Takar Outcome Evaluator

A reward-model adapter for the Telegraph network's EVENT_OUTCOME_RESOLUTION family. It is a rank-8 LoRA plus a one-logit reward head on the shared Foundation Model modernbert-base-lora16-v1 (ModernBERT-base), packaged in Telegraph's lora-v1 safetensors delta format.

It scores how well a miner's answer resolves the event that was asked. It ranks last the answers that have nothing to do with the question, which the network's default judging did not catch.

Package EVENT_OUTCOME_RESOLUTION.safetensors, 6,776,916 bytes
keccak256 (this file, v2) 0xfeac3f314909f8a881f67226552445156d347df6682f4f3743eec60e50eedd3e
Network status rejected (registration 156, v2.1 = same weights, re-hashed). See below
Class hybrid (the answer is the miner's whole response body)
Input "### Task:\n" + question + "\n\n### Answer:\n" + answer, node tokenizer, ≀512 tokens
Output one reward logit; higher is better

Network result: rejected

On 20 Nov 2026 the node's remote check sent 15 public pairs to our executor. The adapter put 7 in the right order (46.7%, floor 70%), so it was rejected. The 99.8% below was measured on pairs we generated; it did not transfer to the network's material. We did not capture or train on the node's pairs. Treat this adapter as a study of five failure modes, not as a working evaluator.

What it catches

Real answers from the Telegraph question feed, scored on the network's own WASM engine against an on-topic answer to the same question:

Question Off-topic answer a ranked miner returned Its score On-topic
Will Heidelberg Engineering ANTERION study complete? Bundesliga goal list βˆ’3.73 βˆ’2.45
Will Phrack 74 be released by 2027? NFL Polymarket listing βˆ’6.48 βˆ’2.18
Will Phrack 74 be released by July 2027? F1 Bahrain race data βˆ’9.46 βˆ’0.93

Training

The pairs are built from real Manifold markets (3,697 resolved, 1,865 open) and 4,000 real off-topic payloads from the network feed. There are five kinds:

  • topic: an answer to a different event
  • consistent: an outcome that contradicts its own evidence
  • premature: a confident answer to an open market
  • settled: dodging a closed market
  • junk: a misrouted payload

Plain YES-versus-NO flips are left out, because a 150M encoder cannot know who won.

Pairs (train / val / test) 4,454 / 542 / 537, split by market
Dataset gate (Telegraph kit) pass. Bag-of-words probe 62.2%, length shortcut 50.2%
Test pairwise accuracy 99.81% (length baseline 50.8%)
Real engine, 40 held-out pairs 40 / 40

The test pairs come from the same generator as the training pairs, so 99.8% measures those five failure modes. It does not predict the network's hidden pairs.

Use

Scoring needs Telegraph's foundation and engine, because validators compare scores bit for bit:

git clone https://github.com/telegraphprotocol/telegraph-wasm-baseline
# foundation_lora_fp32.onnx + foundation_engine.wasm from
# github.com/telegraphprotocol/telegraph-foundation/releases/tag/modernbert-base-lora16-v1
go run ./cmd/adapterexec --package EVENT_OUTCOME_RESOLUTION.safetensors

Source, data, training run and every check: https://github.com/PugarHuda/takar/tree/main/evaluator

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for PugarHuda/takar-outcome-evaluator

Adapter
(48)
this model