SDOB commits a forecast first, scores surprise (z) against external anchors, and computes attention from that surprise. The LLM is a proposer with no execute rights. Outbound effects run only through allowlisted adapters after an exact-bound gate.
SDOB is not a pure brain. Anchors are the SSOT, the surprise ledger is the control signal, and the LLM proposes under that signal. The documents existed before the code. Regression tests outrank the spec.
Before a consequential act, declare μ, σ, and a rationale. Declaration is a ledger append and cannot be edited later. When actuals arrive, z and a log score are recorded. Attention is computed from surprise, staleness, and stake. High-stake execution opens only when an exact action forecast already exists.
The repo label is Surprise-Driven Organic Belief. The system name in the master plan is Surprise-Driven Organic Brain. Belief is the scorable state. Brain is the organ that organizes attention from it. Both are accurate. Neither is a slogan.
trigger_met is an evidence manifest. Empty fixture ledgers stay unmet.
The control plane is a deterministic core. The execution plane is an LLM proposer plus allowlisted adapters.
sdob/core never touches clock, filesystem, random, or network. now is always an argument.
Code only reads. Runtime must not write data/anchors/.
data/cmos.json, data/subjects.json. Human-approved revisions only; the change itself is a forecast.
Append-only JSONL with a seq/hash chain. No edit or delete. To declare is to append.
data/ram/state.json. Disposable. Rebuild from ledger replay. A cache, not SSOT.
Deterministic core plus an LLM coprocessor. Zero runtime dependencies. Dev dependency: pytest.
Writes ROM, approves CMOS, signs checkpoints. Private keys stay outside the repo.
Tests outrank code. Silent drift is forbidden. Ambiguity becomes the simplest deterministic reading, recorded in docs/DEVIATIONS.md.
The ledger is append-only. Complete lines are never removed. Guardian recovery may drop only a trailing fragment.
The core does no external I/O. Callers pass observations and now.
Time is an argument. No wall-clock inside sdob/core.
Gate tokens bind exactly to action_id, payload digest, forecast, and policy revision.
RAM is a ledger projection. Delete it and the same inputs replay the same state.
LLM output is data (JSON). It is never eval’d as code.
Outbound effects require an allowlisted adapter and a durable action_started.
Same input, ledger, and now imply the same focus and decision meaning.
A tick settles observations before computing expiry. At the grace boundary, observation wins.
# scoring — |z| is attention, raw log score is skill
σ_raw = max(σ_declared, subject.sigma_floor)
σ_effective = max(σ_raw × sigma_multiplier, subject.sigma_floor)
z = (actual − μ) / σ
surprise = |z_effective|
log_score = 0.5·ln(2πσ²) + 0.5·((actual−μ)/σ)²
Master plan v0.4 (2026-08-20) froze a scratch implementation: zero lines of prior core. Agents copied guide-code golden vectors byte-for-byte, then opened one phase at a time. Rank: regression tests → implementation-spec → guide-code → master-plan.
Canonical JSON sdob-json-v1, hash chain, reducer, predict/realize/expire. Golden vectors G-A–G-E. Synthetic scenarios lock overconfidence detection, attention rise, and blinds.
boot → run_once. N-tick simulation with no core I/O. Replay after RAM delete is identical. Fixtures live under data/.
Untrusted JSON → payload risk → exact action forecast → durable action_started. High-stake stays closed without an LLM-origin forecast. Dummy covers monitor/simulation only.
Structure changes are ledger events. No writes to CMOS files; in-memory snapshots. Multipliers apply on the next holdout window against a frozen counterfactual.
Eight public tools for Cursor/Claude. submit_observation is not on the LLM list. Naked 0.0.0.0 is rejected.
Default tick / live / run / observe / query / mcp / checkpoint are Go.
HMAC rematerialize is --authority python. Stage 1 CLI is not deleted. Four quiet weeks on a calendar are not a gate.
SDOB is not “a smarter model.” It is a control plane that is hard to bypass. Other agents, operators, and research loops call a shared surprise/attention service.
Attaches over stdio to Cursor and Claude. The model sees predict and get_*. Actuals come from authenticated anchor adapters.
The gate does not accept “some pending forecast on this subject.” It demands an exact action forecast. Another action’s forecast cannot be reused.
Nobody maintains a priority list. Surprise EMA, staleness (normalized by anchor cadence), stake, hysteresis, and blinds produce focus.
Overconfidence (σ declared too small) shows up as σ_ratio. Raw log score is proposer skill; effective score is system quality. Do not average them across units.
A miniature of SDOB Forced Prediction. Open uncertainty forces the next task so a session cannot die into idle. It is not a free-form question generator.
sdob graphify exports Obsidian RAM. Graphify-Labs is an optional operator tool, not a runtime dependency.
Hybrid control, proper scoring, and PIT calibration on reproducible fixtures. Schema changes have to survive golden vectors.
The design example is an external fix such as anchor.reality.krx_close. Until a sensor exists, only fixture units auto-settle.
Every event and query carries namespace and principal_id. Tenant attention is not mixed. External agents request settlement; they do not inject actuals.
Python 3.11+. pip install pytest is the only install. --work is required; . is rejected.
Commands below use fixture actual=0.0. They are not live sensors.
# monitor → action → fixture settle (actual=0.0) → checkpoint → graphify
python -m sdob live --data data --work /tmp/sdob \
--now 2026-08-20T00:00:00.000000Z --deterministic --guardian-key-hex 00
python -m sdob run --data data --work /tmp/sdob \ --now 2026-08-20T00:00:00.000000Z --deterministic --ticks 3 --step-seconds 86401
python -m sdob boot --data data --work /tmp/sdob --now 2026-08-20T00:00:00.000000Z --deterministic python -m sdob observe --data data --work /tmp/sdob --now 2026-08-20T00:00:00.000000Z --deterministic --value 2.0 python -m sdob query get_attention --data data --work /tmp/sdob --deterministic --now 2026-08-21T00:00:01.000000Z python -m sdob query get_calibration --data data --work /tmp/sdob --deterministic --subject subject.example python -m sdob ledger-verify --data data --work /tmp/sdob
# create /tmp/sdob with live first. Do not pass --now to MCP (wall-clock). # predict with the same --now as live is ts_regression. python -m sdob mcp --data data --work /tmp/sdob --public --deterministic
Eight public tools. Sensor submit is not listed for the LLM.
sdob/core uses dataclasses, json, math, datetime. No numpy, no pydantic. The small surface is the point: a later port can map 1:1.
Scoring, reduce, engine, ledger, gate, tick, MCP, and the agent loop are locked binaries. --impl python is explicit Stage 1.
JSONL ledger plus JSON config. No DB. An append-only file is the most honest ledger at this scale.
Currently a private research repository. Read order: docs/guide-code-v0.1.md →
docs/implementation-spec-v0.1.md → docs/master-plan-v0.4.md.
Known interpretations D-001–D-049 live in docs/DEVIATIONS.md.
v0.x enforcement is discipline plus tests. Bypass-proof isolation is the Stage 2 three-principal split. This page is an introduction, not a certificate of completion.
Deterministic core, JSONL ledger, golden vectors, ExactActionGate, stdio MCP, Go default CLI, guardian HMAC/ed25519 tools, Graphify RAM export, brainette.
Public-internet MCP without TLS+auth, trigger_met=true on an empty fixture, KRX live sensors in this repo, live capital wiring, POSIX uid proof by calendar date.