TriusAI/kapteeni-v1.1c

Kapteeni v1.1c — a multimodal Jev-compatible System One decision model

Send a state — optionally carrying an attached image — plus typed questions; get back calibrated probability distributions your code can branch on. No text generation. Kapteeni implements the TypeSafe System One decision-model interface (POST /v1/systemone, the wire format of Jev).

The flow: state (+ optional state.image, one base64 PNG/JPEG) → every question — and, for choice, every option; for score, every level — serialized into its own pass over the same state (chat-template render; the image enters through the processor's vision path as ~400 tokens) → Qwen3.5-4B → each pass's final-token hidden state → tiny per-primitive readout heads (choice = softmax over the row's option passes scored in-loss during training; score = per-level independent BCE; noul = an absolute probability — P(A) + P(not-A) can exceed 1 by design, mirroring the reference) → per-primitive temperature (the only constants fitted anywhere in the line; held-out validation, never benchmark data) → contract shaping → typed answers with exact-1 probability sums. Deterministic end to end; question ids never reach the model; a question is forwarded on its own, so adding or removing questions cannot change another question's answer.

Benchmark: JevBench public half (231 items, official harness, official scorer)

Run of record (2026-10-02), measured through THIS snapshot's packaged distribution (serve --dist/--hf) at jevbench upstream bb05a33; 231/231 attempted, 0 failures, all schema-valid. Context rows: our shipped text-only variant kapteeni-v1-meticulous.

	kapteeni-v1.1c	(context: v1-meticulous)
composite (the benchmark's axes; our hardware assumptions)	63.72	65.71
Intelligence	68.33	60.3
public accuracy (231 items)	0.766 (177/231)	0.710
easy / standard / hard	1.000 / 0.931 / 0.559	1.000 / 0.889 / 0.469
top-label ECE → Calibration	0.0984 → 80.3	0.0496 → 90.1
latency p50 / p95 (raw)	0.26 s / 10.6 s	0.17 s / 1.17 s
mean input tokens/decision	562	597

• n=231 carries a 95% CI of ±5.9pt — the composite difference vs 65.71 is inside single-run noise; don't read it as a win or a loss.
• The hard-tier jump (+9.0pt), the accuracy gain, and Intelligence +8.0 are within-resolution separable, and the family breakdown shows where: multi_hop 0.278 → 0.667, long_policy 0.316 → 0.421 (the rule skills the port targeted), while temporal_numeric (n=15; 0.20 → 0.13) remains beyond every model this lineage has trained.
• Latency/cost are raw numbers on our only hardware (AMD Strix Halo iGPU, ROCm, contended); plain PyTorch/transformers, no vendor-specific code. The board's official Intelligence folds in sealed items (ours renormalizes over the three public tiers, as the text variants' numbers do); the full section incl. the frozen-v1.5-method note is in the repo's docs/JEVBENCH.md.

When to use which of our models (numbers first; this repo makes no leaderboard/placement claims): traffic with images or Chinese → v1.1c, the only variant that handles either; pure-text well-formed numeric/temporal/multi-step decisions → the measured rule skills favor v1.1c (and the text variant -intuit over -meticulous); messy, adversarial, unknown pure-text traffic whose confidence values are consumed downstream → -meticulous's calibration is the documented safest.

Pre-registered gates (all passed; one run, one reading)

gate	reading	bar
synth3-val (n=590, English image decisions)	0.9661	> 0.8068 (frozen base)
synth3zh-val (n=204, Chinese image decisions; reported)	0.9412	—
MNLI-noul (n=150, English text, never trained on)	0.8800	≥ 0.84
OCNLI-noul (n=150, Chinese text, never trained on)	0.8467	≥ 0.83
synth2zh-val (n=403, Chinese rule skills)	0.9132	≥ 0.90
synth2-EN-val (n=609, English rule skills)	0.9048	≥ 0.90
fitted ECE (noul / choice / score)	0.0241 / 0.0175 / 0.0685	≤ 0.10

Quickstart

# serve from this download (merged model — no peft, no training artifacts)
HF_HUB_OFFLINE=1 python3 -m kapteeni.serve_v11c --dist ./ --port 8002
# or directly from this repo id, no download step needed:
python3 -m kapteeni.serve_v11c --hf TriusAI/kapteeni-v1.1c --port 8002
# API: POST http://localhost:8002/v1/systemone — the existing typesafe
# adapter and every Jev harness reach it unchanged.
# Demo website (15 pre-configured cases with hand-checked gold
# annotations, incl. real photographs): http://localhost:8002/

Example request:

{"state": {"image": "<base64 png>", "note": "a cafe menu."},
 "model": "kapteeni-v1.1c",
 "questions": {"q1": {"type": "choice",
                      "instructions": "Which drink is the most expensive?",
                      "criteria": {"Espresso": null, "Latte": null,
                                   "Mocha": null}},
               "q2": {"type": "noul",
                      "instructions": "Is any drink above $9.00?",
                      "criteria": {"true": "above", "false": "not above"}}}}

Image handling: one image per request. Arriving images larger than a 640px longest edge are downscaled to it (aspect preserved; no crop, no paste, no upscale) — small text and fine detail can become unreadable at the bound, and accuracy may differ from what the full-resolution image would give, because the model and all its gates were trained and validated at 640×640; the response carries an advisory notice naming this whenever the bound applies. Precision-critical callers should send images at ≤ 640px on the long edge, or crop to the region of interest. Question ids never reach the model; answers are content-only.

Training data

Synthetic, gold by construction, all pixels/text ours (Apache-2.0): 20 image document families (English + fully-Chinese renders), binned temporal/multi-hop/long-policy rule families (English + Chinese ports), and an English text replay unioned with the v1.2-era surfaces. Teacher soft labels for noul/replay: k-sample agreement from a cloud LLM judge (fidelity documented). Contamination: the 8-token CJK-aware shingle audit vs JevBench text public items + OCNLI shows zero hits (scripts/contamination_audit.py); ImageJevBench items were never part of an item-level audit (items and keys are not distributed), but the image training data is programmatic template families into which benchmark content cannot enter by construction. Val slices, MNLI and OCNLI were never trained on; no gates or constants were adjusted after any measurement.

Licenses / provenance

Code + generators: Apache-2.0 (matching the base model; the serving package ships inside this snapshot). Trained weights: CC BY-SA 4.0 — the ShareAlike term comes from the MultiNLI + FEVER annotation lineage in the text replay; full provenance and attribution guidance in WEIGHTS-LICENSE.md. Backbone: Qwen/Qwen3.5-4B (Apache-2.0). Not affiliated with TypeSafe; "Jev" is their model and trademark.