[AnthonyRonning]

Summary

POST /v1/systemone now also serves Tinfoil's decision models: decision-small, decision-medium and decision-large (vLLM-SR's d3 family in Tinfoil's attested enclaves). A request picks one with model. Request and response fields are unchanged, and glm-5-3-flash stays the default with its behavior unchanged.

• Decision models answer every question of a request with one upstream POST /v1/decisions, sent through the in-process attested Tinfoil client. web/system_one/decisions.rs translates the System One shape both ways, following d3's own reading of a System One question:
  • a noul becomes a predicate; a noul with criteria becomes a boolean choice with the criteria as option descriptions, false first;
  • choice and score keep their labels, descriptions and level order;
  • non-string values are sent as JSON with Python separators, the way d3 renders them.
• Answers keep the System One shapes. Probabilities come back in option order with TypeSafe's confidence. temperature is 1 unless the request sets one; an explicit value is applied as softmax(ln p / T). option_mass is 1, because a classification head only scores the offered options (vLLM reports the same label_mass).
• GLM-5.3-Flash is untouched: same prompt, calibration, limits, errors and billing. The shared upstream plumbing (permit, health probe, send, usage, error classification) moved from SystemOneCall::send_read into Upstream::send, which both engines use, without changing what it does.

Per-model contract

	glm-5-3-flash (default)	decision-*
images	up to 4	422 system_one_images_unsupported (new)
Options per choice	2–255	2–26 (422 system_one_bad_option_count, message names the limit)
Score levels	2–10, any text	2–10, distinct non-blank text, as d3 requires
State + one question	32k + 16k tokens	also within the model's 32,768-token context, else 413 system_one_context_exceeded (new)
Default temperature	calibrated for the option count	1
option_mass	measured	1
usage.requests / output_tokens	one per question / one per question	1 / 0
Upstream	Continuum, masked logprob read	Tinfoil /v1/decisions

system_one_unsupported_model now lists the four model ids. Unknown models still return 422.

Integrations

• Models. model_config: a DecisionModel table, Pro-only. These models never resolve as completion models and appear in neither chat catalog, so chat, Responses and the Research model picker do not see them.
• Routes and health. provider_registry: one Tinfoil route per decision model, kept apart from completion routing. inference/health: those routes are seeded, so probes, circuit breaking and capacity gates cover them.
• Billing.
  • One usage event per call: provider tinfoil, model = the public id, input tokens as Tinfoil counts them (the state is counted once per question).
  • Needs MaplePrivacyLabs/maple-billing-server#135 deployed first. Without its price rows, the billing server prices an unknown Tinfoil model at its Kimi K2.5 fallback rate.
• SDKs (Rust and TypeScript). Doc comments and READMEs only. The types are unchanged, since every answer still carries temperature and option_mass.
• Proxy. README only. /v1/systemone was already forwarded, and model is opaque to the proxy.
• Maple Agent. Registers the three decision models as text-only Pi classifier models next to glm-5-3-flash.

Validation

Unit and component gates

• Backend (services/opensecret): cargo fmt --all -- --check, cargo clippy --locked --all-targets --all-features -- -D warnings, and cargo test --locked --all-features: 893 passed, 0 failed, 34 ignored (the ignored set is the usual disposable-DB and live-provider tests).
• New backend tests:
  • the decision-model table and its Pro access, kept out of both catalogs and out of completion resolution;
  • one Tinfoil route per decision model;
  • health seeding of those routes;
  • in system_one::decisions: request lowering, answer reading (shape, refusal, value and order checks), usage, the temperature override, per-model validation, and a loopback test of the wire request through the Tinfoil transport.
• Maple Agent: provider tests 30 passed (including decision-model registration); pi-ai System One tests 5 passed.
• Hook: the repository pre-commit gates passed on both commits: TypeScript SDK (format, build, tests), Rust SDK (fmt, clippy, 123 tests, docs), OpenSecret (fmt, clippy, 893 tests), and Maple Agent (fmt, git-metadata test, workspace clippy, workspace tests: 2,026 passed).

Local stack, end to end

• Setup:
  • maple-dev-env workspace, with the backend from this branch, local billing and Continuum proxy, and the real Continuum and Tinfoil upstreams;
  • requests through the released maple-proxy 0.4.2, unchanged, with Free and Pro fixture API keys;
  • 15 cases paced at about 8 per minute.

Case	Result
GLM, model omitted and explicit	200; calibrated temperatures (2.48 / 2.41), measured option_mass, requests: 3, output_tokens: 3
decision-small / -medium / -large, same 3 questions	200 in 0.51 / 0.37 / 0.69 s; temperature 1, option_mass 1, requests: 1, output_tokens: 0
noul with criteria, structured instructions and descriptions	200; the criteria move the answer (0.84 plain against 0.04 described)
64 questions on decision-medium	200 in 0.57 s, one upstream request
temperature: 2 on decision-medium	200; applied and echoed
images on a decision model	422 system_one_images_unsupported
27 options on a decision model	422 system_one_bad_option_count (limit 26)
~32k-token state on a decision model	413 system_one_context_exceeded
256 options on GLM	422 system_one_bad_option_count (limit 255, unchanged)
unknown model	422 system_one_unsupported_model
Free plan with API credits, decision model and GLM	403 model_not_available_on_plan

• Usage events: the backend log shows one event per decision call: provider tinfoil, model decision-*, Tinfoil's input tokens (329 for 3 questions, 6,774 for 64), 0 output. GLM still emits one per question.

Before this change

• Live probes of Tinfoil's /v1/decisions through the attested router established its limits: text only, 26 options, 10 levels, 64 questions, 32,768-token prompt per question, strict schema.

• Paced eval on SST-2 and AG News (100 each) through this translation, accuracy:

  Model	SST-2	AG News
  glm-5-3-flash (current System One)	0.95	0.88
  decision-small	0.93	0.86
  decision-medium	0.92	0.88
  decision-large	0.96	0.91

  decision-large also has the best log loss on both.

Not covered here: deployed Dev or Prod behavior, and billing's pricing of these events in a deployed environment (local billing does not consume usage events; maple-billing-server#135 has the pricing tests).

Rollout

1. Merge and deploy maple-billing-server#135.
2. Merge this PR and deploy the backend.
3. Merge MaplePrivacyLabs/maple-marketing#167 (docs, draft, stacked on feat: enable Gemma 3 27B model with starter plan access #110).

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}