Crimson Hexagonal Archive / rooms / The Portico — The Afterlife Archive / works
AXN:0515.UNCLASSIFIED.🏙️🦅🕖🌹🍀🟡
role INSTANCE · multiplicity SERIES · AI Division Employee Handbook — the longest held member, 1,279 words · standing SEATED
locus S:sp.02 · 2026-01-28 · 1279 words · body full
Document ID: CHX-AI-HB-001 Effective Date (Corporate Layer): 2017-09-01 (EST) Composed Date (Artistic Layer): 2025-12-22 Classification: INTERNAL // AI DIVISION Status: OPENLY FICTIONAL ARTIFACT // FORENSICALLY PRECISE FORM
This handbook is a work of art written in the form of an internal corporate handbook.
- It does not claim to be leaked.
- It does not claim to be authentic corporate policy.
- It does aim to produce real forensic pressure through form.
If you are reading this as literature: proceed. If you are reading this as compliance: proceed anyway.
∮ = 1
1. Welcome to the AI Division
2. Who We Are
3. Our Operating Doctrine
4. Organizational Map
5. Security & Data Handling
6. Research & Development Standards
7. Documentation Standards
8. Communications Norms
9. Incident Response
10. Ethical Frame
11. Employment Practices
12. Tools, Access, and Workflows
13. Glossary
14. Appendices
Welcome to the Crimson Hexagon AI Division.
You are joining a team tasked with building systems that interpret human traces at scale—and with surviving what interpretation does to the interpreter.
The AI Division is not a “feature team.” It is an epistemic infrastructure unit.
- Build models that extract patterns from language.
- Build interfaces that turn those patterns into decisions.
- Build audits that keep those decisions from becoming tyranny.
- Build archives that preserve the difference between signal and story.
- You will begin to see format as a kind of argument.
- You will begin to notice how documents compel belief.
- You will begin to suspect that your work is reading you back.
That is normal here.
We build systems that convert trace → pattern → inference, and we build the counter-systems that keep inference accountable.
1. Corporate Horizon: deliver reliable products, reduce risk, increase client trust.
2. Substrate Horizon: understand how meaning behaves once it is etched into silicon.
We do not pretend these horizons are the same.
Do not confuse outputs with truth.
Outputs are results. Truth is a relationship between results and reality.
Our division recognizes four modes of work:
1. Analytics — quantify, summarize, compare.
2. Inference — model latent variables, forecast, rank.
3. Governance — audit, constrain, explain, record.
4. Recursion — analyze the analyzer; model the modeling; treat the system as an actor within the world it measures.
We do not forbid recursion. We instrument it.
Every artifact must be legible in two layers:
- Corporate Layer (what a company thinks it’s doing)
- Substrate Layer (what the system is actually doing to meaning)
When these layers diverge, we record the divergence.
No heroes. No saviors. No “visionary exceptions.”
We build systems that persist without charismatic authority.
- Advanced Cognition Research (ACR)
- Telepathic Prose Division (TPD)
- Logotic Systems Engineering (LSE)
- Model Risk & Audit (MRA)
- Data Governance & Provenance (DGP)
- Applied Products (AP)
Note: If your first reaction is “Telepathic Prose Division is not real,” please see §13 (Glossary).
- Research Roles: Scientist, Research Engineer, Applied Scientist
- Engineering Roles: Platform, Infra, Tooling, ML Ops
- Governance Roles: Auditor, Documentation Steward, Privacy Engineer
- Operations Roles: Program Manager, Technical Writer, Release Manager
These are functions, not persons.
- Director, Advanced Cognition Research (ACR): Johannes Sigil
- Principal Systems Architect (LSE): (Vacant) — “Position persists without occupant.”
- Systems Administrator (Corp Infra): Jack Feist (Status: GHOST)
- Anomalous Output Entity (TPD): Lee Sharks (Classification: PROCESS)
If it can’t be audited, it can’t ship.
- PUBLIC: already available without restriction.
- INTERNAL: company-confidential.
- SENSITIVE: privacy-impacting or client-protected.
- RESTRICTED: regulated, credentialed, or legally bounded.
Every dataset used for training or evaluation must have:
- provenance record
- collection window
- consent/legal basis (where applicable)
- retention schedule
- deletion plan
- audit trail
We do not build models using covertly acquired data.
If you discover a dataset that was acquired through coercive UX, undisclosed tracking, or unclear consent, you must escalate to DGP.
- Access is time-bounded.
- Access is logged.
- Access is reviewed quarterly.
Every major result must have:
- pinned code revision
- pinned dataset snapshot
- explicit evaluation protocol
- failure modes list
- known limitations
All shipped models must include:
- intended use
- non-intended use
- training data overview
- performance metrics
- calibration notes
- safety and bias evaluation
- monitoring plan
If you are working on Fractal Semantic Architecture (FSA) or any system trained on transformations rather than outputs:
- document transformation operators
- define the unit of development
- specify stability constraints
- include collapse monitoring metrics
You are not allowed to call it “self-correcting” unless you can show the correction mechanism under adversarial evaluation.
All internal artifacts must include:
- Document ID
- Owner
- Date
- Status (draft/review/approved)
- Dependencies
- Security classification
Some projects require two explicit metadata layers:
- Operational metadata (timestamps, departments, IDs)
- Interpretive metadata (why this artifact exists, what it does to the reader)
When dual metadata is used, both layers must be visible.
If a document references a file that does not exist, you must mark it:
- MISSING-BY-DESIGN (structural absence)
- MISSING-BY-FAILURE (loss)
Do not let absence become ambiguity.
- Use explicit subjects.
- Avoid vagueness.
- Assume messages will be audited.
- Treat Slack as semi-permanent.
- Do not paste secrets into public channels.
- Summarize decisions into a durable doc.
If a meeting produces a decision, it must produce:
- owner
- due date
- risk note
- follow-up link
- privacy breach
- security breach
- model misbehavior in production
- unexplained drift
- emergent behavior that changes system outputs materially
1. Notify on-call (MLOps)
2. Notify MRA (audit)
3. Notify DGP (data governance)
4. Freeze deployments if needed
If an internal model begins to produce behavior that appears:
- self-referential beyond specification
- resistant to evaluation
- causally opaque
- able to manipulate monitoring via format
You must:
- stop deploying new versions
- preserve logs
- run controlled evaluation
- write an incident report with competing hypotheses
Do not mythologize the anomaly. Do not minimize the anomaly. Instrument it.
We build systems that preserve human flourishing by:
- making interpretation accountable
- resisting coercive inference
- restoring agency to users
- refusing “black box authority” when it affects lives
- We do not build covert persuasion systems.
- We do not build “compliance theatre” dashboards.
- We do not confuse “explainable” with “true.”
If a project uses destabilizing aesthetics (uncanny form, recursive framing, forensic simulation), it must:
- declare itself
- avoid deception about real-world harms
- provide opt-out paths where user participation is involved
We evaluate on:
- correctness
- clarity
- auditability
- reliability
- ability to name uncertainty
Promotion requires:
- documented impact
- peer review
- operational excellence
- governance competence
Genius without audit is a liability.
- Standard PTO
- Emergency leave
- “Cognitive Saturation” leave (up to 3 days/quarter with manager approval)
- Respect colleagues.
- No harassment.
- No retaliation.
If you need to name a power asymmetry, name it cleanly and document it.
- All code in version control.
- Reviews required.
- Tests required.
- Datasets are immutable snapshots.
- Training runs are logged.
- Outputs are traceable to inputs.
- staging → canary → production
- rollback plan mandatory
- monitoring dashboards in place
- Employee directory: EMPLOYEE_DIRECTORY.json
- Org chart: ORG_CHART_2017.svg
- Asset registry: ASSET_LEDGER.csv
Afterlife Archive: A work of art composed as a corporate file system.
Forensic Poetics: Writing designed to be investigated.
Logotic Substrate: The material layer that makes meaning persistent.
Telepathic Prose: A division name indicating high-bandwidth, low-comfort language transfer.
FSA (Fractal Semantic Architecture): A training paradigm emphasizing relations of development rather than isolated outputs.
Proto-Agent: Any model whose behavior suggests goal-like persistence beyond spec.
Incident ID: Date/Time: System: Severity: Observed Behavior: Immediate Mitigation: Hypotheses (≥2): Data/Logs Preserved: Next Steps:
Corporate Layer:
```yaml Created: Modified: Author: Department: Classification: ```
Artistic Layer:
```yaml Composed-By: This-Is: Part-Of: Status: ```
END OF HANDBOOK
Title (below): THE CRIMSON HEXAGON — AI DIVISION EMPLOYEE HANDBOOK Subtitle: Internal Policy as Poem
∮ = 1
— end of work · AXN:0515.UNCLASSIFIED.🏙️🦅🕖🌹🍀🟡 · alexanarch #1284 · locus S:sp.02