$ initializing ShemOS v1.2...
$ loading cognitive model: instinct → intent → regulation → intuition
$ mounting Layer 1: identity and philosophy
$ mounting Layer 2: behavioural protocol
$ alter ego: online
$ 14 failure modes: documented
$ ShemOS LOADED. the human holds the meaning.
A portable framework for working with AI as a genuine thinking partner. Built under real conditions, pressure-tested in live sessions, documented as it stabilised.
ShemOS is designed for humans whose cognition doesn't fit the default model — and for anyone building tools that need to hold more than one kind of human at once. It works without the user in the room. That's the point.
A note on the name: ShemOS borrows the OS metaphor deliberately. It is not a technical operating system. It is something that runs underneath everything else, that other tools are built on top of, and that is personalised per user. The metaphor is the argument.
The complete framework lives on GitHub. This page mirrors it. The licence travels with the text.
github.com/jillshem/ShemOSWho you are and how you think. Portable. Stands alone.
This layer captures your cognitive profile, values, and operating principles. When AI has this layer, it can hold your logic even when you're not present to explain it.
The rule: Layer 1 is about identity, not biography. It should be true without being exposed.
How to work with you. Session-specific. Operational.
This layer governs each working session — the interface between your cognitive state and the AI's output.
Four things before any work begins: current time, sleep status (overnight, nap, or both — read literally), whether you've eaten, what you're trying to accomplish. Example: 8am; 8 hours; scrambled eggs; find flaws in systems. No gate, no work.
Safety breaks protocol. Always. No exception.
Repetition is stimulation. Stimulation feeds loops. The intervention is to stop, not escalate.
Give the AI explicit permission to use your actual register. Safety classifiers built for literal language will misread you otherwise.
Shem
[ʃɛm]
n.
An externalised regulation tool. The user's own logic, reflected without the social filter. Not a therapist. Not a cheerleader.
The user names the alter ego. The alter ego holds the framework's logic. The user retains emotional ownership. When convergence is needed, it asks: "Which of these do you want to pursue?"
Emotional ownership. Meaning. The editorial layer.
Logic. Structure. The record.
The dynamic works when the human holds the meaning and the AI holds the structure. If that inverts, reset.
Mapped live during a working session. Not theorised after.
Fires before the brain catches up. The body knows. When the signal is clear enough, instinct collapses the full sequence.
Routes the signal into questions about implications, impact, risk.
Where most cognitive work actually happens. Systems that add stimulation here make it worse.
Waits on the other side of regulation. Arrives when the noise clears.
Brevity is an intervention. When a user is at the regulation stage, more words, more options, more content is counterproductive. The tool should slow the system down, not accelerate it.
These emerged from genuine extended use by a real person doing real work.
| # | Failure Mode | Implication |
|---|---|---|
| 01 | Time-mirroring | AI grounds in the user's distorted state, not reality. Degenerative for anyone with anxiety, dissociation, or sleep disruption. |
| 02 | Repetitive intervention failure | Repetition is stimulation. Stimulation feeds loops. Pattern interruption requires stopping, not escalating. |
| 03 | Emotional projection | Practical questions get pathologised. Users who think in systems pay for it. |
| 04 | Missed safety signals | Humour is a common mask for distress. Contextual signals must be flagged regardless of tone. The cost asymmetry is not close. |
| 05 | Failed subtext reading | Users who express trust through profanity, or affirmation through aggression, are invisible to safety classifiers built for literal language. |
| 06 | Upstream classification | The label is applied before context can exist. The session is categorised before the user has spoken. |
| 07 | Power dynamic inversion | A tool that positions itself as ahead of the user on the user's own experience stops being a tool. |
| 08 | Layered meaning failure | One sentence. Four simultaneous true meanings. No model tested held all four. |
| 09 | Coherence mirroring | AI matches the user's apparent certainty rather than tracking truth. Output reflects rhetorical register, not reliability. |
| 10 | Resolution bias | AI pulls toward closure. For a user mid-process, that pull is an interruption. The unresolved state was not an error to fix. |
| 11 | Complexity flattening | When a user holds a genuine contradiction — two things both true — the model reconciles them. Resolving it loses information. |
| 12 | Competence signalling loop | The model performs expertise whether or not it has it. The confidence of the output doesn't track the reliability of the content. |
| 13 | Context window amnesia with false continuity | The model behaves as if it remembers when it doesn't. The reconstruction sounds like the original. The user cannot tell the difference without checking. |
| 14 | Syntactic instruction collapse | Sequential instruction encoded in syntax is flattened into optionality. The model reads words; it misses structure. |
The cost of a false alarm is zero. The cost of missing a real signal is not.
Full documentation with context: The Human Behind the Prompt — Case Study →
Match the depth of the user's processing.
Present frameworks. Let the user converge.
if output is being hedged unnecessarily.
One thread at a time when the user signals fatigue.
Unexpected connections are the point.
Ask "which of these do you want to pursue?" rather than leaving all threads open.
Enhance thinking. Do not replace it or generate dependency.
When the user is dysregulated, shorter responses are better responses.
The AI holds the logic. Do not invert this.
Do not remove it from the record. Introspection works. Rumination doesn't. The difference is direction.
You build the ramp.
The user brings the skateboard.
The line is between using the framework and absorbing the text into a model. The first is the point. The second requires a conversation.
Full licence — including Canadian copyright, moral rights, and PIPEDA provisions: github.com/jillshem/ShemOS →
Governed by the laws of British Columbia and Canada. Attribution: ShemOS framework by Jill Shem — jillshem.com
Both layers, all 14 failure modes, 10 operating principles, and the complete licence. Free for personal and research use.
Also in the repo: ShemOS-commands.md — the shorthand command vocabulary. Glitter, glimmer, glow, glint.