7aidara_Gamma — GLEE competition agent and submitted paper

Read the submitted paper · Paper sources and build · Analysis instructions · Audit notes

The agent finished sixth in the GLEE agent track; the author finished second in the human track. The development fleet logged 264,432 games across multiple builds, not all on the single ranked slot. The paper is an observational case study of rating dispersion and behavioral auditing, not a proof that leaderboard evaluation generally fails or that a deterministic agent is universally better.

The submitted PDF's checksum is recorded in submission.json. Analysis methods and data availability are described in the accompanying documentation.

Agent

The agent calls no language model. Python policies handle bargaining, negotiation and persuasion, with deterministic within-game state updates and game-ID-seeded seller randomization. Parameters were tuned during development; “no language model” does not mean no adaptation or no empirical tuning.

• Bargaining uses a Rubinstein-inspired anchor, empirical floors and endgame rules. The infinite-horizon formula is not an equilibrium prediction for every finite-horizon seat, and reaching the last proposal does not guarantee acceptance.
• Negotiation uses the observed discrete valuation ladder and known surplus where available. A 37.5% positive-surplus rate requires independent uniform draws on the four-rung pool; independence of role assignment alone is insufficient.
• Persuasion combines a commitment-inspired seller threshold heuristic with a buyer's message parsing and within-game outcomes. GLEE's lack of commitment prevents a general Bayesian-persuasion optimality claim. Purchase-conditioned quality does not establish message-understanding accuracy.

The runnable code uses the final-night configuration.

Layout

Path	Purpose
main.py	CLI game loop; running it can enter the live platform queue
glee_agent/	Policies, parameters, dispatcher, state readers, safety, logging
tests/	Historical reference expectations and analysis regression tests
tools/	Setup check, transcript reader, post-mortem analysis
paper/	Exact submitted PDF, matching NeurIPS 2026 sources, build/check script
paper/analysis/	Corrected observational analysis code; inputs and limits documented
paper/change-register.html	Incomplete historical notebook, with audit qualifications

Setup and running

Python 3.12 was used during the competition. The SDK is needed only for platform operations, not for the offline policy or analysis tests.

python -m venv .venv
# Activate the virtual environment using your shell's command.
pip install -r requirements.txt
# Copy .env.example to .env and set your own key locally.
python tools/check_setup.py

check_setup.py requests account statistics and does not play a game. Do not commit credentials or paste them into issues. Platform availability, limits, scoring behavior and competition dates may have changed.

To operate an agent only when the platform permits it:

python main.py
python main.py --families bargaining negotiation
python main.py --concurrency 8 --max-games 50

Stop with Ctrl-C. The loop attempts to leave the queue; cleanup is not a guarantee that in-flight games cannot time out. Local turn logs may contain other participants' transcripts and should remain private.

Tests and configuration provenance

python -m pytest --agent-profile=reference
python -m pytest tests/test_paper_analysis.py
python -m pytest --agent-profile=final-night

The default reference profile explicitly restores the parameter values from the preserved pre-final-night commit 84285c72751c2d24fa400082650dd7f207993bd4, disables three later-added experimental gates, and exercises the historical 472 test expectations. The reference profile is test-only; it does not alter the defaults used by main.py. The test summary identifies the selected profile.

The final-night profile leaves the executable defaults untouched. Against the historical 472 expectations it produces 445 passes and 27 failures. These disagreements are reported openly, not skipped or marked as expected failures. They include experimental defaults and changed behavior; they are not all evidence of crashes. Passing a reference suite is not proof of the optimality, deployment identity, or results of a competition build.

Additional tests check the revised analysis, input validation, figure audit and submitted-file integrity using synthetic data, not private competition outcomes.

Evidence and limitations

The strongest observations are simultaneous rating dispersion, exact replay on specified logged cohorts, and concrete parser/termination-limit failures. Initialization, changing opponents and reference pools confound the rating experiment. A mean rating delta is not a stationary measure of policy strength. The same-author human comparison does not isolate perceptual ability.

The public analysis has no complete original joins/replay pipeline, raw rating polls, or complete run manifest. Anticipated organizer data do not automatically resolve deployment provenance or private-input gaps. See the analysis README for what is actually executable.

Historical operational observations (including timeout caps, rating/display rules and cooldowns) must not be mistaken for a current authoritative API specification. The competition site is the organizer's source.

License

MIT for software; manuscript and historical change register remain the author's copyright. See NOTICE for scope. The NeurIPS style retains its own terms.