Datasets:
File size: 6,050 Bytes
536d746 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | # IMD oracle-worker replay pilot
This pilot tests whether a model can use IMD's oracle tools and submit a correct, reproducible answer. It does not predict the historical panel's outcome.
The scope is deliberately small: 20 real Ethereum requests, with ten ERC-20 Transfer counts and ten Transfer sums. Nineteen requests already specify the recipe. Passing this pilot shows basic tool and output compliance; it does not establish which model is best for running a full IMD node.
## What counts as success
A task succeeds only when the model scans the correct contract, event and pinned window; submits the independently verified value with the matching recipe; passes the official answer checker; and finishes within its budget. Incorrect answers, malformed artifacts, abstentions, timeouts and provider errors remain in the denominator.
The primary metric is **verified task success rate**. Report submission coverage, average and p95 task time, and recorded tokens per successful task alongside it. The numerator for tokens per success includes usage from failed tasks. Precision alone is unsuitable: answering one easy task correctly and skipping nineteen gives 100% correctness among submitted answers but only 5% task success.
## Data and verification
Selection is deterministic and saved in `selection.json`, before model evaluation. Eligible requests use Ethereum, a pinned closing block hash, a standard ERC-20 Transfer event, at most 150 inclusive blocks, no indexed filters and raw integer units. One request per normalized question template was retained in the candidate audit; related wording and token substitutions remain, so these are not twenty independent skill families.
Two public RPC providers independently supplied matching logs and closing block hashes. The expected count or sum was recomputed from those logs, rather than copied from panel answers. Original questions and definitions were reviewed for address, operation, inclusive range, units and filters. The official IMD scanner and answer checker passed every task in preflight.
`tasks.jsonl` contains model inputs and provenance. `labels.jsonl` and `fixtures/` are evaluator data. The runner sends only each task's `input` to the model. The RPC replay exposes raw source data through the scanner, without exposing historical panel outputs or grading labels.
## Execution and limits
- An API tool agent calls the pinned official `scan.mjs` and `check-answer.mjs` helpers. This is **not a native Claude Code/Codex IMD worker run**.
- Tools run in a temporary workspace with a minimal environment. Models cannot issue arbitrary shell commands or choose network endpoints.
- Each task gets up to eight model turns, 4,096 output tokens per turn and 180 seconds. Temperature is zero; concurrency is two. Models run sequentially to avoid overlapping local-gateway load.
- Opus and Sol request medium effort. Qwen 27B requests thinking disabled. These are serving setups, not equal-compute architecture comparisons.
- A structurally valid submitted artifact ends the task; wrong answers do not receive correctness feedback. Structural errors may be corrected within the same budget.
- Time includes model calls, scanner execution and artifact validation. RPC data is replayed locally, so timing does not measure production RPC/network performance. Missing usage is reported explicitly.
- One run per model. Confidence intervals are descriptive and do not account for shared templates; close results should not be treated as established differences.
## Node eligibility and untested work
The pinned [official worker documentation](https://github.com/Identity-md/worker/blob/b55f5e4675664a3910f35b740905ace028d820a4/README.md#start-and-pair) lists approved premium configurations for contract/frontend work, including Opus 5.5 medium+, Sol medium+, Astra xhigh+, and Fable 5.1 high+. Subscription access, runtime version, installed tools and worker admission still need to be checked on the operator's machine. Qwen is a research comparison, not a verified premium-worker configuration.
Contract builds, audits, frontend work, open-web oracle research, other chains, all other oracle recipe kinds, production RPC failures, scheduler admission and earnings are untested. Do not turn this pilot into a general node-suitability ranking.
## Reproduce
Python 3.9+ and Node 22+ are required. The runner reads provider credentials from a local `.env`; the reproducibility archive contains no credentials. Set `BENCHMARK_LOCAL_8317_API_KEY` only if the local gateway requires it, and `QWEN_API_KEY` for the optional ModelScope comparison. Gateway model IDs must be available locally.
```sh
python3 scripts/fetch_node_helpers.py
python3 scripts/validate_node_pilot.py
python3 scripts/run_node_pilot.py --model opus --concurrency 2
python3 scripts/run_node_pilot.py --model sol --concurrency 2
python3 scripts/run_node_pilot.py --model qwen-27b --concurrency 2
```
The helper fetcher refuses an upstream revision whose hashes differ from the pinned manifest. The runner refuses to overwrite an existing run. Use a fresh extracted directory for an independent repeat.
## Grading correction
Before publication, all 60 saved outputs were rescored after correcting an overly narrow event-format check: the official checker allows the `event` prefix to be omitted. One Qwen task changed from failure to success (12/20 to 13/20). Original metrics remain in the published metadata; no prompts or model outputs changed, and no model calls were repeated.
## Next evaluation stage
Before recommending a model for node operation, add fresh, independently graded tasks across the operator's intended skill mix; reconstruct starting repositories for contract and audit work; run native worker-compatible runtimes; and repeat evaluations under fixed budgets. Report success separately by skill and eligibility separately from measured performance. Do not optimize thresholds or prompts on the old public test set.
License: unknown. Collected from the publicly accessible IMD API.
|