Evidence Loop Visibility Engine

Evidence before action.

A deterministic, offline reference implementation for turning bounded visibility evidence into one reviewable proposal per site.

The 0.3.x release line was controller-accepted on 2026-07-20 after independent read-only opencode-go/grok-4.5 (high) evaluation returned PASS. Immutable candidate identities are recorded outside the candidate tree so editing a truth document cannot invalidate a fingerprint embedded in that same content. The v0.3.2 GitHub release is the package publication trigger; PyPI availability follows its protected Trusted Publishing workflow.

The 60-second explanation

The engine accepts strict, timestamped evidence packets, chooses one eligible opportunity per site, and produces a proposal without changing the site. It then verifies the proposal's lineage, routing, and approval boundary before it records an atomic receipt.

Every run stops in an explicit state: approval-required, clean-no-op, or blocked. Selection and receipts are deterministic, so the same exact input bytes produce the same decision and digest. The installed runtime uses only the Python standard library and makes no network, browser, provider, or credential calls.

This is a control loop for reviewable decisions—not a black-box promise of rankings, traffic, answer inclusion, citations, or causality.

Who it is for

• Engineers building evidence-first SEO, AEO, GEO, or LLMO tooling.
• Technical SEO and editorial teams that need a reproducible proposal trail.
• Evaluators testing lineage, failure containment, and human approval gates.
• Operators who want a useful offline core before connecting any live system.

Public core and private operation

This repository is a complete, useful offline core. It is not an intentionally crippled demo, and its Apache-2.0 implementation has no artificial lock-in. Real adapters, calibrated evidence and history, evaluation and operator judgment, team workflows, and managed operation can add value around the core without changing what the public package honestly does.

See the open-core boundary for the extension points, clean-room rule, and the line between public behavior and separately operated systems.

Quickstart

Python 3.10+ is required. Runtime dependencies are the Python standard library only.

Released PyPI package

python3 -m venv .venv
.venv/bin/python -m pip install evidence-loop-visibility-engine
.venv/bin/evidence-loop demo --output work/demo
.venv/bin/evidence-loop benchmark

This invokes the currently published PyPI release. The source workflow below exercises the same commands from a checkout.

Run from source

From the root of the source working tree:

python3 -m venv .venv
.venv/bin/python -m pip install -e .
.venv/bin/evidence-loop validate --input examples/normal.json
.venv/bin/evidence-loop normalize --input examples/connector-envelope.json --as-of 2026-12-31T00:00:00Z --output work/normalized
.venv/bin/evidence-loop validate --input work/normalized/normalized.json
.venv/bin/evidence-loop run --input examples/normal.json --output work/normal
.venv/bin/evidence-loop demo --output work/demo
.venv/bin/evidence-loop benchmark

The package artifact exposes the same evidence-loop command. The committed examples/ files are readable fixtures; packaged resources make demo and benchmark work after wheel or source-distribution installation too.

normalize accepts one or more sanitized Connector Exchange Envelope v1 files and deterministically writes a schema-version-2 document. Schema version 1 documents remain supported. The installed core performs no network or provider calls, and normalization alone creates no opportunities. Its --as-of UTC collection bound is required; it never falls back to the wall clock.

For release artifact validation, install the optional build tools and run the same gate:

.venv/bin/pip install -e '.[release]'
.venv/bin/python scripts/artifact_smoke.py

Optional Vercel connector

Vercel users may separately install the Evidence Loop Vercel Web Analytics Connector. Its v0.1.0 GitHub release emits sanitized envelope JSON for this core. It is the optional second package; it is not bundled into the core, and it remains separately unpublished on PyPI.

Companion GitHub Action

Phase 3 adds a companion Action that accepts only sanitized Connector Exchange Envelope v1 files, requires an explicit UTC as-of bound, normalizes with the installed public core, and runs one bounded cycle. It never invokes or bundles a provider connector. The fixed artifact contains only normalized.json, run.json, and the non-blocked last-success.json.

External Action dependencies are pinned to full commit SHAs, the core executes directly from the immutable Action checkout without package installation, and the documented caller permission is contents: read. Push and pull-request workflows completed the hosted Action job successfully at candidate commit bfa544e, with clean-no-op and external_calls=0. This does not prove publication, provider compatibility, or a production result. See the companion Action guide.

One bounded cycle

Observe -> Choose -> Propose -> Verify -> Record

Choose is explicit: fresh, non-missing evidence is eligible; lower numeric priority wins, then the stable opportunity ID breaks ties. Propose is the runtime's Act step and never mutates a site. Verify is a distinct fail-closed boundary before Record. See LOOPS.md and docs/loop-engineering.md.

Input and output

An input document has reserved example URLs, evidence, and opportunities:

{"schema_version":"1","sites":[{"site_id":"site-a","site":"https://a.example","evidence":[{"evidence_id":"ev-1","source_kind":"manual-observation","observed_at":"2026-01-15T10:00:00Z","completeness":"complete","freshness":"fresh","uncertainty":"low"}],"opportunities":[{"opportunity_id":"opp-1","domain":"technical-seo","title":"Review indexability signals","priority":1,"evidence_ids":["ev-1"],"approval_gate":"human-review"}]}]}

run.json preserves site, evidence IDs, source kind, timestamp, completeness, freshness, uncertainty, routed capability, and approval gate:

{"terminal_state":"approval-required","input_sha256":"<SHA-256 digest of exact input bytes>","sites":[{"site_id":"site-a","status":"approval-required","selected_opportunity_id":"opp-1","proposal":{"approval_required":true,"mutation_allowed":false,"evidence_ids":["ev-1"]}}],"safety":{"offline":true,"site_mutation":false,"provider_access":false}}

The CLI prints only safe counts, IDs, terminal state, and zero external calls or cost. last-success.json is atomically replaced only for a non-blocked run.

Capability maturity

These are deterministic proposal templates, not SEO analysis or outcome predictions:

Allowlisted module	Maturity	Proposal boundary
measurement-integrity	Implemented deterministic	Preserve source, window, completeness, freshness, uncertainty
technical-seo	Implemented deterministic	Propose an indexability/crawlability review
search-intent-content	Implemented deterministic	Propose an intent clarification review
aeo-answerability	Implemented deterministic	Propose a question/answer structure review
geo-citation-research	Synthetic observation	Observe a citation surface; never fabricate a GEO score
llmo-sampling	Synthetic observation	Specify fixed-prompt sampling, variance, terms, and cost gates
brand-governance	Approval-gated	Propose claim and voice consistency review
marketing-conversion	Approval-gated	Propose a measured conversion hypothesis review

Unknown capability domains block their entire site lane. Other sites remain isolated.

Commands and exit codes

• validate --input FILE: strict validation; exit 0 when valid, 2 on a global input/path error.
• normalize --input FILE [--input FILE ...] --as-of UTC --output DIR: required explicit UTC-bound offline Connector Exchange Envelope v1 validation and deterministic schema-version-2 output; exit 0 when valid, 2 on an input/output error.
• run --input FILE --output DIR: one bounded cycle; exit 0 for approval-required or clean-no-op, 3 for global blocked, and 2 for an input/output error.
• demo --output DIR: committed synthetic normal, clean-no-op, and contained failure examples; exit 0 when all complete.
• benchmark: deterministic public conformance cases and pass rate; exit 0 when all cases pass.

Security and non-goals

The installed engine/CLI opens no network connection, invokes no browser or provider, reads no credential environment, spawns no process, and mutates no site. It rejects duplicate JSON keys, NaN/Infinity, oversized or deeply nested input, unsafe IDs/timestamps/strings, lexical traversal, and symlink ancestors or children. Output receipts are atomic. The release scanner is separate defense-in-depth tooling and may invoke local Git to enumerate tracked files; its heuristic is not proof of safety.

This is not an autonomous SEO, growth, ranking, traffic, answer, citation, or conversion system. It does not claim special markup, llms.txt, or any file guarantees Google or another system's visibility. It does not publish, schedule, create backlinks, submit pages, or access private repositories. Fixtures are synthetic and use reserved example domains.

Documentation and development

• Architecture
• Loop engineering
• Open-core boundary
• Connector contract
• Optional Vercel connector
• Companion GitHub Action
• Visibility domains
• Security model
• Public claims
• Quickstart
• Release contract
• Roadmap
• CONTRIBUTING.md

Run make check for tests, compilation, and the release scanner. Artifact build/install smoke checks are in scripts/artifact_smoke.py.