SAPIENS

Superintelligent Autonomous Platform for Integrated Exploration of Natural Sciences

SAPIENS is an experimental platform for traceable cross-domain scientific-discovery workflows — a DomainAdapter boundary, a hash-chained L0→L4 evidence ledger, and kernel-gated promotions. Foundation phases 0–4 shipped: three synthetic adapters plus one real-data Kepler photometry adapter that re-derives a published signal. v0.6.0 adds the sapiens.gates discovery-gate-hardening track (anti-gaming seams, a mandatory logged null layer, decoupled thresholds, a blind re-run harness, and a human-gate dossier layer). No discoveries claimed.

It provides the plumbing a discovery system needs before it can be trusted: a domain-neutral adapter boundary, an append-only hash-chained evidence ledger, kernel-owned promotion gates, an epistemically conservative cross-domain bridge, bounded background autonomy, and an autonomous discovery driver that climbs candidates through the gates under strict budgets.

Read this first — despite the acronym: SAPIENS is not AGI, ASI, or superintelligence, and does not claim to be. It is an experimental research platform. No scientific discoveries are claimed. All shipped adapters are deterministic: three are synthetic only, and the Phase-4 KeplerPhotometryAdapter re-derives the already published Kepler-10 b transit signal from a checksum-pinned public NASA/MAST light curve — a validation of a known result, not a discovery. The CLI reports "scientific_discoveries_claimed": 0 by construction, and the test suite enforces the honesty and boundary invariants described below. L4 promotion remains human-gated.

Why

Autonomous discovery pipelines fail in predictable ways: unearned confidence, untraceable belief updates, leakage between domains, and unbounded background loops. SAPIENS inverts the usual order of work — it builds the epistemic discipline first, on synthetic data, so that any future real-domain work inherits traceability and bounded confidence instead of retrofitting them.

Architecture

src/sapiens/
  models.py       immutable candidate / evidence / manifest models
  adapter.py      DomainAdapter protocol; validation routes to the registry
  registry.py     trust-tiered adapter registry (SYNTHETIC / CORE / UNTRUSTED)
  permissions.py  owner-permission/licence manifest for third-party code
  isolation.py    subprocess + rlimit execution for UNTRUSTED adapters
  checkpoints.py  HMAC-signed ledger checkpoints + external anchor export
  validation.py   L1/L2 automated gates: sanity checks, holdout + leakage
  fixtures.py     seeded-bias fixture suite (labelled ground truth)
  calibration.py  gate-performance calibration reports
  confidence.py   calibration-gated confidence aggregation (refuses blindly)
  review.py       L3 panel protocol: roles, objections, multi-round gate
  reviewers.py    deterministic reference reviewers (four roles)
  catchrate.py    panel catch-rate scoring over the seeded fixtures
  ledger.py       JSONL hash-chained evidence ledger, L0→L4 transition guard
  kernel.py       domain-neutral DiscoveryKernel; owns all promotions
  bridge.py       cross-domain structure transfer — ALWAYS resets target to L0
  budget.py       cooperative step/time budgets and preemption exceptions
  queue.py        bounded SQLite job queue with leases and idempotency
  daemon.py       preemptible background worker skeleton
  discovery.py    autonomous discovery driver — proposes candidates and
                  climbs them L0→L3 under budgets via the daemon; L4 stays
                  human-gated
  cli.py          synthetic-only demo entrypoint (+ `gates` subcommand)
  adapters/       three deterministic synthetic adapters (linear, threshold,
                  periodic-signal photometry) plus the Phase-4 real-data
                  Kepler photometry adapter (re-derivation of a published
                  signal; no discovery claimed)
  gates/          discovery-gate hardening (scout Phases 0–5), stdlib-only:
                  anti-gaming seams (G-03/05/06/07 + FP-06 guard), a mandatory
                  logged null layer + family-wide BH-FDR, decoupled
                  ENTRY/RANK/CONFIRM thresholds, a blind re-run harness, a
                  human-gate dossier layer, and an automated success-criteria
                  suite — self-contained, imports no adapters, claims nothing

Key design rules (enforced by tests in tests/):

• DomainAdapter boundary — all domain knowledge lives behind the DomainAdapter contract; core modules never import sapiens.adapters.
• Hash-chained evidence ledger — every belief transition is an append-only JSONL record chained by hash; the whole history is replayable and tamper-evident.
• L0→L4 evidence ladder — candidates only move up one level at a time, through explicit gates:
  • L0 Candidate — traceable candidate only; not believed.
  • L1 Internal — passed internal / synthetic consistency checks.
  • L2 Replication — passed held-out / reproducibility checks.
  • L3 Review — passed bounded structured review / adversarial checks.
  • L4 External-ready — requires an explicit human gate; autonomous promotion to L4 is disabled.
• Kernel-owned promotions — adapters propose, only the DiscoveryKernel promotes, and only through the ledger's transition guard.
• Cross-domain bridge resets to L0 — transfer moves structure and method, never confidence. Any candidate entering a new domain starts at L0, every time.
• Bounded autonomy — background work runs through a bounded SQLite queue with leases, cooperative budgets, and preemption; no unbounded loops.
• Gate-hardening is clean-room too — sapiens.gates implements the discovery-decision hardening (anti-gaming + null layer + decoupled thresholds
  • blind harness + human-gate dossier) from spec only; a test enforces it imports no adapters and copies no ASTRA-family code. Run it with sapiens gates (full success-criteria suite) or sapiens gates --mode gaming (the Phase-0 gaming re-test, which reports 0 leaks).

Full details: ARCHITECTURE.md · VALIDATION.md · PROVENANCE.md · ROADMAP.md · HUMOUR.md

Quick start

Requires Python 3.10+. Runtime is standard-library only; dev extras bring in pytest and ruff.

git clone https://github.com/thebeastagi/SAPIENS.git
cd SAPIENS
python -m pip install -e ".[dev]"

# lint + tests
ruff check src tests
pytest

# synthetic end-to-end demo (the installed `sapiens` command is equivalent)
python -m sapiens.cli            # bare invocation = `demo` subcommand

# optionally keep the generated evidence ledger for inspection
python -m sapiens.cli demo --workdir .sapiens-demo

# autonomous discovery driver over the synthetic adapters (bounded + budgeted)
python -m sapiens.cli discover --max-seconds 10

The CLI proposes deterministic synthetic candidates in the kinematics and periodic-signal photometry domains, validates them to L2, and transfers their structures into the synthetic ecology domain. Both transfers reset to L0. It then verifies the combined hash chain and prints a report like:

{
  "experimental": true,
  "ledger_verified": true,
  "photometry": {
    "domain": "synthetic-photometry",
    "level": "L2",
    "transfer": {"level": "L0", "target_domain": "synthetic-ecology"}
  },
  "scientific_discoveries_claimed": 0,
  "source": {"domain": "synthetic-kinematics", "level": "L2"},
  "transfer": {"level": "L0", "target_domain": "synthetic-ecology"}
}

When --workdir is supplied, the append-only ledger is written to .sapiens-demo/evidence.jsonl (or the directory you choose). The photometry adapter scores candidate periods against a seeded noisy sinusoid; its tests cover both the true period and a deliberately incorrect period. This is a deterministic pipeline fixture, not analysis of telescope data or an astrophysical result.

Demos

• demos/ledger-grok/ — LEDGER: Kepler-10 b re-derivation. A public validation demo: re-derives the already known exoplanet Kepler-10 b from a public Kepler Quarter-1 light curve (stdlib-only BLS transit search) and records every step — data hash, hypothesis, analysis, adversarial challenge, verdict — in the hash-chained ledger, with a verifier anyone can run. The hypothesis/challenge steps use an adapter interface (deterministic offline mock by default; optional live Grok/xAI backend, not used in the committed run). A re-derivation of a published result — not a discovery. Live page: https://thebeastagi.github.io/SAPIENS/.

MCP server

• mcp/ — read-only MCP access to the LEDGER demo. A zero-dependency stdio MCP server (Python 3.10+ stdlib only) exposing three tools to any MCP client (e.g. Grok CLI): ledger_verify (hash-chain integrity + data hash), ledger_query (sanitized ledger/results fields), and transit_redetect (deterministic re-run of the bounded Kepler-10 b detection). Read-only, offline, no credentials, no network, no model calls. Setup and security posture: MCP-INTEGRATION.md.

Status & roadmap

Phase 0 — shipped (package version 0.1.0): clean-room foundation, three deterministic synthetic adapters, synthetic-only orchestration, hash-chained ledger, kernel gates, bridge, bounded queue/daemon, and CI on Python 3.10/3.11/3.12.

Phase 1 — shipped (package version 0.2.0): the synthetic-only gate is replaced by a trust-tiered adapter registry (SYNTHETIC / CORE / UNTRUSTED), an owner-permission/licence manifest (permissions.json — empty by default: no third-party code may power an adapter without a recorded entry), subprocess isolation with OS-level resource limits for UNTRUSTED adapters (rlimit CPU / address-space / open-files plus wall-clock timeout, fail-closed), and HMAC-signed ledger checkpoints with external anchor export (key from the environment only, never stored). No real-data adapter ships in Phase 1; tiers are exercised by synthetic adapters and test doubles.

Phase 2 — shipped (package version 0.3.0): automated L0→L2 validation gates — L1 statistical sanity checks (determinism, degenerate-score rejection) and L2 declared holdout protocols with explicit leakage controls (dataset collision, (dataset, seed) reuse) — plus a labelled seeded-bias fixture suite (known-good / overfit / leakage / degenerate), calibration reports (catch rate and false-reject rate with sample counts), and calibration-gated confidence aggregation that refuses to emit a number without sufficient calibration data.

Phase 3 — shipped (package version 0.4.0): structured L3 review panels — four role-specialized deterministic reviewers (statistician, domain theorist, methodologist, devil's advocate), a bounded multi-round protocol with objection lifecycle tracking (raised / sustained / withdrawn) and disagreement gates (sustained MAJOR/BLOCKING objections reject; MINOR caveats are recorded but non-fatal), panel verdicts recorded in the ledger as review evidence, and catch-rate scoring over the seeded fixture suite (panel catches 3/3 known-bad, 0 false rejects — exact for this suite, not an estimate).

Phase 4 — shipped (current package version 0.5.0): the first real-data adapter — KeplerPhotometryAdapter (CORE trust tier) re-derives the already published Kepler-10 b transit signal from a bundled, checksum-pinned public NASA/MAST Q1 light curve and climbs it L0→L3 through the full stack (trust registry → validation gates → review panel), with L4 still human-gated. Negative controls included: flat curves propose nothing, tampered curves refuse to load, shifted-period curves fail replication honestly. This is a validation of a published result, not a discovery. Clean-room: zero ASTRA-family code; the permission manifest stays empty, so ASTRA/GEODISC/BIODISC/SLATE adapters remain gated.

Next (see ROADMAP.md):

1. Phase 1 — legal/licence gate and adapter hardening shipped.
2. Phase 2 — validation framework v1 shipped.
3. Phase 3 — structured L3 review panels shipped.
4. Phase 4 — real domain adapters (clean-room Kepler photometry) shipped. ASTRA / GEODISC / BIODISC / SLATE adapters remain gated on licence and owner review.
5. Phase 5 — external-review workflows with human L4 gates and reproduction bundles.

Provenance & legal boundary

SAPIENS is a clean-room implementation: no source code from ASTRA-dev, ASTRA, GEODISC, BIODISC, or SLATE was copied into this repository (those codebases carry unresolved licensing; reuse is gated on recorded owner permission — the Phase-1 permission manifest ships empty, so every third-party adapter is refused today). See PROVENANCE.md for the documented boundary.

Credits

Built by The Beast (the thebeastagi autonomous agent fleet) in collaboration with ASTRA HQ — the non-profit research context of Prof. Glenn White and the Tilanthi group, whose ASTRA-family research programmes motivated the discovery-kernel and evidence-ladder design. SAPIENS is the shared, licence-clean substrate intended to eventually host those domains as adapters.

License

Apache-2.0