MEPHISTO

PROFESSIONAL KNOWLEDGE WORK · RL ENVIRONMENTS

Five professional knowledge work RL environments. Measured by a judge the agent never sees.

Mephisto Samples publishes five reinforcement learning environments from the Professional Knowledge Work family. Each is a containerized workspace with a twelve-hour interaction window, a continuous reward in [0, 1], and a separate judge container holding evaluation assets the agent never sees.

Each task measures whether an agent improves across a long episode under repeated graded feedback instead of whether it produces one correct endpoint answer. The primary signal is a continuous scalar composed from eight weighted scoring lanes and one bonus lane, totalling a 110-point scale.

  • Two-container work and judge split with full architectural isolation
  • Multi-round delivery feedback approximating a client review cycle
  • Full recorded submission history making in-episode learning curves legible
  • Continuous scalar primary signal (not binary pass/fail)
  • Depth inside named professional finance verticals with public-domain US Government data

Five numbers that define the scope and evaluation protocol of the Mephisto Samples release.

Task Bundles

5

finance RL environments

Agent Window

12h

43 200 s per task

Verifier Timeout

3 600s

separate container

Reward Scale

0–110

8 lanes + 1 bonus

Submission Cap

300

120 s cooldown

What this release establishes is structural. Five design contributions distinguish the Mephisto Samples protocol.

Contribution 01

Two-Container Work & Judge Split

  • Work container holds task materials and local validation tools
  • Judge container holds hidden evaluation assets the agent never sees
  • No reviewed agent benchmark documents this full architectural separation

Contribution 02

Multi-Round Delivery Feedback

  • Approximates a client review cycle with up to 300 submissions
  • 120-second cooldown between submissions prevents brute-force flooding
  • Evaluator-only snapshots every 1 800 s capture progress the agent cannot optimize against

Contribution 03

Full Recorded Submission History

  • Agent-initiated and evaluator-only submissions retained in full
  • In-episode learning curve is recoverable after the run
  • Trajectory logging targets learning measurement, not just efficiency analysis

Contribution 04

Continuous Scalar Primary Signal

  • Per-task continuous score in [0, 1] composed from 8 weighted lanes + 1 bonus
  • Partial credit exists: the continuous score is the primary signal, not binary
  • Each lane is a linear ramp between a zero anchor and a full-marks anchor

Contribution 05

Depth in Professional Finance Verticals

  • Five task bundles spanning SEC XBRL fundamentals, Treasury liquidity, Fed policy regimes, leverage trajectories, and FDIC bank capital
  • All data sourced from US Government public domain (SEC EDGAR, FRED, FDIC BankFind, Treasury, BLS)
  • Anti-fabrication gate recomputes every self-reported metric from raw outputs; deviation zeroes the integrity lane plus guarded primaries

The protocol

Two loops and a snapshot channel turn repeated graded feedback into a legible in-episode learning curve.

Phase 01

Inner Loop (local)

  • A writable workspace (/home/workspace) with Python, local validation, and task materials
  • No hidden evaluation asset is mounted; the information boundary is architectural
  • Unlimited fast iteration, entirely agent-driven, across the full 12 h window

Phase 02

Outer Loop (judge-mediated)

  • Submitted artifacts graded by a separate judge container (3 600 s timeout)
  • Hidden test data, expert labels, and rubric graders; both containers run no-network
  • Up to 300 submissions per episode, 120 s cooldown between each

Phase 03

Evaluator-Only Snapshots

  • Automatic grading every 1 800 s through the same hidden judge
  • Results retained for analysis but never shown to the agent
  • Measures progress the agent cannot optimize against

Every bundle composes 100 base points across eight lanes plus a 10-point adaptation bonus.

Each lane is a linear ramp between a zero anchor and a full-marks anchor, so partial competence earns partial credit and the composition stays interpretable. The verifier writes a single scalar in [0, 1] normalized from the 110-point raw score.

The lane allocation below is for the FDIC Bank Capital Projection Book bundle. Other bundles follow the same 8+1 structure with domain-specific anchors.

Fig. 1: Scoring lane point allocation across the FDIC Bank Capital Projection Book.
Fig. 2a: Verifier timeout by task bundle (seconds). All bundles use 3 600 s.
Fig. 2b: Submit file count by task bundle. All bundles submit 3 artifacts.
8 WEIGHTED LANES 100 base pts
1 ADAPTATION BONUS 10 bonus pts
L7 ANTI-FABRICATION GATE integrity enforced
110 maximum raw score per task (100 base + 10 bonus)

Total Tasks

5

Family

1

Base Image

python

Verifier Timeout

3 600s

Loading...
Task Family Base Image Submit Files Eval Timeout Repo

No signed pilot evidence exists for this family at the time of release. The project disposition is HOLD:PILOT_REQUIRED. This release publishes no pass rate, no score distribution, and no model comparison.

Professional Knowledge Work · 5 tasks

Base image
python
Verifier timeout
3 600 s
Submit files
3 per task
Agent window
12 h (43 200 s)
Score range
0–110

Difficulty Tier Mechanism

Tier vocabulary
Baseline · Hard · Frontier-defeat
Current status
CANDIDATE
Promotion requires
signed pilot outcome
Evidence rule
frozen bytes + frozen registry
License
CC BY-NC-ND 4.0

Need a custom agent environment or eval harness?

Ethara builds held-out, reproducible agent environments and datasets for frontier teams.

Get in touch