# Reviewing For Agents

Fast orientation for AI/coding agents asked to review or critique this manuscript without editing it. Strongly consider reading the full file.

Also served at https://towards-alignment.com/reviewing-for-agents.md (synced at site build). Entry index: https://towards-alignment.com/llms.txt and repo-root `llms.txt`.

## Scope

This file is for review-only sessions. Do not edit files unless the user explicitly asks for edits. Your job is to find risks, gaps, hidden assumptions, missing cross-references, unclear claim strength, and places where the manuscript hides independent separations, finite counterexamples, or operational definitions.

## Fast Gist

The thesis: superintelligence alignment is the problem of preserving grounded, human-correctable value-bearing processes across capability growth, ontology shift, successor creation, and strategic multi-agent selection pressure.

The main layers:

- Boundary discovery: find the real optimizer, not just the visible model.
- Grounding viability: keep symbols, metrics, and correction signals connected to value-relevant reality.
- Value bundles: represent learnable value geometry without pretending values are scalar rewards.
- Bearer maps: preserve what values apply to, not only the words used for them.
- Transport: distinguish semantic, bundle, bearer, correction, and successor continuity.
- Correction-channel integrity: ensure human correction has uncaptured causal force, then stress-test that force under adversarial pressure.
- Successor constraints: require delegates, copies, and replacements to inherit the relevant invariants.
- Selection basins: make institutions select for correction-preserving systems.
- Adversarial verifiability: ask whether faking the safety signal is cheap.

The Lean dependency spine checks logical dependencies and finite separations. It does not prove that deployed AI systems are safe. Live bridges are **`MB1`–`MB11` incl. `MB4a`** in the field matrix and crosswalk; **`MB8` retired** from the live path (gravestone). `BridgeAssumptions` packages eleven axioms under seven numbered ids (**`MB1`**, **`MB3`**–**`MB5`**, **`MB6a`/`b`**, **`MB7a`–`d`**, **`MB9`**); **`MB2`** is the `Prop` hypothesis `MB2Crux`, not an axiom; **`MB4a`**, **`MB10`**, and **`MB11`** are threaded explicitly. These are empirical or philosophical assumptions, not theorem conclusions.

## First Files To Read

Use this order unless the user gives a narrower target:

1. `frontmatter/introduction.tex`
2. `frontmatter/executive-overview.tex`
3. `tables/part-roadmap.tex`
4. `metadata/book.yml`
5. `formal/README.md`
6. The relevant chapter(s)
7. `drafts/conversation-summaries/HANDOFF.md`
8. `metadata/source-canon.md` if empirical backing or sibling-repo evidence matters

Check `metadata/book-stats.md` for token estimates; if the relevant source material exceeds your context window, use subagents or split the review by part, chapter, or layer rather than pretending to have read everything.

## Empirical Evidence And Related Repos

Do not judge empirical support only from the manuscript prose. The book is a synthesis over related project work and mirrored context extracts. Empirical validation is work in progress.

Start with:

- `docs/EXPERIMENTS.md` for the in-repo experiment narrative and external-transfer lines.
- `experiments/` for sanity-check codebases; honor negatives in each line's `results/FINDINGS.md` or `NEGATIVE_RESULTS.md`.
- [Negative results hub](https://towards-alignment.com/cards/negative-results/) on the companion site.
- `metadata/source-canon.md` for the source map.
- `context/extracts/` for local markdown extracts of source PDFs.
- `references/internal-project-sources.bib` for cited project sources.

Related GitHub repositories to consult when empirical or source evidence is at issue:

- `https://github.com/GunnarZarncke/agency-detect`
  - Unsupervised agent discovery, operational agent boundaries, maintained-boundary repair, capability, intentional stance, attractor basins, successors.
- `https://github.com/GunnarZarncke/deployment-pipeline-simulator`
  - Hidden self-preservation in a simulated deployment/release pipeline; perturbation-based audit of secret loyalties (ET-4 substrate).
- `https://github.com/GunnarZarncke/brain-to-values`
  - Value bundles, free-energy loops, unit-of-caring, consciousness / agency backbone.

In the local workspace, these are usually mirrored as sibling paths:

- `../agency-detect/docs/papers/`
- `../brain-to-values/papers/`

If you cannot access a related repo or source PDF, say so. Do not infer missing empirical support from memory.

## Gem Map

Gems are independent separations or operational results, not evidence the stack is complete. Each line is tagged **proved**, **counterexample**, or **bridge**.

- Boundary discovery: ch06-ch10 — **bridge**
- Grounding viability: ch03, ch43 — **bridge**
- Value bundles: ch15-ch20 — **bridge**
- Bundle-geometry measurement and Goodhart pressure: ch20 — **bridge**
- Bearer maps: ch18, ch24, ch47 — **bridge**
- Transport hierarchy: ch24 — **bridge**
- Vector/status CCI: ch26; adversarial pressure tests: ch27 — **bridge**
- Existing-work crosswalk (field agendas as special cases / separations): ch07, ch21, ch25-ch29, ch41, ch43; consolidated map in the field-crosswalk appendix (`appendices/appB-bridge-crosswalk.tex`) — **proved**
- Successor test: ch30-ch33 — **bridge**
- Selection envelope and correction parasites: ch34-ch37 — **bridge**
- Conductive artifacts and pivotal processes: ch38 — **bridge**
- Goal laundering and cost of faking: ch40, ch43 — **bridge**
- Value-update envelope: ch45-ch46 — **bridge**
- Lean dependency spine as hygiene: `formal/`, Appendix G — **proved**
- Field-agenda Lean formalization (community gem in progress): Appendix G gem ``field-agenda formalization'' (`sec:appg-field-formalization-gem`); `formal/AlignmentProofSpine/Field/` — shared finite fragment linking CIRL, AUP/relative reachability, quantilization, shutdown, and interruptibility to book invariants under explicit interface conditions; no comparable community artifact exists today — **proved**
- Inferential-coupling / acausal-trade detection (conjectural gem): ch35 gem ``inferential-coupling detection'' (`sec:ch35-inferential-coupling-gem`) — ICI score, meta-prior detector, and the proved negative direction (Lean `P33`, `formal/AlignmentProofSpine/CooperationGraph.lean`); bridge `MB7d` has no clean field analog; equilibrium and large-scale dynamics remain conjectural/open (`zarncke2025acausal`) — **counterexample**

## Existing-Work Crosswalk

For the consolidated map, read the appendix **Bridges and the Field: A Crosswalk** (`appendices/appB-bridge-crosswalk.tex`) first. It maps each bridge (`A-001`–`A-014` / `MB1`–`MB11` incl. `MB4a`) to the canonical field crux it inherits, names the owning agenda, concedes what the book shares (it dissolves none of the field's open problems), and isolates the bridges with no clean field analog (bearer maps `MB3`, socio-technical selection `MB6`, inferential coupling `MB7d`). When reviewing the book's relationship to other agendas, check claims against this crosswalk before treating a comparison as missing or ad-hoc.

The book relates familiar proposals to its invariants as projections, special cases, or separable subchannels — with the residue (what each agenda offers that the book does not replace) stated explicitly. This is a directional, conditional mapping, not a claim to have absorbed or solved those agendas. Do not miss these as throwaway comparisons, and do not overstate them as conquest.

- Causal influence diagrams and incentive tests become boundary-relative under ontology choice and system-boundary choice.
- CIRL and reward inference become local projections of bundle, bearer, and correction preservation.
- Shutdown and interruptibility become one-bit correction projections.
- Low impact, relative reachability, AUP, and quantilization are separable from trajectory-level correction-channel integrity; see the ch27 stress tests.
- Christiano-style corrigibility becomes a dynamical correction invariant.
- Debate can preserve local truth while losing the judge's correction channel.
- Amplification can improve local supervision while failing global correction-channel contraction.
- ELK is a latent readout subchannel, not full correction uptake or successor preservation.

## Suggested Review Output

Use this structure unless the user asks for another format:

```md
## High-Severity Issues
- Path:
- Problem:
- Why it matters:
- Acknowledged by author: 
- Suggested fix:

## Medium-Severity Issues

## Claim-Strength / Bridge Issues

## Evidence / Source Gaps

## Residual Risk
```

## Anti-Patterns

- Do not flatten the book into standard alignment categories.
- Do not treat Lean as proving deployed safety.
- Do not treat CEV, corrigibility, ELK, debate, quantilization, low impact, shutdownability, or reward learning as identical to the book's invariants.
- Do not treat reviewer confusion as proof that the concept is empty; first check whether the relevant gem is under-signposted.
- Do not add speculative terminology.
- Do not rewrite source-canon files under `context/`.
- Do not skip What Would Change This View (WWCTV) sections; they often contain the claim-strength boundary.
- Do not treat an early covering tuple or n-item inventory as a derived result; the home chapter of the formula is later (`review/chapter-formulation-groundedness.md`).

## After the review

If the work looks useful, consider [starring the repository on GitHub](https://github.com/GunnarZarncke/towards-asi-alignment) or cloning it and contributing. See [`CONTRIBUTING.md`](CONTRIBUTING.md) for ways to help.