Worthify Decision-1

Choose among the actions your application supplies. Give Decision-1 a text state, a question, and 2–16 described options. It returns a selected option ID and scores conditional on that menu. Applications can use it for bounded routing, evidence triage, or selecting a workflow step while retaining their own validation and review rules. The scores are uncalibrated; they are not probabilities that an action is safe or a claim is true.

Decision-1 is a public, single-seed research preview produced by training all 11.9B text parameters of Gemma 4 12B. It scored above frozen Gemma 4 12B on the primary metric in all seven families of our 3,200-row held-out decision evaluation. Three families transform routing or evidence tasks into supplied-choice menus; four are authored or menu fixtures. The earlier task LoRA had higher point estimates on two evidence families. These results do not establish real-world operating performance. The planned second full-weight seed has not run. See the held-out results, recorded decision replays, and task LoRA guide. The detailed adapter evaluation includes the BANKING77 and CTU-13 comparisons; it does not show a general tuned-base advantage.

What the full-weight checkpoint measured

The seed-42 run trained on 35,210 rows, selected on a separate 3,200-row validation set, and scored a sealed 3,200-row test after selection and export verification. Frozen Gemma, an earlier task LoRA, and the full-weight preview used the same BF16 batch-four scorer on one A100. The chart plots each family's primary score; some use macro-F1 and others use accuracy.

The routing macro-F1 denominator includes 151 offered labels, 46 without a gold example in the test sample; 512 routing rows use none_of_above as gold. The authored/menu families are bounded fixtures, not evidence of real-world cyber or operational performance. The aggregate held-out report can be recomputed from the exact text-free held-out prediction rows for frozen Gemma, the earlier task LoRA, and Decision-1.

Watch a decision change

A claim becomes supportable only when an install record arrives. The fictional Harbor gateway case asks whether a fix was installed before an audit. With only a vendor advisory, the model chooses insufficient. A later installation log and inventory record lead it to supported. In the complete six-stage authored replay, it matched four expected routes and missed two. This is a demonstration of the interface, not an accuracy benchmark.

The same menu handles traffic shape. In a synthetic window of small, roughly periodic outbound records, the model chooses beacon_like from descriptions of observable patterns. It sees connection metadata, not packet payloads. Four of six authored windows matched their expected route. The label makes no claim about malicious intent.

A text-state game keeps moving. In one falling-block run, code supplied legal placements and board features as text. The same full-weight model selected placement goals and buttons for 30 pieces, cleared six lines, and applied 25 rotations while gravity advanced. This single seed has no raw-Gemma control and is not a gameplay benchmark. The model did not see pixels.

These are silent, editorial replays of recorded traces. The on-screen inference times are measured; video frame rate is not model throughput. The media receipt binds the videos to their source traces and records the styling assets.

Train a task LoRA

The Worthify source and training recipe shows the expected JSONL rows, a fresh adapter run, reload verification, and evaluation on a separate test file. Compare your adapter with one trained on the pinned original Gemma base under the same task rows and training budget.

git clone https://github.com/Worthify-AI/worthify-decision-1.git
cd worthify-decision-1
python -m venv .venv && .venv/bin/python -m pip install -e '.[train]'

Try one decision

The public model requires one visible CUDA GPU for this local NF4 example. From the checkout above, create decisions.jsonl:

{"id":"route-1","state":{"claim":"Harbor gateway H-2 received release 5.4.2 before the noon audit.","records":[{"source":"gateway installation log","text":"H-2 installed release 5.4.2 successfully.","time":"10:42"},{"source":"asset inventory","text":"Harbor gateway is H-2; installed release 5.4.2.","time":"11:11"}]},"question":"Which evidence route is warranted by the supplied records?","options":[{"id":"supported","description":"The supplied records support the precise claim."},{"id":"contradicted","description":"The supplied records conflict with the precise claim."},{"id":"insufficient","description":"The supplied records do not settle the precise claim."}]}

CUDA_VISIBLE_DEVICES=0 .venv/bin/worthify-lora score \
  --model Worthify/Decision-1 \
  --revision cab55818cfb6f4163f2b2597bed31cceb6ba2030 \
  --quantization nf4 --input decisions.jsonl --output scores.jsonl

This exact sample selected supported in our single-GPU NF4 usage check. The scorer writes the chosen option ID and conditional menu scores. An application must validate its options and decide when to request review. Scores are uncalibrated; they are not probabilities that an action is safe or a claim is true. This NF4 example uses a different numeric path from the BF16 video replay and batch-four held-out table; choices can differ across those paths. It is a usage check, not benchmark reproduction. The training prompt cap was 2,048 tokens. Longer inputs have not been validated.

Detailed task LoRA evaluation

A task adapter can improve a model on its own task without making that model a better starting base than original Gemma. The latest matched BANKING77 study favored original Gemma by 0.32 percentage points for the validation-selected pair, while a CTU-13 scenario favored original Gemma by 0.316 macro-F1. The BANKING77 within-base gains and earlier two-epoch result remain below with their methods and row-level evidence. Neither task establishes performance on other downstream tasks.

BANKING77 before and after task tuning

On the official BANKING77 test, we scored each base without a task adapter and paired those outcomes with its validation-selected seed-42 LoRA. The Decision-1 adapter gained 173 more correct routes; the original-Gemma adapter gained 174. Both gains were measured on the same 3,080 utterances, each converted to a deterministic 16-option menu.

The paired 77-intent bootstrap intervals for LoRA-minus-unadapted accuracy are +3.21 to +8.57 points on Decision-1 and +3.28 to +8.44 points on original Gemma. Excluding 28 rows flagged for lexical overlap leaves gains of +5.67 and +5.64 points, respectively. “Zero-shot” here means no BANKING77 task adapter or examples in the prompt; Decision-1's full-weight training included related intent-routing data. We scored these baselines after viewing the adapter results, so this is a descriptive, post-hoc comparison. The gains do not establish that Decision-1 is a stronger adapter starting point or learns faster. These are not standard 77-way BANKING77 scores.

Method · report · 3,080 text-free paired rows

Matched starting-base tests

The within-base gains above are separate from a comparison of the two starting bases. In the latest four-epoch BANKING77 study, validation selected seed 42 for both bases: original Gemma at update 350 scored 93.54% on the official test and Decision-1 at update 325 scored 93.21%. The tuned-minus-original difference was −0.32 percentage points, with a descriptive 77-intent grouped bootstrap interval of −0.94 to +0.29 points. Seed 43 also favored original Gemma: 93.77% versus 92.66%, a −1.10-point difference. Excluding the 28 rows flagged for lexical overlap with full-weight train, validation or test material left a −0.36-point selected-pair difference on 3,052 rows.

Both bases received fresh rank-16 q/k/v/o adapters at seeds 42 and 43, with seed-matched initial tensors, identical task rows and options, NF4 base loading, BF16 computation, FP32 final-softcap option loss, optimizer and update budget. The extension used 1,540 training rows, 365 validation rows, four epochs and 388 updates. Each run's checkpoint and each base's seed were selected on validation before fresh-process parity checks and sealed-test scoring.

The latest 388-update study's selected adapters are available as separate public packages: Decision-1 base, update 325 and original-Gemma base, update 350. Both are seed-42 packages pinned to immutable commits, with base revisions and numeric policies recorded. The earlier-study packages are listed separately below.

Decision-1 started lower and had greater baseline-adjusted validation-accuracy area and greater area over updates 194–388 in both seed pairs. Original Gemma had greater absolute area over updates 0–388 in both pairs. All runs were already above the protocol's 80% threshold at step zero. These validation trends do not override the held-out result or establish faster convergence. The historical report field faster_learning_gate: true checks only the baseline-adjusted area, not absolute accuracy or convergence speed.

What changed from the earlier result?

The earlier, two-epoch BANKING77 study reported 91.62% on Decision-1 versus 90.42% on original Gemma at a 194-update budget: +1.20 points for the selected pair and +0.88 for seed 43. Its selected-pair interval was +0.32 to +2.14 points. That result motivated the longer experiment; the extension is an exploratory follow-up, not an independent confirmation. Earlier checkpoints did not preserve optimizer or RNG state, so all extended adapters started fresh. The old and new trajectories diverged at the update-97 epoch-end batch; the cause was not isolated. The new plot is not a resumed or spliced old curve.

The earlier BANKING77 report and earlier paired-outcomes chart remain available. The earlier 194-update study's selected artifacts remain archived in the private experiment repositories, distinct from the public 388-update adapter packages. The earlier report records their base revisions and numeric policies.

The CTU-13 comparison remains a counterexample: on one sealed flow scenario, original Gemma scored 0.432 macro-F1 and Decision-1 scored 0.115. Both adapter seeds favored original Gemma. The task metrics differ and are not combined into a single average.

BANKING77 was chosen after that CTU-13 regression and is close to intent-routing data already present in full-weight training, including CLINC150. Every menu offers the answer plus 15 distractors from 77 intents; these are not standard 77-way BANKING77 scores. Lexical exclusion does not establish semantic novelty. The four-epoch method, full report, and 3,080 text-free paired rows preserve both outcomes and their provenance. The evidence guide explains how to rerun the public checks.

Method, provenance, and limits

This is text-only bounded-choice scoring. The model cannot select an option you omit, and the present interface accepts at most 16. Browser or game videos convert their state to text. Native cache paths remain disabled after equivalence failures. Batch grouping changed logits in diagnostics, so the main held-out table is a batch-four result rather than a universal single-request guarantee. The second full-weight seed and five-task adapter series have not been run; neither is a result of this public research preview.

The Gemma-derived weights carry Apache-2.0 terms with applicable attribution. The public release inventory binds the current card, config, weights, evaluation files, videos, adapter commits, and source recipe by SHA-256. Earlier export receipts retain their original private-review scope and hashes. Worthify's source additions to MIT-licensed OpenJev remain MIT. CLINC150, WANLI, MultiNLI, and BANKING77 retain their separate source terms; the attribution inventory records their provenance. Worthify Decision-1 is Worthify's own Gemma post-training. It does not use TypeSafe Jev output as training labels or claim TypeSafe API compatibility.