Beeline Campaign Agent

Python hackathon agent for AIZheka. It builds a deterministic campaign shortlist, learns from public pilot observations, and selects campaigns under budget and contact limits. The runnable release is offline: no API key, GCP instance, web server, or Vercel hosting deployment is needed.

Release status

Updated 2026-09-23. Milestones 1-5 are implemented. The default Agent.act(env) uses the tested Bayesian strategy and deterministic pilot preflight checks.

portfolio.py is an experimental overlap-aware beam-search selector, not the default, because its measured median regressed. ai_provider.py, semantic_features.py, and verified_cascade.py are experimental modules with offline contract tests. They are not wired into Agent.act. Setting AI environment variables does not enable AI in the submitted agent. Milestones 6-11 are not all complete; publishing this snapshot does not imply otherwise.

Deploy and run

Use Python 3.11. Run all commands from the repository root: the supplied runners resolve data paths relative to the working directory. The repository includes the case inputs and generated shortlist. Keep the original participant package unchanged.

Windows PowerShell

For a new clone (skip cloning if this directory already exists):

Set-Location D:\hackalem
git clone https://github.com/BAITC-Hacks/hack-272a574e-aizheka.git
Set-Location hack-272a574e-aizheka
git switch main
$env:PIP_CACHE_DIR = 'D:\hackalem\.pip-cache'
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.txt
.\.venv\Scripts\python.exe -m unittest discover -s tests -v
.\.venv\Scripts\python.exe local_eval.py
.\.venv\Scripts\python.exe local_eval.py --runs 10
.\.venv\Scripts\python.exe make_submission.py

For an existing clean clone, use git switch main and git pull --ff-only, then the install/test/run commands above. Do not discard local changes to update. If the console cannot encode evaluator output, set $env:PYTHONIOENCODING = 'utf-8'. Calling the environment's Python directly avoids PowerShell activation-policy issues and keeps dependencies on D:.

Linux or WSL

git clone https://github.com/BAITC-Hacks/hack-272a574e-aizheka.git
cd hack-272a574e-aizheka
git switch main
python3.11 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python -m unittest discover -s tests -v
.venv/bin/python local_eval.py --runs 10
.venv/bin/python make_submission.py

Use a separate clone/environment for Linux; a Windows .venv cannot be reused in WSL. If Git on /mnt/d produces chmod/config.lock errors, use Windows Git for that clone, or a Linux-filesystem clone. GPU/cloud compute is unnecessary for the current workload: 20 tests ran in about 2 seconds and ten evaluations in about 4 seconds on the development machine (excluding installation).

Submission deployment

The entry point is agent.py / Agent.act(env), not an HTTP service. For a complete local evaluation deployment, clone the repository and follow setup. For a judging upload, the default agent's explicit file dependency list is:

agent.py
adaptive_pilots.py
action_gate.py
analysis/candidate_shortlist.csv
requirements.txt
submission.csv

Preserve the analysis/ directory. The organizer supplies the environment; local reproduction additionally needs the supplied harness files, dictionaries, customer profile, and data/ from this repository. Confirm with organizers that supporting modules and the shortlist CSV are accepted before submitting. An isolated submission archive has not yet been tested. Do not zip the whole working directory: exclude .env.local, keys, .venv, .cache, and local artifacts.

make_submission.py regenerates seed-42 campaigns. Two runs on this release produced identical bytes and SHA-256:

2c38b5a2e2b2ba26cffa8744de434f34555f346e1bd4c5b2164705ca74b2f53d

The resulting row is adaptive_1_tariff_8_tariff_10, current tariff tariff_8, ARPU segment MID, target tariff_10, channel sms. Hidden-environment decisions may differ; the row is generated, not hardcoded.

Architecture and completed results

Public inputs -> 32-candidate shortlist -> Bayesian adaptive pilots -> resource-aware final selection -> submission. No hidden effects or sampled customer IDs are read by the agent.

• Imported all 16 participant-package files and recorded original hashes in SOURCE_MANIFEST.sha256.
• Data analysis identified disjoint historical/target populations and evidence for 451 of 1,260 transition groups, with median historical support 11.
• Enumerated 5,040 combinations and retained 32 diversified evidence-backed candidates. Historical effects are weak priors, not conversion forecasts.
• Implemented normal-normal posterior updates and value-of-information pilot selection. Exploration is capped at eight attempts, 1,600 contacts, and 20,000 campaign-budget units, with deployment reserves.
• Added a deterministic gate immediately before pilots: validates sample size, tariffs, channel costs, audience freshness, remaining budget, and contacts.
• Current verification: 20/20 offline tests passed. Coverage includes Bayesian updates, resource bounds, rejection fallback, shortlist consistency, AI answer contracts, cache keys, outage handling, evidence claims, and cost accounting. These tests do not establish live model quality.

Before and after: same seeds 0-9

Values below are simulated net ARPU gain, not API spend. Rounded to whole units. Starter measurements are from the earlier baseline report; both current-policy and experimental-portfolio results were rerun on 2026-09-23.

Metric	Starter	Retained Bayesian default	Experimental portfolio
Median	-357,948	488,485	47,307
Minimum	-1,019,431	28,584	43,975
Maximum	-76,493	489,719	488,889
Positive runs	0/10	10/10	10/10
Seed 42 net	-1,035,279	489,248	47,668

Seed	Retained default	Experimental portfolio
0	488,503	45,217
1	489,719	47,362
2	196,745	196,745
3	489,119	46,399
4	488,889	488,889
5	28,707	47,410
6	488,467	43,975
7	489,527	46,935
8	28,584	47,252
9	487,626	487,626

The experimental selector raised the minimum on these seeds but substantially reduced the median, frequently choosing a tiny fallback. It remains available for research, not promoted to the default. Its final-to-final overlap is exact for public ID ordering; pilot overlap is only an expected-coverage approximation because pilot IDs are private.

Current default seed 42: PASS, gross gain 513,448, communication cost 24,200, net gain 489,248, eight pilots plus one final campaign, 2,914 total contacts and 2,644 unique customers. After pilots, 82,680 budget and 13,806 contacts remained.

Previously reported Bayesian seeds 10-29: median 194,221, minimum 319, maximum 637,120, positive 20/20. These were not rerun for this publication and are now known evaluation seeds, not fresh holdout evidence. A rejected confirmation-pilot variant had median -25,172 and only 3/10 positive runs; it was removed.

Mock scores validate mechanics, not future judging performance. Hidden effects differ. No measured Jev or cascade score improvement exists yet.

Vercel / Jev status and credentials

Vercel is used as a potential model gateway, not as project hosting. Optional SDK dependencies can be installed with:

.\.venv\Scripts\python.exe -m pip install -r requirements-ai.txt

Put a Vercel AI Gateway key in ignored .env.local as AI_GATEWAY_API_KEY=..., or supply it through the process environment. Never commit a real value. Do not overwrite an existing .env.local when updating. Verify exclusion with git check-ignore .env.local. Rotate any key previously shared in chat. Ignore rules are not encryption.

The experimental Gateway uses https://ai-gateway.vercel.sh/typesafe and model typesafe-ai/jev, with the TypeSafe SDK appending /v1/systemone. Four independent questions cover offer fit, consumption fit, negative-reaction risk, and evident needs mismatch. Score criteria are ordered arrays; responses retain probability distributions. See the TypeSafe Score contract and Vercel TypeSafe integration.

Live checks on 2026-09-23 found and corrected an invalid Score request shape. The latest corrected request still returned HTTP 403: AI Gateway requires a valid credit card on file. Check billing on the same team that owns the key. No successful inference or provider-reported billed usage was returned; lack of usage data is not proof of zero charges. No OpenRouter fallback or direct OpenAI balance was used. The user's total authorized initial test allowance is $1; this is not an automatically enabled runtime budget.

The helper reserves estimated in-flight costs, limits calls/time, disables SDK retries, and reconciles reported cost when available. These local estimates do not guarantee a provider-side dollar ceiling or persist across processes. Keep cumulative spending records and use provider-side limits before more tests.

The draft -> deterministic evidence check -> Jev verification -> one escalation -> verification helper follows the verified-cascade recipe. It has not been live-validated; draft/escalation models are not selected, and thresholds are not calibrated. Its proposals cannot execute agent actions.

Remaining work

• Revisit portfolio risk/overlap modeling and compare on new evaluation seeds.
• Complete a successful Vercel contract test and representative labeled fixtures.
• Wire optional semantic features into the agent only after paired ablations.
• Pin and test draft/escalation models; assess false acceptance and abstention against human-reviewed labels, not Jev alone.
• Finish broader robustness tests, isolated package reproduction, and organizer confirmation of network, supporting-module, and data-file submission rules.

Detailed records: game plan, baseline, data economics, shortlist, and Bayesian pilots. Earlier reports are historical snapshots; this README describes the current release.