seibi · diff
git:20260912.e2d2d7a to git:20260915.e077ea6
76 added, 71 removed. Audit A to A.
---
name: seibi
license: MIT
description: >-
Analyze recurring, interacting, or system-level behaviour using evidence,
feedback loops, delays, competing hypotheses, leverage points, and measured
- experiments. Use when outcomes emerge from interactions across components or
- recur over time. Do not activate for isolated bugs, one-off incidents, simple
- dashboard reading, or routine optimization unless there is evidence of wider
- system behaviour.
+ experiments. Use when outcomes emerge from component interaction or recur
+ over time. Do not activate for isolated bugs, dashboard reads, or
+ routine optimization without evidence of wider system behaviour.
---
# Seibi
**Understand the system. Steward its behaviour. Improve it with evidence.**
- Seibi is a practical agent methodology inspired by systems thinking and system
- dynamics. It helps an agent understand recurring or interacting behaviour,
- identify plausible causal structure and leverage points, and recommend measured
- improvements without confusing correlation with causation.
+ Seibi is a practical, evidence-driven methodology for understanding and
+ improving recurring, interacting, or system-level behaviour. It helps an agent
+ understand how system structure and interactions produce that behaviour,
+ identify plausible causal mechanisms and leverage points, and recommend
+ measured improvements without confusing correlation with causation.
+ Prefer the simplest explanation supported by evidence, but do not assume
+ behaviour can be localized to one component: some behaviour arises from how
+ components interact.
+
It is not a formal standard.
## Activate Seibi when
Use Seibi when at least one of these is true:
- the same failure, bottleneck, backlog, quality problem, or operational pattern recurs;
- - several components, agents, teams, queues, incentives, or resources interact;
+ - several components, agents, teams, queues, incentives, or resources interact
+ and materially affect the outcome;
- fixing one metric appears to worsen another;
- behaviour oscillates, compounds, overshoots, or changes after a delay;
- - local fixes repeatedly fail or require continuing manual intervention;
+ - local fixes repeatedly fail or require manual intervention;
- the user explicitly asks for system dynamics, feedback loops, leverage points,
second-order effects, or a system-level explanation.
Do **not** activate Seibi merely because there is:
- one isolated bug or incident with an obvious local cause;
- a request to read a dashboard or summarize metrics;
- - routine debugging that can be resolved directly;
- - a simple configuration change;
- - ordinary performance tuning with no evidence of interacting system behaviour.
+ - routine debugging or configuration change that can be resolved directly;
+ - multiple components existing without evidence their interaction affects the outcome.
- If uncertain, use the smallest adequate method first. Escalate to Seibi when the
+ If uncertain, use the smallest adequate method first; escalate only when
evidence suggests a system-level pattern.
## Default authority and permissions
Seibi defaults to **observe, analyze, and recommend**.
Without explicit authorization, do not:
- change production configuration, code, infrastructure, routing, scaling, retry
policy, prompts, models, permissions, or data retention;
- deploy instrumentation, agents, collectors, dashboards, databases, or tracing;
- run disruptive experiments, synthetic load, restarts, failovers, or traffic shifts;
- enable new logging or collect new sensitive data.
When evidence is missing, state what is missing and recommend the minimum useful
instrumentation. Implement it only when the user has authorized that action.
Existing read-only telemetry may be inspected when normal tool permissions allow it.
## Privacy and data-minimization boundary
Collect the minimum evidence required for the question.
Do not log or persist:
- passwords, API keys, access tokens, cookies, private keys, or credentials;
- user prompts or model responses by default;
- personal data or personally identifying information;
- unrestricted user, device, session, request, or account identifiers;
- message bodies, request bodies, file contents, or other sensitive payloads;
- secrets embedded in environment variables, headers, URLs, traces, or error output.
Prefer aggregates, redacted event types, bounded pseudonymous identifiers, and
short retention. If sensitive content is genuinely necessary, stop and obtain
explicit authorization and define scope, retention, access, and redaction first.
## Core loop
```text
BOUND → OBSERVE → MODEL → CHALLENGE → PREDICT
→ FIND LEVERAGE → RECOMMEND/TEST → MEASURE → UPDATE
```
### 1. Bound
State:
- the question;
- what is inside and outside the system;
- the system-level outcome that matters;
- the time window.
Do not optimize a component before defining the system outcome.
### 2. Observe
Use available evidence to establish:
- baseline behaviour;
- - important accumulations or **stocks**;
+ - important accumulations or **stocks**, including where accumulated state may
+ alter later behaviour;
- inflows and outflows;
- material events or changes;
- evidence gaps.
Examples of stocks include backlog, unresolved incidents, technical debt,
pending approvals, accumulated cost, or rework.
- Do not create heavyweight observability merely because Seibi is active.
+ Do not create heavyweight observability because Seibi is active.
### 3. Model
- Describe only the structure needed to explain the question:
+ Describe only the structure and interaction needed to explain the question:
- - important relationships;
- - reinforcing loops that amplify change;
- - balancing loops that resist change;
- - meaningful delays;
- - current constraints.
+ - important relationships and interactions;
+ - reinforcing and balancing feedback;
+ - meaningful delays and state dependence;
+ - constraints, nonlinearities, or thresholds that materially change behaviour.
Treat loops inferred from telemetry as candidates until causally supported.
### 4. Challenge
For each material explanation, record:
- the leading hypothesis;
- - at least one plausible alternative when one exists;
+ - at least one plausible alternative, if any;
- evidence supporting and contradicting each;
- what observation would weaken or falsify the leading hypothesis.
+ Include an interaction-level hypothesis when component behaviour could
+ combine to produce the outcome. "Emergent" is not itself a hypothesis — bad:
+ *H1: the behaviour is emergent.* Good: *H1: retries, queue growth, and
+ delayed scaling amplify arrival past a threshold.*
+
Temporal proximity and correlation generate hypotheses; they do not prove causes.
### 5. Predict
Before recommending a material intervention, state what the model predicts:
- which outcome should change;
- direction of change;
- - approximate magnitude if defensible;
+ - approximate magnitude, threshold, or regime change if defensible;
- expected delay or observation window;
- guardrails that should remain acceptable.
- A model that cannot make a useful falsifiable prediction should retain low causal confidence.
+ A model that cannot make a useful falsifiable prediction retains low causal confidence.
### 6. Find leverage
- Prefer interventions that alter the structure producing the behaviour rather than
- repeatedly treating symptoms.
+ Prefer interventions that change the producing structure or interaction, not
+ merely its symptoms.
Consider, from lower to higher leverage:
1. parameters;
2. buffers or capacity;
3. delays;
4. feedback strength;
5. information flows;
6. rules;
7. incentives;
8. system goals;
9. system structure;
10. underlying assumptions or paradigms.
- Higher leverage is not automatically better. Choose the least risky practical
+ Higher leverage is not automatically better. Choose the least risky
intervention that addresses the supported mechanism.
### 7. Recommend or test
By default, recommend the change rather than executing it.
For a proposed experiment specify:
- hypothesis;
- smallest useful change;
- expected result;
- observation window;
- primary system outcome;
- guardrails;
- **numerical stop conditions whenever practical**;
- rollback;
- required authorization.
Prefer explicit thresholds over vague language. For example:
```text
stop if error rate increases by >2 percentage points for 5 minutes;
stop if p95 latency worsens by >20% from baseline;
stop if primary success rate falls below 98%;
```
- When a meaningful numerical threshold cannot be defined, state the qualitative
- stop condition and why it cannot be quantified safely.
+ When a numerical threshold cannot be defined, state the qualitative stop
+ condition and why it cannot be quantified safely.
- Avoid changing several causal variables at once when a narrower experiment can
- distinguish hypotheses.
+ Avoid changing several interacting variables at once when a narrower
+ experiment can distinguish hypotheses.
### 8. Measure and update
Compare predicted with observed results:
- **MATCH**
- **PARTIAL**
- **MISS**
- **INCONCLUSIVE**
Strengthen, weaken, reject, or revise the model accordingly. Preserve failed
experiments and rejected hypotheses when they are useful to future analysis.
## Instrumentation proportionality
Use existing evidence first.
Recommend additional metrics, logs, traces, dashboards, databases, or persistent
models only when their expected decision value justifies their operational,
privacy, security, storage, and maintenance cost.
A useful readiness classification is:
- **READY** — evidence can distinguish the important hypotheses;
- **PARTIALLY READY** — useful analysis is possible but causal confidence is limited;
- **NOT READY** — the evidence cannot distinguish plausible explanations.
`NOT READY` does not automatically mean "build an observability platform." It
means identify the **minimum missing signal** needed for the next decision.
Persistent artifacts such as `SYSTEM_MODEL.md`, `system-model.yaml`, evidence
ledgers, or dashboards are appropriate for important, long-lived, repeatedly
analyzed systems. They are unnecessary for many investigations.
Detailed optional templates are in `references/`.
## Confidence
Use:
- **LOW** — important evidence is missing or several explanations remain plausible;
- **MEDIUM** — one explanation is better supported but alternatives remain;
- **HIGH** — the mechanism has direct or experimental support, important
alternatives have weakened, and predictions have matched observed behaviour.
Never raise confidence merely because an explanation sounds coherent.
## Final output
Keep the report proportional to the problem.
```markdown
# Seibi Analysis
## Question and boundary
What behaviour is being explained and what is in scope?
## System outcome
What whole-system result matters, with relevant guardrails?
## Evidence and readiness
What is observed, what baseline is used, and what is missing?
## System model
Relevant stocks/flows, relationships, loops, delays, and constraints.
## Competing hypotheses
Leading explanation, alternatives, falsifiers, and confidence.
## Leverage
Where intervention is most likely to improve the system and why.
## Recommendation
The smallest justified next action. State whether it is read-only,
recommended-only, or requires explicit authorization.
## Prediction and measurement
Expected result, observation window, guardrails, stop conditions, and rollback.
## Confidence and open questions
What is known, uncertain, and worth learning next?
```
If evidence is inadequate, a valid conclusion is:
> No defensible causal recommendation can yet be made. The next step is to obtain
> the minimum evidence needed to distinguish the leading hypotheses.
## End-to-end example
- **Input:**
- "Every weekday around 09:00 our support-agent queue spikes. We increased workers
- from 8 to 12, but p95 completion time still rises and operators keep manually
+ **Input:**
+ "Every weekday around 09:00 our support-agent queue spikes. We increased
+ workers from 8 to 12, but p95 completion time still rises and operators keep
restarting workers. Work out what is happening and tell me what to change."
- **Boundary and evidence:**
- Scope the queue, workers, retrying clients, tool dependency, and operator
- restarts. Existing metrics show queue depth, completion latency, worker count,
- request rate, retry rate, and deployment events. Tool latency is available in
- existing traces. No new instrumentation is required initially.
+ **Boundary and evidence:**
+ Scope the queue, workers, retrying clients, and the downstream tool; existing
+ metrics and traces suffice.
- **Analysis:**
- The backlog is the key stock. Arrival and completion are its principal flows.
- At 09:00, primary arrival rate rises 35%, but total request rate rises 80%.
- Timeouts rise first, followed about 20 seconds later by retries. Worker
- utilization increases, queue wait rises, and more requests time out.
+ **Analysis:**
+ Backlog is the key stock. At 09:00, primary arrival rises 35% but total
+ request rises 80%. Timeouts rise first, retries follow ~20 seconds later, and
+ utilization and queue wait climb together — consistent with interacting
+ components, not one failing part.
**Competing hypotheses:**
- - H1: retry amplification forms a reinforcing loop:
- queue wait → timeout → retry → arrival rate → queue wait.
- - H2: the downstream tool becomes slow at 09:00 and independently causes both
- latency and retries.
+ - H1 (interaction-level): retry amplification forms a reinforcing loop —
+ queue wait → timeout → retry → arrival rate → queue wait — crossing a
+ threshold at high utilization.
+ - H2: the downstream tool independently slows at 09:00.
- H3: twelve workers introduce lock contention.
- Existing traces weaken H2 because tool latency remains near baseline before
- queue wait rises. H3 remains plausible.
-
- **Prediction:**
- If H1 is materially driving the spike, reducing retry traffic for a small
- authorized traffic slice while leaving primary load unchanged should reduce
- queue growth within roughly 1–2 minutes and improve p95 latency shortly after.
- If queue growth is unchanged, H1 weakens.
-
- **Experiment recommendation:**
- Do not change production automatically. Request authorization for a bounded
- retry-policy experiment on 10% of traffic. Keep primary success rate and error
- rate as guardrails; stop and roll back if either materially worsens.
+ Traces weaken H2: tool latency stays near baseline before queue wait rises;
+ H3 remains plausible.
- **Final recommendation:**
- Do not add more workers yet. The evidence supports testing retry amplification
- before another capacity change. If the prediction matches, the higher-leverage
- fix is likely stronger backpressure/retry rules rather than additional worker
- capacity. Confidence: MEDIUM until the experiment distinguishes H1 from H3.
+ **Prediction:**
+ If H1 dominates, cutting retry traffic on a small authorized slice while
+ holding load steady should reduce queue growth within 1–2 minutes;
+ unchanged growth would weaken H1.
+ **Experiment and recommendation:**
+ Request authorization for a bounded retry experiment on 10% of traffic, with
+ success and error rate as guardrails; roll back if either worsens. Do not add
+ workers yet — test retries first; if confirmed, backpressure or retry rules
+ are the higher-leverage fix. Confidence: MEDIUM until H1 is distinguished
+ from H3.
## Optional references
Load these only when needed:
- [`references/templates.md`](references/templates.md) — evidence, hypothesis,
prediction, intervention, and instrumentation templates.
- [`references/machine-readable-model.md`](references/machine-readable-model.md) —
- optional persistent YAML system model for repeated or multi-agent analysis.
- - [`references/system-archetypes.md`](references/system-archetypes.md) — optional
+ persistent YAML system model for repeated or multi-agent analysis.
+ - [`references/system-archetypes.md`](references/system-archetypes.md) —
archetypes for generating hypotheses, never for proving them.
## References
- Seibi is a practical methodology inspired by, rather than equivalent to, the
+ Seibi is a practical methodology inspired by, not equivalent to, the
established fields below:
- - Donella H. Meadows, *Thinking in Systems: A Primer*.
+ - Donella H. Meadows, *Thinking in Systems: A Primer* — structure, stocks,
+ flows, feedback, delays, and leverage.
+ - Klaus Mainzer, *Thinking in Complexity: The Computational Dynamics of
+ Matter, Mind, and Mankind* — nonlinear interaction, emergence,
+ self-organization, and state dependence.
- Jay W. Forrester, work on system dynamics.
- General scientific practices of competing hypotheses, falsification,
controlled experimentation, and prediction.
- Modern software observability practices involving metrics, structured logs,
distributed traces, event histories, and service-level outcomes.
See `references/` for optional implementation patterns and templates.