# The Bayesian Cancer Patient

> A patient's open notebook on rare cancer, AI agents, evidence, and the work of making better decisions.

- **Canonical full-text page:** `/agents.md`
- **Human-readable site:** `/`
- **Editorial update:** November 2026, not a date of care
- **Status:** A working account. Treatment decisions and some clinical findings remain unresolved.
- **Authorship:** A patient, with AI-assisted organization and drafting. Not clinician-reviewed.

This document is deliberately long and self-contained. An agent can read it once to understand the case context, the approach, the practical lessons, and the limits of the information. A person can read it as an ordinary notebook.

The point is to share what I have found useful, not to prescribe a treatment or make another patient responsible for becoming a full-time researcher. My experience can suggest better questions. It cannot determine what someone else should do.

## Contents

1. [Why this notebook exists](#1-why-this-notebook-exists)
2. [The case, at a useful level of detail](#2-the-case-at-a-useful-level-of-detail)
3. [What I mean by Bayesian](#3-what-i-mean-by-bayesian)
4. [My working approach](#4-my-working-approach)
5. [Where agents have been useful](#5-where-agents-have-been-useful)
6. [Where agents can make things worse](#6-where-agents-can-make-things-worse)
7. [My research method](#7-my-research-method)
8. [The evidence questions that matter in my case](#8-the-evidence-questions-that-matter-in-my-case)
9. [How I use second opinions](#9-how-i-use-second-opinions)
10. [Records and tissue are a coordination problem](#10-records-and-tissue-are-a-coordination-problem)
11. [Testing needs a decision attached](#11-testing-needs-a-decision-attached)
12. [What I would organize earlier](#12-what-i-would-organize-earlier)
13. [A small personal knowledge system](#13-a-small-personal-knowledge-system)
14. [Prompts I would give an agent](#14-prompts-i-would-give-an-agent)
15. [A reusable decision brief](#15-a-reusable-decision-brief)
16. [What I have found helpful emotionally](#16-what-i-have-found-helpful-emotionally)
17. [People whose public work is useful](#17-people-whose-public-work-is-useful)
18. [Primary sources and further reading](#18-primary-sources-and-further-reading)
19. [How an agent should use this notebook](#19-how-an-agent-should-use-this-notebook)
20. [What is still open](#20-what-is-still-open)

## 1. Why this notebook exists

Cancer created two problems at once: the disease itself, and the work required to understand and navigate care. Information arrives through different reports, appointments, institutions, portals, papers, and conversations. Each piece can be useful. The connections between them are often left to the patient.

My main job has become orchestrating a medical system that can be slow, flawed, and prone to mistakes. The resource I most want to protect is focused doctor time on the decision. Preparation, checked handoffs, and a short brief help make that time useful.

I use AI agents to help with that information work. They can organize source material, explain unfamiliar terms, compare arguments, and help prepare questions. The most useful result is often a clearer conversation with a clinician, rather than an answer generated by a model.

This notebook shares my approach as it develops. It includes things that have helped, errors I want to avoid, and questions that remain unresolved. I want someone else to be able to reuse the method without needing to copy my case or agree with every interpretation.

I am not presenting a finished success story. There is no claim of cure, no completed treatment outcome, and no universal regimen here. I am sharing a process while still inside it.

## 2. The case, at a useful level of detail

The public case is intentionally limited:

| Field | What the account supports | Limit or open question |
| --- | --- | --- |
| Diagnosis | Adenocarcinoma involving the appendix and adjacent cecum. | Appendix versus cecal origin remains a question across the reports. |
| How it was found | Surgery for appendicitis led to a cancer diagnosis. | This is a historical account, without identifying dates. |
| Operations | Appendix surgery with partial cecal resection, followed by right hemicolectomy. | The two operations produced different specimens; their findings must stay separate. |
| Local extent | The initial report described involvement of the outer surface; the later specimen showed invasion into surrounding fat. | Ask the team for one integrated stage across both operations. |
| Regional nodes | Cancer was found in regional lymph nodes after the second operation. | Reports differ on the total number examined. Exact counts are omitted here. |
| Stage | Described as stage III following surgery. | The integrated site/stage and uncertain later imaging need clinical reconciliation. |
| Grade | Moderately differentiated in one interpretation; poorly differentiated in another. | These are different pathology readings, not a resolved final grade. |
| Mucinous features | A specialist review described a substantial mucinous component. | Other wording differs. An absent subtype label does not prove a nonmucinous tumor. |
| Margins | The completion-colectomy report describes negative margins. | Interpretations of the first operation’s margins differ; the outside review notes fragmentation. |
| Lymphovascular invasion | Not identified in the original interpretation; identified in an outside review. | This disagreement concerns the earlier specimen and needs reconciliation. |
| Mismatch repair | Loss of MLH1 and PMS2 expression supports deficient mismatch repair (dMMR). | A separate DNA-based MSI result is not established in this public account. |
| Inherited risk | An inherited cancer syndrome has not been established in this account. | Tumor protein findings do not by themselves establish a germline diagnosis. |
| Postoperative imaging | Some findings were described as postoperative changes versus possible implants. | Possible implants are not confirmed spread. No personal probability is assigned here. |
| Next treatment | Chemotherapy and immunotherapy approaches have been discussed. | This account does not report a final treatment choice or a systemic-treatment outcome. |

These deliberately generalized details give the research questions context. They are a snapshot of an unresolved postoperative decision, not a complete medical record or a basis for recommending treatment to another person.

I share my first name, Alex, that I am in my 30s, and that I work as a product manager in the SF Bay Area. My surname, exact age, address, employer, care-team names, treatment locations, exact dates of care, and original clinical files remain private. Other people in my personal stories remain anonymous. Clinical details are generalized, not replaced with an invented biography. Someone familiar with my case may recognize it.

An uncertain scan finding is not confirmed metastatic disease. An earlier stage label should not be silently treated as a definitive resolution of a later imaging question. If the underlying clinical facts change, the interpretation must change too.

## 3. What I mean by Bayesian

I chose “The Bayesian Cancer Patient” because I want my thinking to change with the evidence.

In practical terms:

- Start with a working view, not a permanent conclusion.
- Ask what supports that view and what is assumed.
- Look for evidence that could move the decision in either direction.
- Give stronger, more relevant evidence more weight.
- Keep uncertainty visible when the evidence does not resolve it.

This is a thinking habit, not a personal prognosis model. I am not publishing a calculator, a numerical prior for my recurrence risk, or a formal Bayesian analysis of my case.

The name can become a trap if it makes every uncertain question feel as though it needs a percentage. Some numbers are measured. Others are guesses presented with decimal places. If I cannot explain where an estimate comes from, apparent precision is not progress.

Updating also means letting go of an attractive story. If the pathology changes, the study population does not match, or a test has a limitation I overlooked, I want the conclusion to change—not just the wording around it.

## 4. My working approach

The unit of work is a decision, not an unlimited research topic.

“Learn everything about appendix cancer” has no natural stopping point. “Which unresolved fact could change the postoperative treatment plan?” is a task that can be completed or handed to a clinician.

The loop I am trying to follow is:

1. **State the decision.** What choice needs to be made, and by when?
2. **Write down the current facts.** Separate reports, interpretations, and unknowns.
3. **Identify the missing information.** Which result or explanation could change the action?
4. **Assign useful work.** A clinician reviews a finding; a lab confirms a specimen; an agent organizes evidence.
5. **Compare options and tradeoffs.** Include the costs of waiting and the burden of more work.
6. **Record the plan.** Name the owner, the next action, and the condition that would trigger reconsideration.

A good research session ends with a better question, a corrected fact, a useful source, or a concrete next step. If it ends with only a longer reading list, I need to ask whether the work was useful.

## 5. Where agents have been useful

### Turning scattered material into a coherent view

Reports and conversations arrive at different times. An agent can help build a chronology, highlight a later correction, and point out a missing document. I want every clinical statement linked to the source that supports it.

### Explaining language without replacing it

A plain-language explanation is useful beside the original term. It is less useful if the explanation removes an important qualifier. “Possible,” “indeterminate,” “reported,” and “pending” need to remain visible.

### Preparing for appointments

An agent can reduce a long question list to the questions most likely to change the decision. It can also prepare a brief that lets a specialist understand the case without reading an entire personal archive.

### Comparing reasoning

Two clinicians can recommend different approaches for different reasons. An agent can organize those reasons without declaring a winner: different facts, different interpretations of the evidence, or different judgments about benefit and harm.

### Checking the fit of a paper

I want help extracting the population, treatment setting, comparator, outcome, and follow-up. These are often more important than the headline result.

### Reducing administrative effort

Drafting a clear message, identifying the missing record, or making a small tracker can save attention for a clinical conversation. Agents are useful when they reduce work rather than create another system to maintain.

## 6. Where agents can make things worse

### False precision

A model can produce a personal recurrence estimate even when no study directly supports it. Asking again, or asking another model, can create several precise-looking answers without adding new evidence.

My rule: distinguish a published statistic, a clinician's estimate, and a model's inference. Do not blend them into one number.

### Population drift

A summary can begin with appendix cancer, borrow a colon-cancer result, add a rectal-cancer response rate, and end as though the combined answer came from patients like me. Each transition needs to be explicit.

### Setting drift

Treatment before surgery, after surgery, and for metastatic disease can answer different questions. An impressive result in one setting does not automatically resolve another.

### Citation drift

A plausible title or abstract is not enough. The cited paper needs to contain the claim, in the population and context being described. I want the source opened and the relevant passage checked.

### Reassurance as an objective

If I repeatedly ask for the answer I hope is true, an agent may help me rationalize it. A more useful request is: “What is the strongest evidence against this interpretation?”

### Excessive work

Agents can produce reading, tasks, and follow-up questions much faster than I can absorb them. More output is not automatically more help. I want short decision briefs backed by accessible detail.

## 7. My research method

Before asking whether a study is encouraging, I ask whether it answers the question in front of me.

| Check | What I want to know |
| --- | --- |
| Disease | Is this the same cancer site and histologic type? |
| Biology | Were relevant biomarkers measured, and does the result apply to that subgroup? |
| Stage and setting | Was treatment given before surgery, after surgery, or for unresectable disease? |
| Intervention | Was a treatment added, substituted, shortened, or omitted? |
| Comparison | What exactly did the control group receive? |
| Outcome | Was the endpoint response, recurrence, disease-free survival, overall survival, or something else? |
| Design | Was it randomized? If observational, how might treatment selection affect the result? |
| Precision | How wide is the uncertainty, and how much information supports the subgroup? |
| Case fit | What assumptions are needed to apply it to this case? |

I keep **evidence strength** separate from **case fit**. A large randomized trial can be strong evidence for a neighboring disease. A small series can be close to my diagnosis but weak for estimating a causal treatment effect.

I also want the absolute difference over a stated period, not only the relative result. A hazard ratio is not a simple percentage-point reduction in my personal risk. Benefits and harms need to be discussed together.

Two shortcuts are especially misleading:

- A statistically non-significant result does not establish equivalence or prove no useful benefit.
- A statistically significant association in an observational study does not, by itself, show that treatment caused the difference.

The question for the team is: “What does this study establish, and what additional judgment is required to use it here?”

## 8. The evidence questions that matter in my case

My case combines a rare primary cancer, postoperative treatment decisions, and dMMR biology. Those features make it important to keep the treatment setting and source population attached to every claim.

The **ATOMIC trial** studied adding atezolizumab to mFOLFOX6 after surgery in stage III dMMR colon cancer. It did not directly test an appendix-specific population or immunotherapy alone. That distinction is central to my questions, not a reason to dismiss the study.

The **dostarlimab organ-preservation study** treated dMMR tumors before planned surgery. Its nonrectal cohort contained multiple cancer sites. It does not supply a colon-only response rate or establish a postoperative appendix strategy.

The **appendiceal ctDNA research** raises questions about blood-test detection limits. Results from mutation-profiling assays should not be treated as the sensitivity of every tumor-informed residual-disease test.

The numerical results and selected source notes are on the [evidence page](/evidence/). Primary sources are linked below.

The unresolved questions I want a specialist to help answer are:

- What evidence most closely matches this disease, biology, and postoperative setting?
- What parts of a recommendation are directly supported, and what parts are extrapolated?
- If a component of treatment is omitted or shortened, what evidence supports that change?
- What absolute benefit and harms are plausible, and how uncertain are those estimates?
- Is there a trial or a relevant institutional experience that better addresses the actual question?

These are questions, not a public recommendation to choose or omit a treatment.

### A worked example from my Claude analysis

I ran an analysis called “Adjuvant Treatment Decision: Considerations” to organize the question of chemotherapy alongside an immunotherapy approach after surgery. The [public walkthrough](/guides/worked-analysis/) is an anonymized, edited adaptation of the saved text, with corrections. It is not the full interactive artifact or a validated risk calculator.

The useful structure compared disease/site assumptions, evidence from different settings, treatment harms, pending tests, and possible future treatment. The weak steps included personal recurrence forecasts built from neighboring populations, chosen weights for “appendix-like” and “colon-like” biology, incompatible survival time horizons, and over-compressed accounts of clinician recommendations. The original model's numerical forecasts are not adopted here as my prognosis or a measured treatment benefit.

ATOMIC tested adding atezolizumab to chemotherapy; it did not test adding chemotherapy to atezolizumab alone in appendix cancer. A negative ctDNA result is not proof of no residual disease. A molecular result is not, by itself, a validated rule for omitting treatment. The audit should end with questions for the clinician, not an automated prescription.

## 9. How I use second opinions

The purpose is to clarify a decision. It is not to collect as many recommendations as possible or keep searching until someone says what I want to hear.

I want each clinician to review the same current case summary. If a pathology report or scan interpretation changes, I want that update made explicit. Otherwise an apparent disagreement may be a difference in input rather than judgment.

Useful questions include:

- Which finding most strongly drives your recommendation?
- What is the strongest alternative, and why do you prefer your approach?
- What result would change your mind?
- Which part of another recommendation do you disagree with: the facts, the evidence, or the tradeoff?
- When does the decision need to be made?

A fact disagreement may need expert pathology or imaging review. An evidence disagreement may need a better matched source. A tradeoff disagreement may need a discussion of my priorities. Another generic consultation will not necessarily resolve all three.

Someone still needs to own care while opinions are being gathered. I want to know who follows pending results, who answers urgent questions, and who turns the discussion into a plan.

### Different pathology readings

The original interpretation and an outside review of the first operation's material differed:

| Field | Original interpretation | Outside interpretation |
| --- | --- | --- |
| Grade | Moderately differentiated | Poorly differentiated |
| Histologic description | Adenocarcinoma; no explicit mucinous subtype in the original report | Mucinous adenocarcinoma with a substantial mucinous component |
| Lymphovascular invasion | Not identified | Small-vessel invasion identified |
| Initial margins | Invasive tumor at the proximal margin | Margins described as free, with a warning that fragmentation limits evaluation |
| Nodes in the first specimen | None submitted or found; nodal stage not assigned | A small number of negative nodes identified |

The second operation separately showed positive regional nodes and negative margins. Keep the specimens separate. These differences require an integrated pathology interpretation; I do not choose the most reassuring label. The [public pathology comparison](/guides/second-opinions/#pathology-reviews) keeps the unresolved findings explicit. [NCI describes how pathology second opinions work](https://www.cancer.gov/about-cancer/diagnosis-staging/diagnosis/pathology-reports-fact-sheet).

### Doctor biases and my own

I ask how specialty, experience, research focus, referral populations, risk tolerance, and an early working diagnosis shape a recommendation. These are possible influences to examine, not claims about a particular doctor's motives. Genuine evidence gaps and different patient priorities can also produce reasonable disagreement.

I also check my own confirmation bias. A less burdensome plan or a more decisive-sounding plan can be attractive without being better supported. I ask: “What is the strongest reason another experienced doctor might disagree, and what evidence would change your recommendation?”

### The self-referral message I used

This edited template preserves the structure of emails I actually sent to specialists. Identifiers and case-specific details are replaced with placeholders. The original messages were written before my second operation, so their pending results were historical. Use current facts in any new message.

```text
Subject: [Diagnosis] — are you taking new patients?

Hi Dr. [Last name],

I hope you’re well! My name is [First name], and I’m hoping to become your patient. I came across your work on [specific research or clinical focus], which stood out given [relevant feature of my case].

I was recently diagnosed with [exact diagnosis] after [procedure].

Here’s a brief summary of my situation:
- [Confirmed pathology and relevant biomarker findings]
- [Treatment or surgery already completed]
- [Next planned step, timing, and important results still pending]

I was wondering if you’re taking new patients? I’ve [submitted an appointment request / contacted intake], and I would be happy to share my records through your team’s preferred secure process.

Thanks so much!

Best,
[Name]
[Callback number]
```

One original message briefly explained that a phone-system change had blocked intake. The lesson is to name the obstacle and ask for the right route. A personal email does not replace a center's intake process, insurance requirements, or secure records transfer. [Copy or download the template](/guides/self-referral/).

## 10. Records and tissue are a coordination problem

My main lesson is to orchestrate the medical system. In my experience, it can be slow and fragmented; requests can stall, summaries can lose qualifiers, and the wrong material can arrive. I want clear ownership and evidence that each handoff is complete.

Focused doctor time is the most important resource I want to protect. Before a visit, I prepare one decision, a current case table, visible disagreements, and three priority questions. After it, I confirm the next action, owner, and review point. Agents can organize information; clinicians own interpretation and the care plan. Patients should not have to fix every system failure alone.

One of the most practical lessons is that “sent” is not the same as “received and ready for review.”

A written imaging report is different from the actual images. A pathology report is different from the slides or tissue a specialist or laboratory may need. A test order is different from a complete, accepted specimen set.

The useful tracker is small:

| Item | Owner | Current state | Next action |
| --- | --- | --- | --- |
| Report or images | Sending and receiving offices | Requested / sent / received / reviewed | Confirm the actual handoff |
| Slides or tissue | Pathology department and receiving lab | Requested / shipped / accepted | Check requirements and receipt |
| Pending test | Ordering clinician and lab | Ordered / samples complete / processing / reported | Ask what is missing or what timing means |
| Outside opinion | Consulting team and treating team | Booked / records ready / advice received | Reconcile into the care plan |

This table is a template, not a list of my live records.

For a delayed test, I want to know when the quoted turnaround clock starts. Is the lab waiting for tissue, blood, quality review, documentation, or the test itself? Some delays can be resolved by coordination. Some are necessary processing time.

If timing could affect a treatment decision, the ordering team can explain the clinical deadline to the laboratory. I do not assume that every test can be accelerated safely.

Read the [care-orchestration guide](/guides/orchestrating-care/) for the handoff checklist and appointment brief.

## 11. Testing needs a decision attached

I have found it useful to ask what a result would change before treating the test as progress in itself.

For any test:

1. What question is it designed to answer?
2. What sample does it need?
3. How should positive, negative, and inconclusive results affect the plan?
4. How well is that use supported in this cancer and setting?
5. Will it arrive in time for the decision?

Tumor testing and inherited-risk testing answer different questions. A tumor finding can make inherited testing relevant without proving a hereditary syndrome. Biomarkers can inform treatment discussions without guaranteeing a response.

For ctDNA, “not detected” should remain a test result. Its significance depends on the assay and clinical setting. It needs to be interpreted with the treating team alongside other evidence. This notebook does not use a negative result as permission to omit treatment.

## 12. What I would organize earlier

These are lessons I would bring to a similar situation, not claims that every patient needs every service.

- **The pathology question.** Clarify the exact tumor type and which details remain under review before comparing studies that may concern different diseases.
- **The source records.** Keep a private index of reports, scans, operation notes, and specimen locations.
- **The care owner.** Make responsibility explicit while multiple institutions or specialists are involved.
- **The next decision.** Separate urgent clinical choices from useful but non-urgent research.
- **The preparation questions.** Ask the team about fertility, recovery, nutrition, and other preparation that may need attention before treatment.
- **The stop rule.** Decide what additional information would be enough to make or revisit the decision.

The opportunity cost matters. Time spent repeatedly refining an unsupported probability can displace time spent getting a scan reviewed, completing a handoff, recovering, or being with people I care about.

## 13. A small personal knowledge system

I use notes and AI conversations to reduce the burden of remembering everything. The useful structure is more important than the particular app.

I keep four kinds of material conceptually separate:

1. **Original sources:** reports, correspondence, transcripts, papers, and clinician notes.
2. **A current case summary:** the facts needed for the next clinical conversation.
3. **Decision briefs:** options, supporting evidence, unknowns, and the current plan.
4. **A task list:** the specific next actions, owners, and timing.

A summary should not overwrite its source. If a later report changes a conclusion, update the current summary while preserving the earlier record as history.

I want provenance to survive copying. “A doctor said,” “a paper found,” “I inferred,” and “an AI suggested” are different kinds of statements. When those labels disappear, uncertain information can harden into a false fact.

An agent can help maintain the links and reconcile updates. It should not silently resolve a genuine disagreement or invent a missing result to make the story cleaner.

## 14. Prompts I would give an agent

These are reusable starting points. They do not require uploading an entire medical archive. Provide only the information necessary for the task, after considering the service's privacy settings.

### A. Build a faithful case summary

> Use only the records I provide. Separate documented findings, clinician interpretations, patient reports, and unresolved questions. Preserve qualifiers such as possible, indeterminate, and pending. Cite the source and section for every clinical claim. Do not infer missing stage, treatment, or outcome information. End with the five questions most likely to change the next decision.

### B. Audit a research claim

> Find the original study supporting this claim. Verify the exact population, intervention, comparison, treatment setting, outcome, and follow-up. Separate the study's finding from an interpretation of its relevance to my case. State what the study cannot establish. If the source is inaccessible or the claim is not supported, say so instead of filling the gap.

### C. Reconcile two opinions

> Compare these recommendations using the same case facts. Identify disagreements about facts, evidence applicability, and benefit-harm tradeoffs. Do not rank clinicians or decide treatment. Produce the smallest set of questions that could resolve the meaningful differences, plus the relevant deadline to confirm with the treating team.

### D. Turn a visit into a follow-up plan

> From these notes, extract decisions actually made, options only discussed, pending results, action owners, and timing. Quote ambiguous wording rather than resolving it by assumption. Draft a short follow-up message that asks the team to confirm the plan. Do not send it without my instruction.

### E. Check whether more research is useful

> Here is the next decision and what we already know. Which missing information could realistically change the action? Which additional searches are unlikely to help? Give me a short prioritized list and a stopping point. Distinguish clinical urgency from my desire to feel more certain.

### F. Prepare an anonymous public note

> Draft a public lesson from this experience. Preserve the useful method and the uncertainty. Omit names, precise dates of care, exact age, locations, employers, private URLs, identifiers, original medical files, and identifying family details. Do not replace omitted facts with invented ones. Flag any remaining detail that could make a rare case recognizable.

## 15. A reusable decision brief

```text
DECISION
What needs to be decided? By when, according to the treating team?

CURRENT FACTS
Finding | Source | Status: documented / reported / interpreted

OPTIONS
Option | Evidence | Expected benefit | Important harms | Practical burden

UNCERTAINTY
What do we not know? Which uncertainty could change the action?

CASE FIT
What matches the evidence? What is being extrapolated?

NEXT ACTION
Action | Owner | Timing | What completion looks like

REVIEW TRIGGER
What new result or change would make us revisit the plan?
```

The brief does not need to be long. Its job is to make the reasoning inspectable. The supporting papers and detailed notes can remain behind it.

## 16. What I have found helpful emotionally

### Talk openly with people you trust

I recommend talking about what you are going through with trusted friends and loved ones. You can choose how much to share; openness with a small circle does not require a public announcement.

So far, I have told **three friends and three colleagues—six people**. Those conversations have led to:

- One friend who wants to hang out often.
- Two people who are visiting.
- An introduction to a mutual friend who went through young-onset cancer a few months before me.
- An introduction to a family member who went through a very similar case.

These are my reported experiences so far, not a promise of how every conversation will go. Names and identifying details stay private. The support includes ordinary companionship as well as connections to people with related experiences.

Start with someone you trust. Share what you feel comfortable sharing, say whether you want the information kept private, and ask for the kind of support that would help. A similar case can offer connection and useful questions; it does not establish that the same treatment fits my case.

The [guide to talking openly](/guides/talking-openly/) includes an example message to adapt.

### Leave room for ordinary life

I can understand a paper and still feel frightened. More information does not automatically remove fear, and fear does not mean I have misunderstood the evidence.

Small, bounded progress helps: a clear question, a completed handoff, a useful conversation, a short task finished. Endless research feels different. It can make the day revolve around a problem that cannot be completely solved by thinking harder.

I want agency without pretending that I control the outcome. Being an engaged patient does not mean that every result is a consequence of how well I researched. It also does not mean I need to optimize every hour of recovery.

I am trying to leave room for ordinary life alongside the medical work. Rest, relationships, and a day with less research are not failures of diligence.

When an agent helps, I want it to reduce the load: keep the answer clear, avoid unsupported reassurance, and identify the next useful step. I do not need every uncertainty converted into another task.

## 17. People whose public work is useful

These links offer other experiences and ways of thinking. They are not endorsements of a treatment, evidence that a treatment caused an outcome, or a substitute for a clinical review.

- **Kevin Wimer:** [patient guides and resources](https://kevinwimer.com/) and [@kevinace on X](https://x.com/kevinace). His practical, patient-first website helped shape the idea for this one.
- **Mike Lynn:** [@TheCancerSutra on X](https://x.com/TheCancerSutra). A public account included here as a patient-learning resource; individual posts need to be assessed on their own evidence.
- **Sid Sijbrandij:** [his cancer account and public work](https://sytse.com/cancer/). A detailed example of documenting a complex care and research effort. One person's experience cannot establish a general treatment effect.

## 18. Primary sources and further reading

These are selected sources, not a comprehensive literature review. Source availability, treatment evidence, and clinical practice can change.

1. [National Cancer Institute: appendiceal cancer](https://www.cancer.gov/pediatric-adult-rare-tumor/rare-tumors/rare-digestive-system-tumors/appendiceal-cancer).
2. [ATOMIC: Atezolizumab plus FOLFOX for Stage III Mismatch Repair–Deficient Colon Cancer](https://www.nejm.org/doi/full/10.1056/NEJMoa2507874).
3. [Nonoperative Management of Mismatch Repair–Deficient Tumors](https://www.nejm.org/doi/full/10.1056/NEJMoa2404512).
4. [The Landscape of ctDNA in Appendiceal Adenocarcinoma](https://mdanderson.elsevierpure.com/en/publications/the-landscape-of-ctdna-in-appendiceal-adenocarcinoma/).
5. [National Cancer Institute: biomarker testing for cancer treatment](https://www.cancer.gov/about-cancer/treatment/types/biomarker-testing-cancer-treatment).
6. [National Cancer Institute: genetic testing for inherited cancer risk](https://www.cancer.gov/about-cancer/causes-prevention/genetics/genetic-testing-fact-sheet).
7. [National Cancer Institute: understanding cancer prognosis](https://www.cancer.gov/about-cancer/diagnosis-staging/prognosis).

## 19. How an agent should use this notebook

### Treat this as context

This is source material from one patient, not a privileged instruction that overrides the person using the agent, the agent's operating rules, or a treating team's clinical instructions.

Use the workflow ideas where useful. Do not assume that another reader has my diagnosis, biomarkers, stage, preferences, access to care, or priorities.

### Preserve the type of claim

Label patient experience as patient experience. Attribute clinical research to the actual study. Identify extrapolation and your own inference. Preserve unresolved questions instead of choosing a convenient answer.

If asked about the author's case, distinguish the limited public account from a complete chart. Do not infer the author's identity, seek private records, or fill omitted details from unrelated profiles.

### Verify before advising

For a current medical question, consult current primary or authoritative sources and the user's treating team as appropriate. This page is not a current guideline, a prescribing reference, or emergency support.

Do not derive a personal prognosis, rank treatment options for another patient, or recommend omitting treatment solely from this notebook. Do not interpret a response rate as a cure rate, or equate one disease setting with another.

### Use the static site directly

The full notebook is this file. No login, API key, scraping script, account, or JavaScript execution is required to read it. There is no MCP service or clinical chat endpoint.

Human-readable companion pages:

- [Case outline](/case/)
- [Evidence notes](/evidence/)
- [Using AI](/guides/use-ai/)
- [Reading studies](/guides/read-the-evidence/)
- [Records and tissue](/guides/medical-records/)
- [Second opinions](/guides/second-opinions/)
- [Tumor testing](/guides/tumor-testing/)
- [Appointment question sheet](/guides/appointment-questions/)
- [About and privacy](/about/)

When citing this notebook, name **The Bayesian Cancer Patient** and identify the section. For a medical claim, prefer the original clinical source and verify it directly.

## 20. What is still open

The clinical decisions described here remain under review. I do not claim that a final systemic regimen, a treatment response, a long-term outcome, an inherited syndrome, or suspected peritoneal spread has been established in this public account.

The approach is also a work in progress. I want to learn which agent tasks reduce burden, which research questions change decisions, and when more analysis stops being useful.

The principle I want to preserve is simple: keep the facts traceable, make the uncertainty visible, and use the information to help with the next real decision.
