CAPAS INTELIGENTES

Verify it yourself — free, neutral infrastructure (don't trust us, reproduce it)

A claim is only worth the badge if a third party can reproduce it. Every public claim here is checkable on infra we don't control:

• Conformance on neutral CI, signed. GitHub Actions runs python3 benchmarks/conformance.py on every push and Sigstore-signs the result into a public transparency log (keyless). The badge above is that run — green means the load-bearing invariants hold on GitHub's runners, not just a laptop.
• Run a claim yourself — the Binder / Colab badges open notebooks/run_a_claim.ipynb: install nothing, run the gate in your browser, get the same verdict and audit hash we publish.
• Independent supply-chain score — the OpenSSF Scorecard badge is a third party's assessment.
• Every number is re-derivable — see docs/CLAIMS_REGISTRY.md; each claim carries the command that reproduces it. Governance: GOVERNANCE.md.

Your turn (account-gated, one-time): connect the repo to Zenodo + cut a release → mints a citable DOI (.zenodo.json is ready); add a PyPI Trusted Publisher → pip install goes live via publish.yml; push spaces/ to a Hugging Face Space → a neutral hosted demo; register at bestpractices.dev for the OpenSSF Best Practices badge.

CAPAS does not determine truth. It evaluates whether supplied evidence licenses a specific claim under a declared admissibility contract — deterministically, with no language model in the verdict, and re-derivably (same input → same verdict → same audit hash). A claim + its structured evidence goes in; ACCEPT / REWRITE / REJECT / HOLD comes out.

CAPAS ships open-core + open standard (Apache-2.0). The defensible asset is the schema, the admissibility calculus, the certificate, and the benchmark corpus — not the code. It is a reference implementation in the lineage of OPA / SPDX / RO-Crate, not a closed SaaS. The CAPAS name, certification, and official benchmark are reserved (see NOTICE).

It gates its own claims: every public number is CLOSED (proven by a test), BACKED (regenerates from a command), or SCOPED (a declared estimate) — see docs/CLAIMS_REGISTRY.md. Nothing bare.

What is validated, as of 2026-10-13 (and what is not)

• Blind self-gating corpus (real research claims, sister project): 233 claims coded blind by a separate session; engine frozen at a git tag before coding; verdicts sealed by sha256 before outcomes were opened; double-counted. Blind X1–3: bad-not-ACCEPT 53/82, good-not-REJECT 52/63. Out-of-sample X4+X5: 17/21 and 16/17 (Wilson 95% [0.60–0.92], [0.73–0.99]). Outcomes are the project's own later corrections, not an external oracle. python3 research/ginibre_corpus/score_sealed.py
• Not a fraud detector, measured: on 888 clean control papers vs 6,348 retracted-for-data papers, recomputable numeric inconsistencies did not separate the groups (2.13% vs 1.58% severe). python3 research/retraction_audit/grade_severity.py
• Cross-family counter-signature (GOVERNANCE §6): every public claim was attacked by a model of another family (8 runs); 3 of 7 claims were refuted the same day and each fix is frozen as a test or a sealed file. Verbatim transcripts in research/section6_package/. Nothing is marked proved.
• Physical law for quantum calibration (Atlas): calibration_snapshot separated 4/4 real GHZ runs (GOOD vs DEGRADED) blind, and prospectively chose a layout that halved the per-layer error on ibm_kingston with identical circuits. benchmarks/test_calibration_snapshot.py
• Known limits: the in-browser JS replica is not the engine (parity 130/233; offline → HOLD); the GIGO ceiling stands (a self-consistent fabrication passes); a declared field is the coder's responsibility — the gate does not read prose.

It is not a new provenance standard, benchmark suite, workflow engine, or VVUQ methodology. CAPAS is a profile/costurero over existing standards:

RO-Crate/PROV packaging + sealed route/result trace + VVUQ-style physical evidence + witness independence + honest no-evidence/failure/rejection states.

Who This Is For

CAPAS is for people auditing AI-generated scientific-computation outputs:

• AI-for-science agent benchmark builders,
• scientific workflow provenance / RO-Crate users,
• quantum many-body verification and benchmark users,
• teams deciding whether a computation supports a strong claim, a weaker rewrite, rejection, or hold.

CAPAS is not for users who only need a faster simulator or generic workflow lineage.

Hybrid Pipeline Role

CAPAS is the deterministic middle of a hybrid AI-for-science verification pipeline:

LLM / extractor / scientific verifier upstream
        -> claim + evidence JSON
        -> CAPAS deterministic claim gate
        -> ACCEPT / REWRITE / REJECT / HOLD

CAPAS now includes an explicit upstream MVP for retrieval/parsing/alignment, but the final decision remains deterministic. Local corpus retrieval, web retrieval with --allow-web, and PDF parsing can prepare candidate evidence; they do not silently invent missing evidence or turn CAPAS into a broad scientific oracle.

The standalone direction has started with a local upstream MVP:

local text / solver log / code excerpt
        -> explicit evidence extraction
        -> deterministic claim-text alignment
        -> CAPAS claim gate

It is intentionally explicit-only and deterministic. See docs/STANDALONE_PRODUCT_ROADMAP.md.

External Review Packet

For reviewers and potential users:

• Technical note: docs/CAPAS_TECHNICAL_NOTE.md
• Request for feedback: docs/REQUEST_FOR_FEEDBACK.md
• QMB100 / PhysVEC one-pager: docs/CAPAS_ONE_PAGER_QMB100.md
• Global SotA / market audit: docs/GLOBAL_SOTA_MARKET_AUDIT.md
• Hybrid pipeline positioning: docs/HYBRID_PIPELINE_POSITIONING.md
• Input hardening notes: docs/INPUT_HARDENING.md
• Product scope and debt status: docs/PRODUCT_SCOPE_AND_DEBTS.md
• Public demo: https://fomv9354lve.github.io/capas-inteligentes/
• Current release: https://github.com/fomv9354lve/capas-inteligentes/releases/tag/v0.1.1
• Public feedback issue: https://github.com/fomv9354lve/capas-inteligentes/issues/1
• License: Apache-2.0 (open-core); the CAPAS mark + official certification are reserved — see LICENSE and NOTICE.

The current validation question is not "is CAPAS broadly useful?" It is narrower: do the explicit evidence fields help audit or exchange one scientific-agent computation result, or are they already covered by existing artifacts?

Try In 60 Seconds

python -m pip install -e .
capas decide --input examples/external_claim_rewrite.json
capas batch --input examples/external_claim_batch.json --json
capas provenance-check --input examples/external_claim_fine_tune_ready.json --json
capas pipeline --input examples/standalone_pipeline_semantic_hold.json
capas inspect trace_039
capas validate

Expected behavior:

• decide returns REWRITE when local checks pass but a universal anchor fails.
• batch applies the same deterministic gate to multiple claim/evidence objects.
• pipeline can block a numeric ACCEPT when claim.text overstates the scope of the evidence.
• inspect trace_039 shows a motor-backed claim-transition trace.
• validate runs the product gates and profile/RO-Crate checks.

Release assets:

https://github.com/fomv9354lve/capas-inteligentes/releases/tag/v0.1.1

Product Demo

Install the local package entrypoint:

python -m pip install -e .

Run the product surface:

capas demo
python3 benchmarks/verify_capas_product_demo.py

The demo writes:

• outputs/capas_product_demo_report.json
• outputs/capas_product_demo_report.md

It demonstrates the current product contract: CAPAS reads sealed scientific computation traces and emits evidence-typed claim decisions (ACCEPT, REWRITE, REJECT, HOLD) without marking the corpus as fine-tune-ready.

Run the core product validators:

capas validate

Run the local integration API:

CAPAS_API_TOKEN=change-me CAPAS_AUDIT_DIR=outputs/api-audit \
  capas serve --host 127.0.0.1 --port 8765
curl -H "Authorization: Bearer change-me" http://127.0.0.1:8765/health

The API is local and deterministic. It exposes GET /health, GET /decisions, POST /decide, POST /batch, and POST /provenance-check. CAPAS_API_TOKEN enables bearer-token auth; CAPAS_AUDIT_DIR writes workspace-scoped JSONL decision logs keyed by X-CAPAS-Workspace. It is not a hosted SaaS service.

The canonical payload schema is published in the static docs tree:

https://fomv9354lve.github.io/capas-inteligentes/schema/v3/capas_claim_payload.schema.json

Enterprise integration notes live in:

• docs/ENTERPRISE_INTEGRATION_PACK.md
• docs/PROVENANCE_REGISTRY_OPERATIONS.md
• docs/SECURITY_COMPLIANCE_APPENDIX.md
• docs/PILOT_ROI_BUSINESS_CASE.md
• docs/PAPER_INGESTION_CONNECTORS.md
• docs/UPSTREAM_GITHUB_ACTIONS_POLICY.md
• docs/DAILY_SOTA_UPDATE.md

Run CAPAS in GitHub Actions:

jobs:
  claim-gate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"
      - uses: ./.github/actions/capas-claim-gate
        with:
          mode: batch
          input: examples/external_claim_batch.json
          output: outputs/capas-action-report.json

The manual workflow .github/workflows/claim-gate.yml runs a batch smoke check through workflow_dispatch. Other workflows can reuse the composite action directly with mode: decide, mode: batch, or mode: pipeline.

Inspect a trace:

capas inspect trace_039

Decide an external claim/evidence JSON:

capas schema
capas check-input --input examples/external_claim_accept.json
capas decide --input examples/external_claim_accept.json
capas decide --input examples/external_claim_rewrite.json
capas decide --input examples/external_claim_hold.json

Run the standalone upstream MVP:

capas retrieve --input examples/standalone_pipeline_multisource.json
capas retrieve --input examples/standalone_pipeline_web_source.json --allow-web
capas extract --input examples/standalone_pipeline_accept.json
capas extract --input examples/standalone_pipeline_pdf_source.json
capas align --input examples/standalone_pipeline_semantic_hold.json
capas reason --input examples/standalone_pipeline_semantic_hold.json
capas pipeline --input examples/standalone_pipeline_semantic_hold.json
python3 benchmarks/verify_standalone_pipeline.py

This is not broad scientific reasoning. It is local explicit parsing plus a deterministic semantic-scope guard before the CAPAS gate.

extract also returns evidence_spans with source_id, line number, snippet, and parser for each extracted field, so a reviewer can audit where the evidence came from.

Local corpus retrieval accepts JSON, JSONL, text, Markdown, or a directory of those files and returns auditable snippets matched to claim terms and required evidence fields. Web retrieval is opt-in with --allow-web. Local PDF parsing is supported when the optional standalone dependency is installed:

python -m pip install -e ".[standalone]"

If web permission or PDF parser support is missing, CAPAS records that as an extraction note instead of silently inventing evidence.

The published MVP input contract is docs/schema/capas_claim_payload.schema.json. check-input validates the payload shape; decide then checks whether the supplied evidence licenses the claim, needs a rewrite, contradicts it, or must hold.

Generate the static local UI:

capas ui
python3 benchmarks/verify_claim_gate_ui.py

The UI exposes ACCEPT, REWRITE, REJECT, HOLD, and INVALID samples. Structurally invalid payloads are shown as HOLD with schema_errors, matching capas decide. It also supports batch evaluation for JSON arrays or objects with items / claims, displays schema v3 for audit trails, and includes a keyboard help modal with the main shortcuts and pipeline scope.

Schema v3 supports 11 deterministic claim types, including causal mechanism, systematic review, evidence conflict, and multimodal evidence gates.

Prepare a non-mutating GitHub release plan:

python3 scripts/publish_github_release.py

Publishing a new release still requires a GitHub remote, valid gh auth, a pushed tag, a passing external CI run, and a GitHub release URL.

The public v0.1.1 release is available at:

https://github.com/fomv9354lve/capas-inteligentes/releases/tag/v0.1.1

Prepare/check external reviewer feedback:

python3 benchmarks/verify_external_user_validation.py

The feedback template is examples/external_reviewer_feedback_template.json. Completed external feedback belongs under outputs/external_validation/.

What It Produces

The corpus builder emits:

• sealed JSON traces: benchmarks/gold_traces/*.json
• W3C PROV-shaped exports: benchmarks/gold_traces/*.prov.json
• RO-Crate metadata: benchmarks/ro_crates/*/ro-crate-metadata.json
• audit table: audits/gold_trace_audit_template.csv

Evidence Fields

CAPAS adds domain evidence fields over standard provenance:

• physical_evidence_level
• verification_independence
• witness_stack
• reference_truth
• evidenceStatus
• abs_error
• expected
• value
• observable
• local_property_tests
• universal_anchor
• invariant_caught
• claim_scope

The important distinction is that not all traces are training gold. Some traces exist to prove the format can honestly represent uncertainty, failure, or rejection.

Validate The Published Evidence Corpus

Use a Python environment that has the declared public corpus stack available. Check it first:

python3 scripts/check_reproducibility_env.py

The public repository ships the evidence traces and validators. Private/local scientific engines used during exploratory corpus generation are intentionally not packaged in the public release. The public command runs:

1. RO-Crate validation
2. CAPAS physical-evidence profile validation
3. witness independence validation
4. evidence claim validation
5. universal anchor matrix validation
6. audit summary

python3 scripts/build_evidence_corpus.py

coverage_ready=True is expected. fine_tune_ready=False is also expected until blind inference review is completed.

Validate RO-Crate Coverage

python3 benchmarks/validate_ro_crates.py

The validator checks that evidence is present where it should be and absent where it would be dishonest:

• analytic success: present
• cross-sim success: present
• no-evidence success: none_declared
• backend failure: not_applicable_failed
• router rejection: not_applicable_rejected

Validate CAPAS Physical Evidence Profile

python3 benchmarks/validate_capas_profile.py

This checks the local CAPAS profile over Workflow Run RO-Crate-style crates: WRROC profile URIs, ComputationalWorkflow, top-level CreateAction, parameter realization, capas:evidenceStatus, and capas:PhysicalEvidence semantics. This is a local profile validator, not official profile registration.

External RO-Crate Validation

python3 -m pip install -r requirements-validation.txt
python3 benchmarks/validate_ro_crates_external.py

This uses the ResearchObject rocrateValidator package and writes benchmarks/ro_crates/official_validation_report.json. Current crates validate as RO-Crates without warnings. CAPAS emits a recognized .cwl workflow descriptor for packaging and records Python as the implementation language for the costurero.

Positioning

See:

• docs/PROJECT_DASHBOARD.md
• docs/REPRODUCIBILITY.md
• docs/SOTA_POSITIONING.md
• docs/SOTA_DAILY_WATCH.md
• docs/DAILY_SOTA_UPDATE.md
• docs/WORKFLOW_RUN_RO_CRATE_ALIGNMENT.md
• docs/profile/CAPAS_PHYSICAL_EVIDENCE_PROFILE.md
• docs/FORMAL_BOUND_AXIS.md
• docs/WITNESS_INDEPENDENCE_AXIS.md
• docs/OPTIMIZATION_BRIDGE.md
• docs/EXPERIMENTAL_EVIDENCE_AXIS.md
• docs/UNIVERSAL_INVARIANT_ANCHORING.md
• docs/METAMORPHIC_TESTING_POSITIONING.md
• docs/SCIAGENTGYM_AUDIT.md
• docs/QMB100_AUDIT.md
• docs/VVUQ_QUANTUM_AUDIT.md
• docs/COVERAGE.md
• docs/EXTERNAL_MVP_LAUNCH_PLAN.md

Short defensible claim:

CAPAS is an evidence-typed claim gate over scientific-computation traces: it reads or packages provenance-aligned evidence and decides whether that evidence licenses ACCEPT, REWRITE, REJECT, or HOLD.