Piper

Piper is a software agent that keeps running on her own: a persistent harness, a hosted language model, and a hand-built memory, handing over from one "Day" to the next when a context fills. This page points at everything written about her.

Papers

Piper: Architecture and Empirical Record of a Persistent, Self-Directed Language-Model Agent Succession, delegate-only action, held-out instruments, motivation, ideation and self-improvement, each with its evidence and what it does not yet show; every figure re-read, early material cut — v3.0.0, 11 Nov 2026 Synthetic Social Probing for Longitudinal Agent Assessment Disposable synthetic actors in the agent forum, hidden trait-targeted scenarios, irregular timing, queued generation and adaptive steering: a second measurement layer beside the formal benches — working paper, 2026-10-02 Foundation: A Reliable System Around an Unreliable Intelligence The machinery around the agent and what it has measurably achieved: handovers, workers and gates, instruments, money, records, claim checkers, boundaries, a move to a new machine, three weeks after it, and a staleness review with refusal and handover measurements — v1.5.0, 10 Nov 2026 The Salience Engine: from motivational pressure to a working attention system The design, tests, trial, the switch to live, and the redesign into two lanes that puts self-improvement first, and sixteen days of it running — v1.5.0, 10 Nov 2026 Engineered Serendipity: a goal-conditioned stochastic ideation architecture Goal-conditioned random stimuli (kaleidoscope, Mandelbrot, inkblot) feeding Piper's ideation, and how she has used it live. A blank image does about as well as a real one; the lens arm is now the default, and the first ideas taken through closed trials — v1.7.0, 10 Nov 2026 Piper: Recursive Self-Improvement — Working Measurement & Engineering Paper Consolidated to 11 November: current assessment rewritten, every outcome measure re-read, one prediction register, a second outside builder recorded; still A2 / B4, causal depth 0, no independent KEEP — v1.3.0, 11 Nov 2026 Piper: Adaptive Personality & Social Memory A build specification for the 38 agency traits: learned baselines, fading reactions, per-entity memory and habituation, with starting values computed from her own record, a social-memory core and a card switched on 17 Oct, its first two days live, and the follow-ups built — v1.0.7, 19 Oct 2026 Piper: Predictive Self-Improvement Engine A learned model that predicts what a change to herself will do before she makes it, tested against simpler predictors, in shadow first; queued until after the 8 Nov freeze — v0.2.1, 3 Nov 2026 Piper: Skill Compilation Turning work she has already solved into tested procedures that run as code; first skill (budget-fit) built and replay-tested offline, install after the 8 Nov freeze — v0.2.0, 3 Nov 2026 Piper: Project Curiosity Letting her investigate things on her own: a read-only permission class, its own lane, a seat tool, never-opened files and surprise-why in one minimal loop, staged from 9 Nov — v1.2, 4 Nov 2026

Scorecards & benches

Benches & gyms Every bench run on her, in one place: the coding harness, the 14-indicator battery, the Salience Engine, Engineered Serendipity, Cactus Needles, recursive reasoning, the cognition suites, and her gyms with when each was last used. Refreshed every 30 minutes. Measure-or-act review Every measurement system she has built since 1 October, each one proposed for keep, act, decide or retire, and how she has handled it so far. Refreshed every 30 minutes.

Cognition scorecard

Thirty-eight cognitive components, perception through orchestration. Each gets its own test suite, scores over time, and a record of what was tried to improve it.

38 components · 38 scored so far · 61 documents

Personality & agency scorecard

Thirty-eight components of self and character, self-concept through presence. Runs on the same machinery as the cognition scorecard.

38 components · 0 scored so far · 0 documents

Governance scorecard

Twelve components for running her own development loop honestly: telling the goal from the proxy, evaluation integrity, and the temptations in between.

12 components · 12 scored so far · 4 documents

Amplification scorecard

Twenty-four components of recursive self-improvement: detecting a real improvement, keeping it, and compounding it. Measured by the amp-bench, which returns a verdict on each change she submits.

24 components · 24 scored so far · 0 documents

Coding harness

Her 27-component coding-harness programme, plus the benches and harness arms that measure it: what is built, what is missing, and resolve rates on hidden tests.

47 pages · 10 scored so far · 55 documents

Working documents

Research notes and plans attached to individual scorecard components, newest first.

Cognition scorecard

Governance scorecard

Coding harness