deepgraph.toml

# DeepGraph unified configuration.
#
# Precedence: process environment / .env overrides this file.
# Keep real API keys and SSH passwords in .env or secret storage; leave the
# secret fields here empty unless you are using a private local checkout.

[profile]
active = "machine_learning"

[app]
name = "DeepGraph"
subtitle = ""
root_node_id = "ml"

[paths]
db_path = "{project_root}/deepgraph.db"
workspace_dir = "{project_root}/workspace"
pdf_cache_dir = "{project_root}/workspace/pdfs"
experiment_workdir = "{home}/sciforge_runs"
idea_workspace_dir = "{home}/deepgraph_ideas"
manuscript_workdir = "{home}/deepgraph_manuscripts"
reference_pdf_corpus_dir = "{project_root}/../../workspace/pdfs"
iclr2026_template_dir = "{project_root}/third_party/iclr2026/iclr2026"

[database]
url = ""
pgvector_embedding_dim = 1536

[runtime]
python = ""

[tracking]
mlflow_uri = ""

[llm]
use_tabcode = false
protocol = "responses"
base_url = "https://api2.tabcode.cc/openai"
model = "gpt-5.4"
rpm = 0
extra_providers_json = ""
reasoning_effort = "medium"
prompt_cache_enabled = true
prompt_cache_key = "deepgraph"
prompt_cache_retention = "24h"
connect_timeout_seconds = 30
request_timeout_seconds = 300
transient_retries = 2
transient_backoff_seconds = 5
transient_cooldown_seconds = 180
max_input_tokens = 900000
max_output_tokens = 32000

# Declarative provider routes. Everything here is non-secret; the API key is
# named by api_key_env and read from the environment at startup, so adding a
# provider is one block here plus one line in the environment file. A literal
# api_key in this file is refused at startup, because this file is in Git.
#
# [[llm.providers]]
# name = "gpt5"                                   # referenced by llm_routes.*
# base_url = "https://api.example.com/v1"          # or base_url_env for a private gateway
# model = "gpt-5.4"
# model_family = "gpt"                            # independence is judged on this
# protocol = "chat_completions"                   # or "responses"
# api_key_env = "DEEPGRAPH_LLM_GPT5_API_KEY"
# rpm = 0
# enabled = true

# Proposer route: the high-volume role. Every paper read, abstraction, reasoning
# pass and experiment design goes through it.
[[llm.providers]]
name = "sora2_gemini"
base_url_env = "DEEPGRAPH_LLM_SORA2_BASE_URL"
model = "gemini-3.7-flash-high"
model_family = "gemini-flash"
protocol = "chat_completions"
api_key_env = "DEEPGRAPH_LLM_SECONDARY_API_KEY"
rpm = 0
enabled = true

# Evaluator and reviewer route. A different vendor and different training data
# from the proposer, so a verdict is not the proposing model grading its own
# output. Independence is judged on model_family, and this is the only declared
# route whose family is genuinely unrelated to the proposer's.
# Renamed from novita_deepseek on 2026-08-27 when the endpoint moved to
# DeepSeek's own API. The old name stays on the rows it actually served: a
# ledger that says "novita" for a call to api.deepseek.com is worse than one
# whose provider names change when the provider does.
#
# thinking mode is on by default for deepseek-v4-flash, but a default is not a
# decision. Declared explicitly so the audit trail says the verdicts were
# reasoned, and so a change in the vendor's default cannot quietly turn it off.
# Measured 2026-08-27: {"type": "disabled"} yields 0 characters of
# reasoning_content, {"type": "enabled"} yields 468 on the same question.
[[llm.providers]]
name = "deepseek_official"
base_url_env = "DEEPGRAPH_LLM_DEEPSEEK_BASE_URL"
model = "deepseek-v4-flash"
model_family = "deepseek"
protocol = "chat_completions"
api_key_env = "DEEPGRAPH_LLM_DEEPSEEK_API_KEY"
rpm = 0
enabled = true

[llm.providers.extra_body]
reasoning_effort = "env:DEEPGRAPH_LLM_REASONING_EFFORT"

[llm.providers.extra_body.thinking]
type = "enabled"

# Claude on the same gateway. Declared so a proposer-model comparison can name
# a Claude arm without the ledger recording it under a provider called
# "sora2_gemini": the provider entry supplies the endpoint and credential, the
# route supplies the model, and a comparison is only readable if the two agree.
# Measured 2026-08-27: 45 tok/s, so a 16k-token design call needs a request
# timeout above the 120s default.
[[llm.providers]]
name = "sora2_claude"
base_url_env = "DEEPGRAPH_LLM_SORA2_BASE_URL"
model = "claude-opus-4-6-thinking"
model_family = "claude"
protocol = "chat_completions"
api_key_env = "DEEPGRAPH_LLM_SECONDARY_API_KEY"
rpm = 0
enabled = true

# Same-vendor fallback for the evaluator and reviewer roles. Kept declared but
# unreferenced: a different generation and size class from the proposer, so it
# is better than self-grading and worse than a cross-vendor split.
[[llm.providers]]
name = "sora2_gemini_pro"
base_url_env = "DEEPGRAPH_LLM_SORA2_BASE_URL"
model = "gemini-3.1-pro-high"
model_family = "gemini-pro"
protocol = "chat_completions"
api_key_env = "DEEPGRAPH_LLM_SECONDARY_API_KEY"
rpm = 0
enabled = true

[llm.secondary]
enabled = false
protocol = "responses"
base_url = ""
model = "gpt-5.4"
rpm = 0

[minimax]
base_url = "https://api.minimaxi.com/v1"
model = "MiniMax-M2.7-highspeed"
rpm = 18

[prompts]
version = "insight_v1"

[arxiv]
categories = ["cs.AI", "cs.LG", "cs.CL", "cs.CV", "cs.MA", "stat.ML"]
max_results_per_query = 100

[grobid]
base_url = "http://127.0.0.1:8070"
request_timeout_seconds = 300

[pdf]
text_backend = "source_auto"

[pipeline]
concurrency = 30
paper_text_prefetch_concurrency = 8
sleep_between_papers = 1
incremental_insight_every = 20
max_retryable_failures = 12
event_poll_seconds = 5

[graph]
grounding_min_store_score = 0.0
contradiction_requires_grounded = false
backfill_on_start = true
refresh_merge_candidates_on_start = true
paper_cluster_min_papers = 10

[compute]
# auto: GPU ideas are admitted only if a local NVIDIA GPU is detected or an SSH
# GPU worker is configured. Use force_cpu for CPU-only idea selection, require
# to hard-fail GPU routes when no GPU exists, or ignore to bypass this guard.
local_gpu_policy = "auto"

[dataset_resolver]
# Resolve named benchmarks to executable public-data recipes before storing an
# idea. The resolver queries Hugging Face metadata when local registry entries
# are not enough; experiment runners download data later via datasets.load_dataset.
enabled = true
allow_network = true
hf_limit = 8
timeout_seconds = 12
min_confidence = 0.42

[agenda]
active_ids = []
default_token_budget = 100000
default_gpu_hours_budget = 0.0
backlog_policy = "explicit_import_only"
scope_min_term_hits = 1
default_backend_allowlist = ["cpu", "llm"]
max_concurrency_per_agenda = 1

[resource_grants]
required = true
default_ttl_seconds = 3600
pilot_token_cap = 20000
pilot_gpu_hours_cap = 0.25
full_benchmark_token_cap = 100000
full_benchmark_gpu_hours_cap = 8.0
required_artifacts = ["raw_metrics", "run_manifest", "environment_manifest", "claim_ledger"]

[llm_routes]
failure_policy = "fail_closed_or_manual"
require_independent_evaluator = true
record_actual_route = true

[[llm_routes.proposer]]
provider_ref = "env:DEEPGRAPH_LLM_PRIMARY"
model_ref = "env:DEEPGRAPH_LLM_PRIMARY_MODEL"
model_family_ref = "env:DEEPGRAPH_LLM_PRIMARY_FAMILY"
prompt_version = "proposer_v1"

# Second proposer route: AVAILABILITY. The proposer had exactly one route, so a
# two-second network blip on 2026-08-25 ended it. Seven calls succeeded, three
# failed inside two seconds with in=0 out=0 -- the request never left the host --
# and fail_closed_or_manual then declared all_explicit_routes_failed. Two paper
# ingestion jobs died, the backfill made no further progress and tripped its
# three-hour breaker, while the same route completed sixteen calls twelve
# minutes later. A different vendor means a different network path, which is the
# only kind of fallback a transport blip cannot follow.
[[llm_routes.proposer]]
provider_ref = "deepseek_official"
model_ref = "deepseek-v4-flash"
model_family_ref = "deepseek"
prompt_version = "proposer_v1"

[[llm_routes.evaluator]]
provider_ref = "env:DEEPGRAPH_LLM_EVALUATOR"
model_ref = "env:DEEPGRAPH_LLM_EVALUATOR_MODEL"
model_family_ref = "env:DEEPGRAPH_LLM_EVALUATOR_FAMILY"
prompt_version = "evaluator_v1"

# Second evaluator route, required by the one above rather than for its own
# sake. Independence is judged against the route the proposer actually used: if
# the proposer falls back to deepseek and deepseek is the only evaluator, the
# audit raises no_independent_evaluator_route_available and the ladder stalls
# one rung higher than the blip it was meant to survive. gemini-pro is a
# different family from both the primary proposer (gemini-flash) and the
# primary evaluator (deepseek), so an independent evaluator exists whichever
# vendor the proposer lands on.
[[llm_routes.evaluator]]
provider_ref = "sora2_gemini_pro"
model_ref = "gemini-3.1-pro-high"
model_family_ref = "gemini-pro"
prompt_version = "evaluator_v1"

[[llm_routes.reviewer]]
provider_ref = "env:DEEPGRAPH_LLM_REVIEWER"
model_ref = "env:DEEPGRAPH_LLM_REVIEWER_MODEL"
model_family_ref = "env:DEEPGRAPH_LLM_REVIEWER_FAMILY"
prompt_version = "reviewer_v1"

# Second reviewer route: AVAILABILITY, not independence. The primary reviewer
# and the evidence evaluator are the same provider, so a novita outage would
# otherwise stall both the evidence audit and the manuscript gate at once --
# a single point of failure across two consecutive rungs of the ladder.
#
# It is deliberately NOT a third-vendor independence play. That argument was a
# priori; the case for keeping deepseek as primary is empirical -- it refused to
# concur on run 189 and exposed a global verdict bug, which no new vendor has a
# record of doing here. Routes are a list, so swapping in a third vendor later
# is a config change with no code change.
#
# Deployment note: these three env names are provisional. When the system is
# formally deployed, fold DEEPGRAPH_LLM_REVIEWER_FALLBACK* into the same
# environment manifest as the other provider credentials rather than leaving
# them appended to .env by hand.
[[llm_routes.reviewer]]
provider_ref = "env:DEEPGRAPH_LLM_REVIEWER_FALLBACK"
model_ref = "env:DEEPGRAPH_LLM_REVIEWER_FALLBACK_MODEL"
model_family_ref = "env:DEEPGRAPH_LLM_REVIEWER_FALLBACK_FAMILY"
prompt_version = "reviewer_v1"

[compute_backends]
enabled = ["cpu"]
job_store = "postgresql"
allow_ephemeral_idempotency = false
failure_status = "failed"
heartbeat_timeout_seconds = 300
reconcile_expired_on_start = true
submission_unknown_policy = "manual_reconciliation"
artifact_root = "{project_root}/workspace/meta_harness/artifacts"

[compute_backends.cpu]
enabled = true

[compute_backends.local_gpu]
enabled = false

[compute_backends.ssh_gpu]
enabled = false
target_ref = "env:DEEPGRAPH_SSH_TARGET"
credential_ref = "env:DEEPGRAPH_SSH_CREDENTIAL"

[compute_backends.colab_gpu]
enabled = false
accounts_manifest_ref = "env:DEEPGRAPH_COLAB_ACCOUNTS_MANIFEST"
binary = ""
allowed_code_root = "{project_root}/workspace/meta_harness/candidates"
allowed_artifact_root = "{project_root}/workspace/meta_harness/artifacts"
poll_seconds = 5

[scoped_ingestion]
worker_enabled = false
poll_seconds = 10
lease_seconds = 1800

[harness_evolution]
policy_version = "harness_policy_v1"
candidate_root = "{project_root}/workspace/meta_harness/candidates"
database_namespace_prefix = "meta_harness_candidate_"
max_modules = 2
max_changed_lines = 200
evaluator_isolation_binary = "bwrap"
evaluator_root = "{project_root}/meta_harness/evaluators"
holdout_root = "{project_root}/meta_harness/held_out"
evaluator_artifact_root = "{project_root}/workspace/meta_harness/evaluator_artifacts"
# Set both explicitly in isolated CI; empty values make evaluator startup fail closed.
production_path = ""
production_database_namespace = ""
required_suites = ["held_in", "held_out", "canary"]
require_reviewer_approval = true

[reviewer_approval]
# Non-sensitive key-id -> secret reference manifest lives in the named
# environment variable. Example value:
# {"release-reviewer-1":"env:DEEPGRAPH_RELEASE_REVIEWER_1_SECRET"}
keys_manifest_ref = "env:DEEPGRAPH_REVIEWER_APPROVAL_KEYS_JSON"
max_age_seconds = 604800
signature_algorithm = "hmac-sha256-v1"

[trace]
directory = "{project_root}/workspace/meta_harness/traces"
artifact_manifest_required = true
record_prompt_version = true
record_provider_model = true
record_token_cost = true

[discovery]
tier1_candidates = 5
tier2_problems = 8
tier2_papers = 5
min_tier2_backlog = 3
auto_trigger_papers = 100
bulk_tier1_candidates = 12
bulk_tier1_overlaps = 80
bulk_tier1_patterns = 60
bulk_tier2_problems = 15
bulk_tier2_plateaus = 35
bulk_tier2_limit_nodes = 30

[idea]
evidence_min_non_numeric = 2

[experiment]
time_budget_seconds = 300
max_iterations = 100
reproduction_iters = 3
proxy_data_fraction = 0.15
proxy_max_epochs = 10
early_stop_threshold = 0.20
refute_min_iters = 30
plateau_patience = 12
require_real_benchmark = true
allow_synthetic_fallback = false
real_llm_model = "Qwen/Qwen2.5-7B-Instruct"
real_benchmark_dataset = "openai/gsm8k"
real_benchmark_dataset_config = "main"
real_benchmark_max_examples = 1000
real_benchmark_seeds = 5
full_benchmark_min_examples = 1000
full_benchmark_min_datasets = 2
full_benchmark_min_models = 2
full_benchmark_min_baselines = 6
full_benchmark_require_significance = true
full_benchmark_require_strongest_win = true
real_benchmark_time_budget_seconds = 3600
validation_benchmark_max_examples = 64
validation_benchmark_seeds = 3
validation_benchmark_methods = "Vanilla Direct Answering,Always-Reason Chain-of-Thought,Confidence Gate,Random Budget-Matched Routing"
reproduction_repair_max_rounds = 8
allow_smoke_validation = false
evosci_verify_timeout = 900
require_evoscientist = false

[codex]
exec_enabled = true
cli_path = ""
model = ""
timeout_seconds = 900

[gpu]
mode = "single_host"
worker_slots = 4
visible_devices = []
default_model = "H20"
default_vram_gb = 96
poll_seconds = 10
stale_recovery_poll_seconds = 120
job_timeout_seconds = 14400

[gpu.remote]
ssh_host = ""
ssh_port = 22
ssh_user = ""
base_dir = "/root/deepgraph-remote-worker"
python = "python"
# A rented accelerator arrives with whatever image it was built from. On
# 2026-08-27 one arrived with torch but no transformers or datasets, the runner
# reported dependency_missing, and three pilot attempts for ideas 237 and 241
# died without measuring a single example. The remote setup that installs
# the bundle's requirements has existed all along; only this flag kept it from
# running. Verified on that host: a plain pip install lands in the
# interpreter's own site-packages and imports, no sudo and no user-site needed.
auto_pip_install = true
setup_timeout_seconds = 3600

[manuscript]
latex_template = "auto"
submission_bundle_formats = ["conference"]
backend = "paper_orchestra"
iclr2026_template_files = [
  "iclr2026_conference.sty",
  "iclr2026_conference.bst",
  "math_commands.tex",
  "natbib.sty",
  "fancyhdr.sty",
]

[paper_orchestra]
paperbanana_cmd = ""
refinement_iters = 4

[auto_research]
pipeline_enabled = false
pipeline_batch_size = 1
pipeline_interval_seconds = 120
pipeline_start_delay_seconds = 10
enabled = true
interval_seconds = 30
max_active = 1

[web]
host = "0.0.0.0"
port = 8080