S002 train · S003 test · S004 test
provenance analysis/split.json

fact.site site protected

provides site_assignment: S001 A · S002 A · S003 B · S004 B
provenance manifest/site.csv

fact.label_source time label_lineage protected

provides label_source_time: S001 2024-01-01 · S002 2024-02-01 · S003 2025-06-01 · S004 2024-02-15
provenance labels/lineage.json

fact.decision_cut time protected

provides training_decision_time = 2025-01-01
provenance run/config.json

fact.preprocess_fit preprocessing protected

provides preprocess_fit_scope = all_subjects_before_split
provenance pipeline/preprocess.yaml

fact.specimen_dates specimen time protected

provides specimen_dates: S001 2023-12-01 · S002 2024-01-15 · S003 2025-05-20 · S004 2024-01-20
provenance specimens/manifest.csv

fact.scanner site scanner

provides scanner_assignment: S001 X1 · S002 X1 · S003 Y9 · S004 Y9
provenance imaging/dicom_manifest.csv

fact.policy policy protected

provides data_policy = research-only, no-identifiable-export
provenance governance/policy.json

fact.negative_duplicates negative_evidence identity protected

provides declared_duplicate_screen = passed_by_filename_only — negative evidence: the screen that claimed to pass
provenance qc/duplicate_screen.txt

The 6 selected factors (of 756)

factor.identity_check (deterministic_rule) → identity_leakage
factor.site_check (deterministic_rule) → site_leakage
factor.temporal_check (deterministic_rule) → temporal_leakage
factor.preprocessing_check (deterministic_rule) → preprocessing_leakage
factor.policy_check (deterministic_rule) → policy_validity
factor.claim_support (decision_rule) → split_integrity_status

(JavaScript renders the full factor records here; they are verbatim in the embedded data block.)

How the compiler got there — the pass ledger

From bioprism context explain, backend backward_factor_slice_reference:

protected_closure — retained 11 · “10 protected tags requested”
backward_slice — retained 6 · “17 variables reachable from targets”
policy — retained 11 · “1 clause(s) in force, 0 candidate(s) declared a requirement, 0 withheld”
temporal_cut — retained 11 · “0 facts withheld at the decision cut”
oracle — retained 4 · “status invalid”
plan_selection — retained 11 · “backend backward_factor_slice_reference retained 0.0145 of facts; portfolio costed 3 candidate(s), argmin faq_inside_out at 2.1439e3 against a baseline of 3.7240e6 (1737.0x); no costed plan is runnable: faq_inside_out declined: factor "factor.claim_support" carries no table; a world-derived region must be given a valuation before it can be executed”
influence_bounds — retained 0 · “0 of 0 withheld fact(s) bounded, 0 group(s) informative, worst informative bound none”

Seven passes were deferred, each with a stated reason (obstruction tests, abstract interpretation, role-and-purpose filter, information-flow export, decision quotient, rate-distortion, adaptive acquisition) — the reference schemas do not yet carry the bindings they require. The reasons are verbatim in the embedded data block:

(JavaScript renders the deferred-pass reasons here; they are verbatim in the embedded data block.)

3 · The receipt — the Context Certificate

The certificate (fiber-context-certificate/0.1) is the machine-readable receipt: what was kept, what was omitted, why, and under which oracle verdict — hashed over canonical bytes so anyone can recheck it.

Omission accounting

Field	Value
omissions.total_facts	750
omissions.exploratory_facts	750
omissions.classification	`no_backward_dependency_path_or_temporally_inaccessible`
omissions.inaccessible_selected_before_cut	`[]` — empty: nothing selected was time-inaccessible

The explain plan groups all 750 omissions under one manifest entry: reason “no backward dependency path to any target under the declared factor graph”, influence zero, bound 0.0. The envelope's sufficiency flag — supports_sufficiency_claim: true — rests on that accounting.

Compile plan

Field	Value
plan.backend	`backward_factor_slice_reference`
plan.compiled_fact_count / total_fact_count	11 / 761
plan.compiled_factor_count / total_factor_count	6 / 756
plan.max_selected_factor_arity	5
plan.fallback	`null`

Protected closure — 11 facts, satisfied

Every fact reachable from the 10 protected tags survived selection (protected_closure_satisfied: true). The closure is the full selected set:

fact.cohort · fact.decision_cut · fact.label_source · fact.negative_duplicates · fact.policy · fact.preprocess_fit · fact.scanner · fact.site · fact.specimen_dates · fact.split · fact.subject_aliases

Oracle verdict

status: invalid · oracle deterministic_split_integrity_v1 — the proposed radiogenomic split does not support a valid external-generalization analysis. Four witnesses:

identity_leakage — alias ALT-77 is shared by subjects S001 and S003, which sit in different splits (train, test): the same person is on both sides.

site_leakage — site_by_split: train = A, test = B: split is confounded with site.

temporal_leakage — decision_time 2025-01-01, but S003's label source is dated 2025-06-01: a label from the future.

preprocessing_leakage — “preprocessing fit used all subjects before split”.

Limitations — verbatim, never hidden

The certificate carries its own limitations, quoted exactly:

“Reference slicer uses dependency reachability and protected tags; it does not yet implement sheaf cohomology, FAQ-width optimization, abstract interpretation, or formal influence bounds.”

Hashes

world_sha256: b3809731cf93040fcd8aef43deb2a552492064b49154e07ea58caa724c10cbb5
query_sha256: d06368c62dd878f483d30ebfd56f38426b1234d37e46ca77d4536db81f59c684
decision_section_sha256: 7439b2262c52c1c794b59be86d922b723a2ea5646362d529f57fb11b5f7e93ce
certificate_sha256: c0da17ffc80465258345c8a538171bfd868100cd883e9a20780a0dc5477e7ea4

Raw certificate JSON

(JavaScript renders the full certificate here. Without it, the same JSON sits verbatim in this page's embedded data block — view source.)

4 · The check

The receipt is only worth something if a third party can recheck it. bioprism context verify recomputes the digest from canonical bytes — without linking the engine that produced the certificate — and exits non-zero if it does not verify. The verbatim result for this certificate:

ok: true · verification: digest verifies · schema fiber-context-certificate/0.1

Raw verify JSON · raw compile envelope

(JavaScript renders the verify result here; it is verbatim in the embedded data block.)

(JavaScript renders the compile envelope here; it is verbatim in the embedded data block.)

To reproduce from a checkout of the repository (the world and query on this page ship as fixtures):

bioprism context compile --world fixtures/fiber-v0.1/radiogenomic_world.json --query fixtures/fiber-v0.1/leakage_query.json --certificate-out cert.json --section-out section.json
bioprism context verify --certificate cert.json

Run it yourself. Every artifact above regenerates from the shipped fixtures with the two commands just shown — install AURORA Agent (Claude Desktop bundle, Claude Code plugin, or cargo build from source), then compile and verify. The certificate digest you get should be c0da17ffc80465258345c8a538171bfd868100cd883e9a20780a0dc5477e7ea4.

Research and developer infrastructure: it does not diagnose an individual, recommend treatment, triage care, enroll participants, or claim medical-device functionality.