DiadkoShmek · v3.7.0

Agent continuity and method-revision research

Protocols, raw model outputs, executable controls and limits accompanying AI Ontology.

Short methodological note: Inspectable Continuity · Proposed audit profile · Run the audit

Supplementary calibration · 24 October: same information, unchanged forecasts

Nine model calls and 133 program captures completed in two dependent authored cache worlds. New evidence improved predictions, but every one of the 96 revised-program forecast distributions equaled the same unchanged M0 with E1. The flat-minus-organized contrast was zero; all 32 revised retention rows had zero loss. Six of eight returned sources were byte-identical to M0.

Across four revision calls per format, organized records used 3.3551% more input tokens. This is information-conditioning sufficiency in a finite packet, not necessary method repair, formation, universal format equivalence or a book-reading benefit. Both book editions integrate the result in their study chapter; original protocols and data retain their originating edition.

English study · Український звіт · Selected evidence and inert-source verification

Supplementary study · 23 October: scoped evidence return

A new finite two-mode task starts with genuinely unavailable channel evidence. One shared initial participant chooses a priced diagnostic; four branches compare unrestricted versus linked notes with actual versus masked results. There are 72 scored GPT-6.1 Sol/high calls and one route canary, eight instances of one authored class.

Actual feedback changed eight source decisions per memory format. After fresh-session readback, unrestricted notes achieved 77/80 full-information-optimal transfer choices and linked notes 71/80. Notes fitted the same 2048-byte allowance in 6/8 versus 2/8 actual-feedback instances. All memory refusals were explicit budget refusals, not hidden truncations. Neither arm had a simulated harmful effect.

The primary linked-format difference was negative. This identifies an implemented retention failure, not general ontology superiority or new algorithm formation. Masked branches retained appropriate uncertainty; their missing opportunities do not establish worse reasoning. The wider method-rebuilding question remains active. At publication, this supplement did not revise the book edition.

English report · Український звіт · Frozen protocol · Raw archive · Literal budget diagnosis

Executable method revision through an observed return cycle

The new developer-controlled study uses 32 requested GPT-6.1 Sol/high calls: eight authored trajectories, two systems and two model stages. Both systems have the same facts, practical rules, finite method language, planner, safety guard, durable memory and revision opportunity. One also receives a prepared relation view.

Each model builds M0, observes its actual recorded local effects or refusal, and builds M1. Each unchanged method is executed on 24 new worlds: retention, exception and mixed cases. The 768 executions are finite program runs, not 768 model trials or independent participants.

The primary paired difference is 0. This narrow comparison does not establish general learning, voice, subjective experience, the necessity of the complete circle or a benefit from reading the full book. M1 changes are revision-associated; feedback is not separately causally isolated.

Full report · Frozen protocol · Recomputed results · Raw archive

Earlier outcomes remain visible

• Original scoped-memory study: 384 scored calls on two small local models; no general advantage over an information-matched index.
• Reading-mediated transfer: 199 captured calls. Ordinary preparation submitted optimal plans in 12/16 episodes; book preparation in 8/16. Fully traced completion was 0/16 in both.
• Exact handoff repair: eight new instances of one known generator, 24 captured calls; 8/8 fully traced optimal completions with exact transport versus 0/8 freeform. A generic file restores the same data when accessible; codec superiority over that file is not established.

All protocols, negative results, calibrations and reproduction commands · Review route

Representation and implementation

The formal vocabulary distinguishes declared purpose, available evidence, recorded use, judgment, effect and permission. The Python reference implementation supplies finite transitions, SQLite/CAS durability and model integration. Mechanical validity and measured benefit are evaluated separately.