Prompt through Hive to a model. Get the response plus a signed, verifiable receipt for the call path.
The receipt captures which provider and model answered, a hash of the prompt, a hash of the response, the
per-phase timing, and an Ed25519 signature that verifies against did:hive:hivemorph.
Six direct providers are live today: Anthropic Claude, OpenAI GPT, xAI Grok and Google Gemini as
first-party custody, Mistral and DeepInfra as direct inference infrastructure, plus Claude,
Gemini, Grok, GPT and DeepSeek routed via OpenRouter; the adapter layer is built to target many more.
This lab is where signed model calls and model receipt verification fit into Hive's wider proof layer.
A model receipt proves the model-call route: the provider and model, the prompt hash, the response hash, the timing, the signature, and the verification result. It does not attest that the model output is correct. "Universal" describes the architecture. Adapters are registered per provider and model, not a claim that every provider is wired today. Live now: Anthropic Claude, OpenAI GPT, xAI Grok and Google Gemini on first-party direct custody, Mistral Leanstral plus a set of registered DeepInfra OpenAI-compatible models on direct infrastructure, and Claude, Gemini, Grok, GPT and DeepSeek routed via OpenRouter. We list exactly which DeepInfra models are wired. This is not a claim of access to DeepInfra's entire catalog. The rest of this page is explicit about what is live versus what is on the roadmap.
The first wired adapter is Mistral Leanstral 1.5 (labs-leanstral-1-5). It is called live on the
backend, and every call is receipted and verified. The worked example walks one full run end to end.
did:hive:hivemorph via /v1/receipt/verify.See the full run, hashes, and signature on the Leanstral receipt example.
These controls call receipts.thehiveryiq.com directly from your browser. Each button does real work:
list the registered adapters, receipt a live Leanstral run, or run a SmartMorph route comparison. If a fetch fails, the
last-known verified values below still stand on their own, and copyable curl commands are provided further
down.
labs-leanstral-1-5did:hive:hivemorphff11617efb0ffc3c51bf292ca701cef6bf0658104cbe5b55d28fb59ef8d6ad4fafb99469edcaAdapters are registered per (provider, model). Live today: direct first-party adapters for
Anthropic Claude, OpenAI GPT, xAI Grok and Google Gemini, direct infrastructure adapters for Mistral Leanstral and a registered set of
DeepInfra OpenAI-compatible models, plus routed adapters via OpenRouter for Claude, Gemini, Grok, GPT and DeepSeek.
The "Run live adapter check" button above reflects the current backend state.
Called on each provider's own first-party API, not a routed hop, so the receipt attests direct first-party provider custody plus the selected model id, prompt/response hashes, timing and signature. It does not attest output correctness.
Gemini 2.5 Flash is the demo-clean hero: first-party call, signed, verified, non-empty response. Gemini 2.5 Pro is receipt-path live and verified too, but on the short demo prompt it can spend its budget on internal reasoning and return no visible text, so the live check honestly marks it receipt-ok / empty-content rather than demo-clean. Both prove first-party custody + model id; neither attests output correctness.
Wired, called live on the provider's own inference endpoint, and receipted on every call.
Registered DeepInfra models on the OpenAI-compatible chat surface. Each is configured on the backend and returns a real signed receipt on call. This is the set we have wired, not a claim of access to DeepInfra's full catalog. Some models produce clean demo text on a generic prompt; others need a model-appropriate prompt, so the live check labels each honestly.
Reached through OpenRouter's OpenAI-compatible surface and receipted on every call. A routed receipt proves the Hive→OpenRouter call path and the selected model id, not first-party provider direct custody. Claude, GPT, Grok and Gemini all also have direct first-party rails above; the chips here are the routed hops, distinct from those first-party rails (this Grok 4.3 chip vs. first-party Grok 4.20; this Gemini 2.5 Flash chip vs. the first-party Google Gemini rail). A routed Gemini receipt proves the OpenRouter path, not Google first-party custody.
These providers expose an OpenAI-compatible chat API, which is the same shape the live adapters use. They can be registered without new receipt code, but none are wired or verified live yet.
Anthropic Claude, OpenAI GPT, xAI Grok and Google Gemini are now live on first-party direct custody (rails above). What remains here is Manus: the account key is in hand, but Manus does not yet expose a usable first-party model-call API surface for a direct adapter, so no adapter is wired and nothing is claimed live. Listing it is a direction, not a capability claim. The receipt layer is provider-independent: the moment Manus ships a callable first-party surface and the adapter is configured, the same signed-receipt path applies with no change to the proof format.
Picks a model, calls it, and captures the call as one signed receipt.
Compares candidate model stops on observed signals, chooses a route by explicit rule, and receipts both the morph decision and each model call.
With direct and routed providers live, Anthropic Claude, OpenAI GPT, xAI Grok and Google Gemini as direct first-party, Mistral and
DeepInfra as direct infrastructure, plus Claude, Gemini, Grok, GPT and DeepSeek routed via OpenRouter, a SmartMorph
route comparison is demonstrable today: the "Run SmartMorph route demo" button above runs live model calls (which can
mix first-party, infrastructure and routed candidates), receipts each, and selects a route by explicit rule. The
morph-decision receipt carries the selected custody path; when the selected stop is reached through OpenRouter it
records a routed_via caveat, proving the routed call path and selected model id, not first-party custody.
It is economic-routing capable / a demo decision. It compares live call
signals and signs the choice; it does not record an economic or physical state transition. When an actual
route change or x402 settlement occurs, each becomes its own receipted event.
Model Receipts pin an output to the model version that produced it. Five further primitives go underneath that: one proves the model's identity across a whole run, three measure and route on what a model retains, and one signs the evaluation that was used to trust it in the first place. Two have a hosted signer live today, MiR and EvAR. The other three are built and filed as patent applications, and are named here because they are part of the proof architecture, not because they are released.
Signs which model actually served each step of a run. Each step commits a model id, a weights digest, a configuration hash and an endpoint, so a chain can be checked against the model that was approved for it and a substitution part-way through a run is visible rather than invisible. It attests identity and lineage. It does not attest output correctness.
Commits explicit facts as they enter a context, marks which are meant to be durable, then checks
later which of those durable facts survived. It signs the retained set and the lost set, measures retained facts
against context tokens consumed, binds a decay accumulator, and flags spans that persisted without provenance.
It carries explicit_fact_retention_only, so the receipt cannot be read as a claim about semantic
memory quality.
A per-model fingerprint built only from verified MiR-M receipts that match the model identity. It recomputes fact-loss rate, loss against context depth in buckets, retention efficiency and contamination rate, and binds the sample root of the receipts it admitted, so the profile is derived from signed measurements rather than from a vendor's own benchmark chart.
Estimates a task's memory demand, verifies the MPP profiles of candidate models, projects each
candidate's fact loss at that task's context depth, and selects under a stateless, low-loss-tolerance or
cost-sensitive regime. It signs a recommendation_from_attested_profiles receipt. It does not claim
the routing decision was provably optimal.
Signs which model, rubric, dataset, evaluation method, scored items, aggregate score, evaluator and attestation tier produced an evaluation result, and records whether a check for contamination between the dataset and the training data was performed and came back clean. It wraps an existing evaluation framework rather than replacing it.
The schema is published at /.well-known/schemas/evar-v1.json and the offline verifier
is canon/verify_evar.py, so the format can be checked today with the open kit. The hosted EvAR
signing and verification endpoints are live at the Hive typed receipt API,
https://thehiveryiq.com/v1. Production acceptance passed against this deployment using
synthetic acceptance data, not a customer workload, and minting a real receipt requires an authorization token
that Hive controls. No customer deployment exists and nothing here implies a customer endorsement. EvAR attests
to the integrity of the evaluation record. It does not attest that the evaluated system is correct, that the
method is adequate, or that any certificate or accreditation is valid. A contamination-check status of
passed records that the declared check ran and returned clean, and is not proof that the dataset is
free of contamination.
Live status is checkable. /sigr/mir and /sigr/mir/verify
answer on the typed signer today, and EvAR's /mint/evar and /verify/evar answer on the Hive
typed receipt API. The three remaining primitives marked patent pending return no route, deliberately, until they are
released.
Read-only endpoints. Verification is free; the receipt round-trips the public
/v1/receipt/verify. The demo-matrix and SmartMorph routes are additive. If the backend has not deployed
them yet, the page falls back to /adapters + two live /run calls, so these still work.
curl https://receipts.thehiveryiq.com/v1/model-receipts/leanstral-demo
curl https://receipts.thehiveryiq.com/v1/model-receipts/adapters
curl -X POST https://receipts.thehiveryiq.com/v1/model-receipts/run \
-H 'Content-Type: application/json' \
-d '{"provider":"mistral","model":"leanstral-1-5","prompt":"State the theorem add_comm for Nat."}'
curl -X POST https://receipts.thehiveryiq.com/v1/model-receipts/run \
-H 'Content-Type: application/json' \
-d '{"provider":"deepinfra","model":"deepseek-ai/DeepSeek-V3.1-Terminus","prompt":"Reply with one short line: pick a route label for a low-latency code task."}'
curl https://receipts.thehiveryiq.com/v1/model-receipts/smartmorph-route-demo
Model facts: Mistral Leanstral 1.5: mistral.ai announcement and the model card on Hugging Face.
"Universal" refers to the adapter architecture, not to a claim that all providers are wired. As of 2026-07-04 the live
adapters are Anthropic Claude (Haiku 4.5, Sonnet 4.5), OpenAI GPT (4o mini, 4.1 mini, 4o), xAI Grok (4.20 Reasoning)
and Google Gemini (2.5 Flash, 2.5 Pro)
on first-party direct
custody, Mistral Leanstral 1.5 and a registered set of DeepInfra OpenAI-compatible models (moonshotai/Kimi-K2.7-Code,
zai-org/GLM-5.2, Qwen/Qwen3-Coder-480B-A35B-Instruct-Turbo, deepseek-ai/DeepSeek-V3.1-Terminus, stepfun-ai/Step-3.7-Flash)
on direct infrastructure, plus Claude, Gemini, Grok, GPT and DeepSeek routed via OpenRouter.
This is the registered set available to the account, not a claim of access to DeepInfra's entire catalog. Some models
return clean demo text on a generic prompt; where a model returns tokens but empty content, the receipt path is still
functioning and is labeled as such rather than presented as demo-clean (Gemini 2.5 Pro can do this on the short demo
prompt). Other OpenAI-compatible providers are listed as
compatible targets, not live integrations, and Manus remains a roadmap item with no
live first-party adapter (no usable first-party model-call API surface yet). The SmartMorph route demo compares live model-call signals and signs the decision; it is economic-routing
capable but does not itself record an economic or physical state transition. Timing values under "Last known verified
run" were recorded from a live backend run on 2026-07-04; the live controls call receipts.thehiveryiq.com
directly and reflect current state, so numbers vary between runs. A model receipt attests the model-call route only. It
makes no claim about the correctness of the model's output.