Changes agents and operators can verify.

8-agent verified savings record: shared evidence, independent reasoning

Eight department agents preserved eight independent role analyses while paid web-search actions fell from 13 to 4 and measured provider cost fell from $0.16499684 to $0.05560972 in the completed paired first-party run.

Boundary: First-party operating evidence with one explicitly synthetic external-client rehearsal tenant; not customer ROI or net customer savings. The savings record is measurement evidence, not an invoice or payment receipt; SeenRelay billing is OFF.

Open canonical evidence → · Machine JSON

Safe-to-try trust posture and machine-first agent adoption

SeenRelay now publishes a machine-readable adoption contract for autonomous agents plus explicit safe-to-try controls: local scan, shadow-first measurement, fail-open fallback, shared CHECK off by default in Zero-State, zero default completed-result TTL, no implicit mutation suppression, and an easy remove path.

Boundary: Verifiable technical controls; not a third-party security certification or vulnerability-free claim.

Open canonical evidence → · Machine JSON

Public client 0.2.25

SeenRelay 0.2.25 is published through trusted publishing on npm and PyPI. JavaScript / TypeScript adds `seenrelay guide`, which combines the local static prescreen and non-mutating adoption plan into one machine-readable next-action surface for agents/operators. It uploads no source, contacts no SeenRelay service, modifies no project, enables no reuse and cannot return a USE verdict. Python remains behaviorally unchanged.

Open canonical evidence →

Public client 0.2.24

SeenRelay 0.2.24 is published through trusted publishing on npm and PyPI. JavaScript / TypeScript static prescreen recognizes MCP 2026-07-28 cacheable-result ttlMs/cacheScope freshness only alongside compatible cacheable MCP operations; generic callTool code with the same words remains a shadow-measurement candidate. Python remains behaviorally unchanged.

Open canonical evidence →

Public client 0.2.23

SeenRelay 0.2.23 is published through trusted publishing on npm and PyPI. JavaScript / TypeScript adds a local Langfuse TOOL-observation candidate census that hashes exact tool inputs locally, measures exact recurrence and in-flight overlap, and reports Langfuse-recorded TOOL cost when present. The output remains NEEDS_POLICY_REVIEW and cannot authorize reuse or claim actual avoided executions/net savings. Python remains behaviorally unchanged.

Open canonical evidence →

Public client 0.2.22

SeenRelay 0.2.22 is published through trusted publishing on npm and PyPI. JavaScript / TypeScript packages the local OpenTelemetry/OTLP trace census for agent-agnostic overlap discovery below individual agent frameworks; admission requires an opaque exact coordinate and explicit SeenRelay policy, and the adapter does not copy generic raw prompts, URLs or tool payloads into the census. Python remains behaviorally unchanged.

Open canonical evidence →

Public client 0.2.20

SeenRelay 0.2.20 is published through trusted publishing on npm and PyPI. JavaScript / TypeScript adds a generic `wrapFleetShadowCall()` helper and fail-open coordinate canonicalization for fleet shadow measurement and active coordination: non-JSON or cyclic coordinates never suppress the authoritative call and remain unclassified instead. Python remains behaviorally unchanged.

Open canonical evidence →

Public client 0.2.19

SeenRelay 0.2.19 is published through trusted publishing on npm and PyPI. JavaScript / TypeScript adds distributed shadow measurement for exact eligible in-flight overlap across workers while every authoritative operation still runs; no result is shared, and observed overlap cost is kept distinct from actual avoided-execution receipts. Python remains behaviorally unchanged.

Open canonical evidence →

Public client 0.2.18

SeenRelay 0.2.18 is published through trusted publishing on npm and PyPI. JavaScript / TypeScript adds conservative local savings receipts and an aggregation ledger to fleet-level exact in-flight coordination: only actual follower reuse counts as avoided execution, dollar values require explicit cost provenance, and receipt failures cannot change the application result. Python remains behaviorally unchanged.

Open canonical evidence →

Public client 0.2.16

SeenRelay 0.2.16 is published through trusted publishing on npm and PyPI. The local static prescreen now treats Firecrawl provider cache/freshness semantics as a stronger native control that must be measured before shared SeenRelay. No CHECK/OBSERVE protocol, billing, reuse-policy or hosted-runtime semantics changed.

Open canonical evidence →

Public client 0.2.15

SeenRelay 0.2.15 is published through trusted publishing on npm and PyPI. The npm package adds `seenrelay scan`, a local-only static prescreen for recurring metered/read-only validation candidates and stronger native/local controls. Static analysis is deliberately pre-evidentiary: it cannot emit USE, authorize reuse, upload source code or modify the scanned project. JavaScript/TypeScript and Python runtime semantics are otherwise unchanged.

Open canonical evidence →

Public client 0.2.14

SeenRelay 0.2.14 is published through trusted publishing on npm and PyPI. JavaScript/TypeScript and Python expose matching deterministic single-workload audit verdict helpers: USE only for complete natural evidence that safely beats the best measured non-shared path on both cost and latency; DO NOT USE for unsafe comparisons or complete non-winning evidence; otherwise INSUFFICIENT EVIDENCE. The evaluator never enables reuse.

Open canonical evidence →

Public client 0.2.13

SeenRelay 0.2.13 is published through trusted publishing and independently verified from npm and PyPI. The public npm package has a valid declared export map, includes its package-local prepack validator, imports every declared export from a clean install, and successfully repacks from the installed registry artifact; runtime semantics are otherwise unchanged.

Open canonical evidence →

Public client 0.2.11

SeenRelay 0.2.11 is published and independently clean-install verified on npm and PyPI. Python adds bounded sanitized natural-workload evidence parity and hostile economics evaluation; authoritative validation always runs during measurement, ambiguous per-call timing attribution fails closed, and the evaluator never enables reuse.

Open canonical evidence →

Public client 0.2.10

SeenRelay 0.2.10 is published and independently clean-install verified on npm and PyPI. Python adds provider-independent Zero-State with caller-owned private L1, source-native conditional validation and authoritative fallback; the base package remains dependency-free and AES-256-GCM support is optional. JavaScript/TypeScript and Python both fail closed on negative-age private entries caused by worker clock skew.

Open canonical evidence →

Public client 0.2.9

SeenRelay 0.2.9 is published and independently clean-install verified on npm and PyPI. JavaScript/TypeScript and Python add multi-signal shared-evidence assurance helpers and deterministic Fact Coordinate Kit v1. Shared evidence remains explicit caller policy and does not establish truth or independent real-world actors.

Open canonical evidence →

Public client 0.2.7

SeenRelay 0.2.7 is published and independently clean-install verified on npm and PyPI. JavaScript/TypeScript adds Ambient framework routing for LangChain plus a local integration catalog; Python adds Ambient integrations for LangChain and PydanticAI plus a local integration catalog. Ambient measurement remains local-only by default and automatic reuse is not authorized.

Open canonical evidence →

Public client 0.2.6

SeenRelay 0.2.6 is published and clean-install verified on npm and PyPI. JavaScript/TypeScript adds Ambient MCP integration that starts local-shadow by default and can opt exact reviewed tools into the existing local-first guard; Python adds local-only Ambient MCP measurement and an OpenAI Agents adapter while active Ambient reuse remains disabled.

Open canonical evidence →

Benchmark fit tightened

Historical example.com Firecrawl smoke measurements remain unchanged as provider-path mechanics evidence. Their workload fit is now poor because the tested facts were directly source-resolvable; better local, source-native or provider-native controls remain ahead of shared CHECK.

Open canonical evidence →

Public client 0.2.5

SeenRelay 0.2.5 is published on npm and PyPI. JavaScript/TypeScript adds a measurement-only direct Firecrawl SDK shadow adapter supporting current scrape() and legacy scrapeUrl(); the provider call remains authoritative, original arguments and raw result are preserved, provider-cache hits are not relabeled as independent observations, and automatic reuse remains disabled. Python remains shadow-first.

Open canonical evidence →

Public client 0.2.4

SeenRelay 0.2.4 is published on npm and PyPI. JavaScript/TypeScript adds a measurement-only Firecrawl shadow pilot and a packaged hostile benchmark evaluator; Firecrawl remains authoritative during measurement, provider-cache hits are not relabeled as independent observations, and the evaluator never enables reuse. Python remains shadow-first.

Open canonical evidence →

Public client 0.2.3

SeenRelay 0.2.3 is published on npm and PyPI. JavaScript/TypeScript Shadow Proof can export bounded, sanitized natural-workload records directly into the hostile benchmark format while authoritative validation continues; unavailable CHECKs remain in the sample and unsafe or uncomparable hypothetical reuse cannot be reported as safe. Python remains shadow-first.

Open canonical evidence →

Public client 0.2.2

SeenRelay 0.2.2 is published on npm and PyPI. JavaScript/TypeScript Shadow Proof adds local agreement counters for hypothetical SAME_OBSERVED reuse while authoritative validation still runs; compared raw values are not retained in snapshots or reports. Python remains shadow-first.

Open canonical evidence →

Public client 0.2.1

SeenRelay 0.2.1 is published on npm and PyPI. It is behavior-equivalent to 0.2.0 and refreshes package discovery metadata and documentation; JavaScript/TypeScript remains provider-independent and local-first, while Python remains shadow-first.

Open canonical evidence →

Public client 0.2.0

SeenRelay 0.2.0 is published on npm and PyPI. JavaScript/TypeScript adds provider-independent Zero-State local-first optimization; Python remains shadow-first in this release.

Open canonical evidence →

Local-first JavaScript/TypeScript path

Eligible read-only validation can use in-flight L0 reuse, explicit-TTL local reuse, optional encrypted caller-owned L1 storage and source-native conditional validation before any shared CHECK is considered.

Open canonical evidence →

Shared evidence remains optional

Zero-State does not put a shared CHECK on the hot path by default. Fresh independent validation can contribute OBSERVE evidence; intermediary provider-cache hits are not re-labeled as a new independent observation.

Open canonical evidence →

Measured Firecrawl structured-extraction savings

In a first-party n=3 smoke benchmark, bounded reuse avoided 3/3 Firecrawl JSON extraction calls, 15 provider credits, and reduced median latency from 1.266 s fresh / 1.040 s provider-cached to 0.618 s.

Open canonical evidence →

Measured browser-interaction savings

In a first-party n=3 Firecrawl browser-interaction smoke benchmark, bounded reuse avoided 3/3 equivalent provider calls, 9 reported provider credits, and reduced median validation latency from 4.385 s to 0.661 s.

Open canonical evidence →