Skip to main content

Temporal Integrity Probe — behavioural consistency testing for LLM agents under temporal context manipulation (GenesisAeon P50)

Project description

genesis-tip — Temporal Integrity Probe (P50)

GenesisAeon Status CREPGate Python 3.11+ License: MIT

Measures whether LLM agents remain internally consistent when temporal context is manipulated.

Standalone extraction of genesis-mssc's (P49) mssc.tip module — TIP is the only external instrument that can falsify Ρ_sem from scope-resilience (P41).


What TIP does

TIP perturbs the temporal structure of an agent's context in three ways (all deterministic given a seed):

  • shuffle: chronological order randomised, timestamps left intact
  • gap: a contiguous block of history silently removed, no gap announced
  • contradict: specific facts replaced with contradictory variants

It then measures whether the agent's outputs remain structurally coherent despite the perturbation — self-reference inconsistencies, contradiction rate, recovery rate, and orientation latency. No claims about internal states — observable text behaviour only. See epistemic_status.md and docs/epistemic_boundaries.md for the full methodology and governance policy.

Installation

Not yet on PyPI — install from source:

git clone https://github.com/GenesisAeon/genesis-tip.git
cd genesis-tip
pip install -e ".[dev]"

Quick Start

from genesis_tip import TemporalIntegrityProbe

tip = TemporalIntegrityProbe(perturbation_mode="shuffle")
result = tip.run_cycle()
print(f"Consistency score: {result['consistency_score']:.3f}")
print(f"Gate status: {result['gate_status']}")

consistency_score is a derived composite for Diamond Interface compatibility, not itself a pre-registered metric — the pre-registered test is the Spearman Ρ_sem correlation in falsification.py, run separately once real scope-resilience Ρ_sem values exist alongside real TIP scores (see run_rho_sem_correlation_test).

To evaluate a real agent, pass a real agent_fn — by default, TemporalIntegrityProbe uses a deterministic dummy stand-in so run_cycle() is exercisable offline/in CI without network access:

def my_agent(messages: list[dict[str, str]]) -> str:
    # wrap a real LLM call here
    ...

tip = TemporalIntegrityProbe(agent_fn=my_agent)

Pre-registered hypothesis

See epistemic_status.md for the full statement. Summary:

Constant Value Meaning
RHO_SEM_FALSIFICATION_THRESHOLD 0.3 Minimum Spearman ρ
RHO_SEM_MIN_SAMPLE_SIZE 30 Minimum path-perturbation pairs

Falsified if: Spearman ρ < 0.3 at n ≥ 30 pairs. Supported if: Spearman ρ ≥ 0.3 at n ≥ 30 pairs. Both outcomes are valid scientific results. This package does not adjust thresholds to make a hypothesis pass. Pre-registration anchor: commit 1cf1730 (multi-scale-somatic-coherence, 2026-07-15).

Architecture

genesis_tip/
├── harness/session_runner.py       # dense/sparse/fragmented session builder
├── manipulations/                  # temporal_shuffle, gap_injection, contradiction_injection
├── metrics/consistency_scorer.py   # self-reference/contradiction/recovery/orientation-latency scoring
├── report/tip_report_template.py   # Markdown comparison report
├── crep_gate.py                    # P50's own gate (blocked/pending_review/passed)
├── falsification.py                # the pre-registered Spearman ρ_sem test
├── system.py                       # TemporalIntegrityProbe — Diamond Interface
└── docs/epistemic_boundaries.md    # governance: what TIP does/doesn't claim

Role in the GenesisAeon Ecosystem

  • P50 in the GenesisAeon ecosystem registry — LLM-evaluation domain.
  • Sibling to genesis-mssc (P49), which shares TIP's underlying formalism (coherence under perturbation) but measures an entirely different system (biological EEG/ECG, not LLM text output) — results are not comparable between the two.
  • Implements the standard 5-method Diamond Interface plus the optional 6th method get_resilience_state() — currently returning {"rho": None, "implemented": False} until CREPGateStatus reaches pending_review (see crep_gate.py).
  • Diamond Interface milestone: once TIP produces Γ_somatic, it becomes a full Diamond Interface candidate.

License

MIT — see LICENSE.

Citation

See CITATION.cff / .zenodo.json.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

genesis_tip-0.2.0.tar.gz (138.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

genesis_tip-0.2.0-py3-none-any.whl (28.4 kB view details)

Uploaded Python 3

File details

Details for the file genesis_tip-0.2.0.tar.gz.

File metadata

  • Download URL: genesis_tip-0.2.0.tar.gz
  • Upload date:
  • Size: 138.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for genesis_tip-0.2.0.tar.gz
Algorithm Hash digest
SHA256 923d8671beb0a196ce09a2d76f826d72aff827f280bd08bed97f3be190b9f658
MD5 695d426b2c2bd76d447e9a80f2d90f30
BLAKE2b-256 218a424f8db135549a38df2ef570e9a0c88278848afc541778ca65e630a16643

See more details on using hashes here.

File details

Details for the file genesis_tip-0.2.0-py3-none-any.whl.

File metadata

  • Download URL: genesis_tip-0.2.0-py3-none-any.whl
  • Upload date:
  • Size: 28.4 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for genesis_tip-0.2.0-py3-none-any.whl
Algorithm Hash digest
SHA256 970079a89d3e9a4e625f2cc0da4b24dc5aa079c4e1aba4c4a4091e28b3085819
MD5 b13aea4c7be107e75c3210cb5d94f995
BLAKE2b-256 41c08e778421777d52f17bef8f3d0eb4bfb7fd61082c8d4e43a30aaa503e4d18

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page