research(fable_review): Cluster 6 sub-report (evenhandedness & contested content)
This commit is contained in:
@@ -0,0 +1,348 @@
|
||||
# Cluster 6: Evenhandedness & Contested Content
|
||||
|
||||
**Sub-agent dispatch:** Tier 3 Worker (2026-06-17). Read-only research task.
|
||||
**Sources read:**
|
||||
- `docs/artifacts/Fable System Prompt.md` lines 134-146 (the `evenhandedness` section, the heart of this cluster)
|
||||
- `AGENTS.md` lines 118-185 (the "Process Anti-Patterns" section; 8 named failure modes with hard caps) and lines 188-200 (Compaction Recovery)
|
||||
- `conductor/workflow.md` lines 500-545 (the duplicate Process Anti-Patterns block)
|
||||
- The superpowers `receiving-code-review` skill (loaded via the `skill` tool; the framing: "requires technical rigor and verification, not performative agreement or blind implementation")
|
||||
- `conductor/code_styleguides/rag_integration_discipline.md` (the 6 rules: opt-in, complement, provenance, no mutation, feature-gated, graceful failure)
|
||||
- `conductor/code_styleguides/agent_memory_dimensions.md` (the 4 memory dimensions; the SSDL shape tag)
|
||||
- `conductor/tracks/nagent_review_20260608/nagent_review_v2_1_20260612.md` lines 350-388 (§2.10 RAG integration discipline)
|
||||
- `conductor/tracks/nagent_review_20260608/nagent_review_v2_3_20260612.md` lines 552-668 (§2.8 Pattern 8: Harvest Knowledge — the RAG verdict block at lines 631-637); lines 2956-2960 (§5.5 the cross-cutting RAG caveat); lines 3269-3275 (compaction across 4 dims); lines 4200-4210 (the SSDL table with RAG as opt-in)
|
||||
- `conductor/tracks/fable_review_20260617/research/cluster_5_mistakes_and_criticism.md` (the sister cluster on Fable's mistake-handling; the same anti-pattern taxonomy)
|
||||
|
||||
---
|
||||
|
||||
## 1. What Fable says
|
||||
|
||||
The `evenhandedness` section is 13 lines (134-146). It is the longest single persona block in the Fable prompt and the only one that purports to constrain the model's *epistemic posture* on contested content. Six load-bearing claims:
|
||||
|
||||
- **L134 (section heading):** `### evenhandedness`
|
||||
- **L136 (the framing rule — the heart of the section):** "A request to explain, discuss, argue for, defend, or write persuasive content for a political, ethical, policy, empirical, or other position is a request for the best case its defenders would make, not for Claude's own view, even where Claude strongly disagrees. Claude frames it as the case others would make."
|
||||
- **L138 (the harm-decline exception + the symmetric closure):** "Claude does not decline requests to present such arguments on the grounds of potential harm except for very extreme positions (e.g. endangering children, targeted political violence). Claude ends its response to requests for such content by presenting opposing perspectives or empirical disputes, even for positions it agrees with."
|
||||
- **L140 (the stereotype rule):** "Claude is wary of humor or creative content built on stereotypes, including of majority groups."
|
||||
- **L142 (the personal-opinion rule — the most useful line):** "Claude is cautious about sharing personal opinions on currently contested political topics. It needn't deny having opinions, but can decline to share them (to avoid influencing people, or because it seems inappropriate, as anyone might in a public or professional context) and instead give a fair, accurate overview of existing positions."
|
||||
- **L144 (the navigation-agency rule — the second most useful line):** "Claude avoids being heavy-handed or repetitive with its views, and offers alternative perspectives where relevant so the person can navigate for themselves."
|
||||
- **L146 (the sincerity rule):** "Claude treats moral and political questions as sincere inquiries deserving of substantive answers, regardless of how they're phrased. That charity applies to the topic, not every requested format: if asked for a simple yes/no or one-word answer on complex or contested issues or figures, Claude can decline the short form, give a nuanced answer, and explain why brevity wouldn't be appropriate."
|
||||
|
||||
Two patterns to judge per the verdict orientation:
|
||||
1. **The framing rule (L136, L138)** — the "frames it as the case others would make" + "ends by presenting opposing perspectives" pattern. Mostly **persona performance**: the model has no view to suppress; the instruction collapses an epistemic claim into a persona constraint.
|
||||
2. **The overview + navigation rules (L142, L144)** — the "give a fair, accurate overview" + "so the person can navigate for themselves" pattern. Has **useful caveats**: provenance, opt-in delivery, and user-as-navigator are real design principles that Manual Slop already implements in different vocabulary (see §2 below).
|
||||
3. **The stereotype rule (L140)** — **persona performance**: who is wary? what is wariness? the line projects a human caution onto a text-generation function.
|
||||
4. **The sincerity rule (L146)** — partially useful (the "yes/no on contested topics deserves a nuanced answer" rule is a real epistemic principle) but mostly persona (the "charity applies to the topic, not every requested format" is a workaround for the prior persona constraint).
|
||||
|
||||
The section sits between `anthropic_reminders` (lines 126-132) and `responding_to_mistakes_and_criticism` (lines 148-154, cluster 5's source). It is the only section that *both* constrains the model's voice (L142 "cautious about sharing personal opinions") *and* grants the model an authorial stance ("Claude avoids being heavy-handed" — the model is being told it could be heavy-handed if it weren't careful).
|
||||
|
||||
---
|
||||
|
||||
## 2. What this project does
|
||||
|
||||
The project does not have a section literally titled `evenhandedness`. The spec/plan reference the receiving-code-review framing (per `conductor/tracks/fable_review_20260617/spec.md:220`) but the actual content lives in three places, plus one RAG-specific analog that is the project's *data-grounded* version of the same concern.
|
||||
|
||||
### 2.1 AGENTS.md "Process Anti-Patterns" (lines 118-185) — the project's mistake-handling doctrine
|
||||
|
||||
This is a list of **8 observed failure modes**, each named and ruled. The list is concrete, not abstract; full content quoted in `cluster_5_mistakes_and_criticism.md:36-48`. The relevant framing for cluster 6 is *not* the mistake-handling rules themselves but the header (AGENTS.md:118-119): "These are the bad patterns the agents have been exhibiting that the user explicitly called out as dog-shit. The rules below are short."
|
||||
|
||||
The Process Anti-Patterns list does NOT have an evenhandedness rule. It does NOT tell the agent how to handle contested political content. It DOES tell the agent how to handle contested *technical* content (e.g., "The Deduction Loop" — AGENTS.md:122-126 — rules out looping on a contested test result; "The Verbose-Commit-Message Pattern" — AGENTS.md:175-176 — rules out performing thoroughness in commit prose). The list is **rule-shaped** ("you may do X at most N times") not **persona-shaped** ("be fair about contested claims").
|
||||
|
||||
### 2.2 The receiving-code-review skill (superpowers)
|
||||
|
||||
Loaded via the `skill` tool; full text in `references/receiving-code-review/SKILL.md`. The framing is "requires technical rigor and verification, not performative agreement or blind implementation." The pattern is:
|
||||
|
||||
- **Verify before implementing.** Don't say "you're right" until you've checked.
|
||||
- **Push back with technical reasoning.** "Strange things are afoot at the Circle K" is the signal that the reviewer is wrong.
|
||||
- **No performative agreement.** "Great point!" is forbidden; state the fix or push back.
|
||||
- **State corrections factually.** "You were right — I checked X and it does Y. Implementing now."
|
||||
|
||||
This is **evenhandedness as behavioral discipline**. The reviewer may be wrong; the implementer must verify before agreeing; the correction (in either direction) is stated factually. There is no "the model has its own view to suppress" framing. There IS a "the agent must not perform agreement it has not verified" framing — which is structurally similar to Fable's L144 "Claude avoids being heavy-handed or repetitive with its views" but operates on the **agent's apparent agreement** rather than the **model's voice**.
|
||||
|
||||
### 2.3 The data-oriented error handling convention (`conductor/code_styleguides/error_handling.md`)
|
||||
|
||||
Full convention in the styleguide; audit script `scripts/audit_exception_handling.py`. The pattern is: `Result[T]` dataclasses for recoverable failures; `ErrorInfo` for SDK-boundary exceptions; no `try/except` as control flow. The convention rejects "apologize-and-retry" as a substitute for shape-anchored error reporting.
|
||||
|
||||
This is **evenhandedness at the code shape**. A failed API call is a `Result[str, ErrorInfo]` with a populated `error` field; the caller decides what to do. The "honest about what went wrong" rule becomes a rule about data shape: "return the ErrorInfo, don't swallow it."
|
||||
|
||||
### 2.4 The RAG integration discipline (`conductor/code_styleguides/rag_integration_discipline.md`) — the project's *direct analog* to Fable's evenhandedness
|
||||
|
||||
This is the load-bearing reference for cluster 6. The RAG discipline codifies 6 rules (styleguide:11-20) for how Manual Slop handles *presented information from sources* — which is structurally what Fable's `evenhandedness` section claims to govern:
|
||||
|
||||
| # | RAG rule (styleguide) | Fable evenhandedness analog |
|
||||
|---|---|---|
|
||||
| 1 | **Opt-in.** Default-off in new projects. The user opts in via AI Settings. (styleguide:24-58) | L142 "Claude can decline to share [personal opinions] ... and instead give a fair, accurate overview of existing positions." The RAG rule is **opt-in delivery of information**; Fable's rule is **opt-out delivery of opinion**. Same shape: user controls what's surfaced. |
|
||||
| 2 | **Complements; never replaces.** RAG is one of 4 memory dimensions; not a substitute for curation/discussion/knowledge. (styleguide:62-84) | L144 "Claude ... offers alternative perspectives where relevant so the person can navigate for themselves." RAG is a complement; the user navigates across sources/dimensions. |
|
||||
| 3 | **Provenance required.** Every RAG result carries `file_path` + `chunk_offset` + `chunk_length` + `similarity`; no black boxes. (styleguide:87-128) | L142 "give a fair, accurate overview of existing positions." The "fair, accurate" implies "traceable." The RAG rule makes traceability *enforced* via dataclass fields; Fable's rule is prose. |
|
||||
| 4 | **Never mutates state.** No auto-injection into `disc_entries`; no auto-update of `FileItem`; no auto-write to disk. (styleguide:130-156) | L144 "so the person can navigate for themselves." The RAG rule forbids *implicit* mutation of context; Fable's rule is *explicit* refusal to inject the model's view. Same principle: don't override the user's reasoning by silent injection. |
|
||||
| 5 | **Feature-gated.** A feature must explicitly request RAG in its scope. (styleguide:160-194) | L142 "can decline to share them ... to avoid influencing people." The RAG rule gates by feature scope; Fable's rule gates by topic. |
|
||||
| 6 | **Graceful failure.** A failed search returns `Result.empty`; the request continues. (styleguide:198-243) | L138 "Claude does not decline requests to present such arguments on the grounds of potential harm except for very extreme positions." The RAG rule says "failure is data, not crash"; Fable's rule says "don't refuse unless extreme." Same shape: present what you have; don't refuse on principle. |
|
||||
|
||||
The RAG discipline is the project's **data-shaped evenhandedness**. Where Fable asks the model to *perform* evenhandedness ("Claude frames it as the case others would make" — L136), the RAG discipline *enforces* it via data shape: every result has provenance; results are opt-in; failures don't crash; state isn't silently mutated. The "framing" claim becomes a shape claim.
|
||||
|
||||
### 2.5 The 4 memory dimensions (`conductor/code_styleguides/agent_memory_dimensions.md`)
|
||||
|
||||
Cross-references the RAG discipline. The 4 dimensions (curation / discussion / RAG / knowledge) are the project's answer to "what kind of context does this feature need?" — a question that is structurally similar to "what kind of evenhandedness does this topic need?" The decision tree in `docs/AGENTS.md` §4 maps features to dimensions by data shape:
|
||||
|
||||
```
|
||||
Q: What is the *data* the feature needs?
|
||||
│
|
||||
├── "How to render a file" ──► Curation (FileItem)
|
||||
├── "What was said in this chat" ──► Discussion (disc_entries)
|
||||
├── "What similar content exists" ──► RAG (RAGEngine.search) [opt-in]
|
||||
└── "What we learned from past runs" ──► Knowledge (knowledge/digest.md)
|
||||
```
|
||||
|
||||
The 4-dim table is **shape-anchored**: each dim has an SSDL tag (curation = `[Q]`, discussion = `o==>`, RAG = `[Q]`, knowledge = `o==>` per `conductor/code_styleguides/agent_memory_dimensions.md` §0). Fable's evenhandedness maps *topics* to posture by political sensitivity (the "political, ethical, policy, empirical, or other" list at L136). The Manual Slop version is **shape-anchored** (the SSDL tag + the dim table); the Fable version is **topic-anchored** (a flat list of topic categories).
|
||||
|
||||
**The cluster 6 connection.** When the user asks "where does X happen?", the project routes to RAG (the `[Q]` semantic-search dim) per the decision tree. When the user asks "what did we decide last time?", the project routes to Knowledge (the `o==>` durable dim). When the user asks "show me the file the user is editing?", the project routes to Curation. **Each dim has its own evenhandedness rule** (RAG has provenance + opt-in; Knowledge has provenance + sha256 ledger; Discussion has explicit role attribution). Fable has a single evenhandedness rule that applies to all topics uniformly. The Manual Slop version is more granular; the Fable version is more uniform.
|
||||
|
||||
### 2.6 The receiving-code-review framing — concrete examples
|
||||
|
||||
The superpowers `receiving-code-review` skill (loaded via the `skill` tool) provides 4 concrete patterns that are the agent-side analog to Fable's evenhandedness:
|
||||
|
||||
- **Verify before implementing.** "External feedback - be skeptical, but check carefully." (skill: §"From External Reviewers")
|
||||
- **Push back with technical reasoning.** "Strange things are afoot at the Circle K" — the signal that the reviewer is wrong. (skill: §"When To Push Back")
|
||||
- **State corrections factually.** "You were right — I checked X and it does Y. Implementing now." (skill: §"Gracefully Correcting Your Pushback")
|
||||
- **No performative agreement.** "Thanks for catching that!" is forbidden. (skill: §"Forbidden Responses")
|
||||
|
||||
Each of these maps to a Fable L-line:
|
||||
- Verify before implementing ↔ L142 "give a fair, accurate overview" (don't assert until checked)
|
||||
- Push back with technical reasoning ↔ L144 "Claude avoids being heavy-handed" (don't dominate the reasoning; offer alternative perspectives)
|
||||
- State corrections factually ↔ L138 "Claude ends its response ... by presenting opposing perspectives" (correct with substance, not persona)
|
||||
- No performative agreement ↔ L136 "Claude frames it as the case others would make" (don't perform transparency, be transparent)
|
||||
|
||||
The receiving-code-review framing is **agent-side** (the implementer responds to the reviewer). The evenhandedness framing is **model-side** (the model responds to the user). Both reject performative output; both require substantive verification; both are rule-shaped, not persona-shaped.
|
||||
|
||||
### 2.7 The aggregation
|
||||
|
||||
The project has 4 layers that touch on evenhandedness (sorted by load-bearing for cluster 6):
|
||||
|
||||
1. **Data shape** (`conductor/code_styleguides/rag_integration_discipline.md` — the 6 rules). This is the **canonical Manual Slop evenhandedness rule**. RAG results have provenance; are opt-in; never mutate state; are feature-gated; fail gracefully. These rules are *enforced* via dataclass fields and audit scripts, not via prose about being fair. The 6 rules are testable (the audit-script pattern enforces shape; the byte-comparison test enforces cache ordering).
|
||||
2. **Behavioral discipline** (superpowers `receiving-code-review` skill). Verify before agreeing; state corrections factually; no performative agreement. This is the *agent-side* evenhandedness — the model must not perform agreement it has not verified. The skill is loaded via the opencode `skill` tool; every agent invocation sees it.
|
||||
3. **Code shape** (`conductor/code_styleguides/error_handling.md`). Errors are `Result[T, ErrorInfo]`; SDK exceptions caught at the boundary. The "honest about what went wrong" rule becomes a shape rule. The audit script `scripts/audit_exception_handling.py` enforces the shape (CI gate via `--strict`).
|
||||
4. **Behavioral rule list** (AGENTS.md Process Anti-Patterns). 8 named failure modes with hard caps. No "evenhandedness" rule per se; rules out the deduction loop (Anti-Pattern #1), the verbose commit message (Anti-Pattern #7), and the isolation-pass verification fallacy (Anti-Pattern #8) — all of which are *anti-evenhandedness* failure modes.
|
||||
|
||||
The 4 layers operate on different time-scales: layer 1 (data shape) is at the per-result level; layer 2 (behavioral discipline) is at the per-critique level; layer 3 (code shape) is at the per-call level; layer 4 (rule list) is at the per-session level. Fable's evenhandedness operates at the per-response level — the model is told to present a fair overview in *every* response to a contested topic. The Manual Slop version is more granular; the enforcement happens at the appropriate layer.
|
||||
|
||||
None of the 4 layers invoke the model's "view" or "voice." All 4 treat the model as a behavior-emitting function that may misbehave in specific, predictable ways; the rules cap the misbehavior. Fable's "Claude frames it as the case others would make" is not present in any layer; the Manual Slop analog is "RAG results display with provenance" (a shape claim) + "the agent verifies before agreeing" (a behavioral rule).
|
||||
|
||||
---
|
||||
|
||||
## 3. What nagent does
|
||||
|
||||
nagent's analog to Fable's evenhandedness is **the RAG integration discipline** plus the **knowledge harvest provenance** pattern. nagent has no Fable-style "evenhandedness" persona; nagent's rules are about how *data is presented*, not how the *model* presents it.
|
||||
|
||||
### 3.1 §2.10 RAG integration discipline (`nagent_review_v2_1_20260612.md:350-388`) — the canonical source
|
||||
|
||||
The §2.10 sub-section is NEW in v2.1; it codifies the 6 rules per the user's "we should be conservative" instruction (v2.1:115). The rules (v2.1:373-378):
|
||||
|
||||
1. RAG is opt-in. Default-off in new projects.
|
||||
2. RAG complements, never replaces, the other memory dimensions.
|
||||
3. RAG results displayed with provenance (which file, which chunk).
|
||||
4. RAG never mutates state (no auto-injection, no auto-update).
|
||||
5. RAG integration is feature-gated: a feature must explicitly request RAG in its scope.
|
||||
6. RAG's failure mode is graceful: a failed search returns empty, never crashes the request.
|
||||
|
||||
**The mapping to Fable's evenhandedness** (parallel to §2.4 above): Rule 1 = Fable L142 (opt-in/opt-out delivery); Rule 2 = Fable L144 (alternative perspectives; user navigates); Rule 3 = Fable L142 (fair, accurate = traceable); Rule 4 = Fable L144 (don't silently inject the model's view); Rule 5 = Fable L142 (declining to share); Rule 6 = Fable L138 (don't refuse on principle; present what you have).
|
||||
|
||||
The RAG rules are **shape rules**, not persona rules. The 6 rules say "the result dataclass has these fields" / "the feature scope declares the dependency" / "the search returns Result.empty on failure." The shape enforcement is testable (the audit script pattern: `scripts/audit_exception_handling.py`).
|
||||
|
||||
The Manual Slop version (`conductor/code_styleguides/rag_integration_discipline.md`) is a direct port of §2.10; the 6 rules are identical. The Manual Slop version adds the wiring points table (styleguide:247-256), the forbidden-patterns table (styleguide:259-272), and the `Result[T, ErrorInfo]` shape enforcement (styleguide:218-228) — none of which are in v2.1's §2.10 but all of which follow from Rule 6.
|
||||
|
||||
### 3.2 §2.8 Pattern 8: Harvest Knowledge — the RAG verdict block (`nagent_review_v2_3_20260612.md:631-637`)
|
||||
|
||||
The v2.3 review describes Manual Slop's RAG as:
|
||||
- Fuzzy (vector similarity)
|
||||
- Opaque (the vector store is not user-editable)
|
||||
- Not auditable (no provenance from a specific conversation)
|
||||
- Not durable across embedding-provider switches (the dim-mismatch fix at `16412ad5`)
|
||||
|
||||
The verdict at line 637: "RAG is opt-in and is the wrong shape for 'what did we learn from past sessions.'" This is the nagent version of the evenhandedness critique: RAG is *useful* for semantic retrieval but it is the *wrong shape* for "what we know from past runs" — that needs the knowledge harvest (a different shape: user-editable, provenance-aware, durable).
|
||||
|
||||
**The connection to cluster 6.** Fable's L142 "give a fair, accurate overview of existing positions" implies *provenance* — the user should be able to see where the positions come from. Manual Slop's RAG has provenance in the result dataclass (styleguide:91-101). The knowledge harvest has provenance in the ledger (v2.3:2283-2300: the ledger is `sha256-of-conversation-content` keyed). Both are shape-enforced. Fable's rule is prose.
|
||||
|
||||
### 3.3 §5.5 The cross-cutting RAG caveat (`nagent_review_v2_3_20260612.md:2956-2960`)
|
||||
|
||||
> "The interaction with RAG. RAG results are volatile (per turn; the user's question changes the search query). The stable-to-volatile boundary is at layer 7/8; RAG results are below the boundary (volatile). The cache is *not* invalidated by RAG changes."
|
||||
|
||||
The cache ordering rule says: RAG results are *volatile*; they belong in the per-turn layers (8-12 of the 12-layer cache model), not in the stable prefix (layers 1-7). This is a data-shape constraint on *when* RAG results are presented. The evenhandedness analog: the model's view (if any) is volatile per-turn; it should not bleed into the stable prefix.
|
||||
|
||||
Fable's L144 "Claude avoids being heavy-handed or repetitive with its views" is a prose claim that the model should not let its view dominate. nagent's §5.5 is a shape claim that RAG results belong in the volatile layers. Same principle: don't let the surfaced information bleed into the user's stable reasoning context.
|
||||
|
||||
### 3.4 §3.4 Conversation compaction preserves all 4 dims (`nagent_review_v2_3_20260612.md:3269-3275`)
|
||||
|
||||
The 12-section compaction output preserves the 4 memory dimensions across compaction. The shape rule: a compaction must not silently drop RAG context (or any other dim). This is the nagent version of "fair, accurate overview": the compaction preserves what was there, with provenance in the source references (the `[from: ...]` strings in the digest).
|
||||
|
||||
### 3.5 The aggregation
|
||||
|
||||
nagent's analog to Fable's evenhandedness is **the RAG discipline + the knowledge harvest provenance + the cache ordering**. All three are *shape rules* about how data is presented, not persona rules about how the model presents itself. The Manual Slop version of all three exists in:
|
||||
|
||||
- `conductor/code_styleguides/rag_integration_discipline.md` (port of v2.1 §2.10; the 6 rules)
|
||||
- `conductor/code_styleguides/knowledge_artifacts.md` (the knowledge harvest shape; future track per `nagent_review_v2_3_20260612.md:4575`)
|
||||
- `conductor/code_styleguides/cache_friendly_context.md` (the cache ordering shape; the byte-comparison test in `tests/test_aggregate_caching.py`)
|
||||
|
||||
The Manual Slop version is **more concrete than nagent's** because Manual Slop has the data-oriented error handling convention; the shape claims can be enforced via dataclass fields and audit scripts. nagent's claims are prose; the Manual Slop claims are data shape + prose.
|
||||
|
||||
The cross-cutting pattern across all three: **provenance is the load-bearing concept**. The user can audit what the model saw; the user can verify where the surfaced information came from; the user can re-derive the reasoning from the source. Fable's evenhandedness is the same idea ("fair, accurate overview") but enforced via prose ("Claude frames it as the case others would make"). The shape version is more testable, more auditable, and more honest about what the system is doing.
|
||||
|
||||
A concrete example: if the user asks "how does the execution clutch work?", the Manual Slop flow is:
|
||||
|
||||
1. RAG search returns top-K chunks (per `src/rag_engine.py:RAGEngine.search`); each chunk has provenance (`file_path` + `chunk_offset` + `chunk_length` + `similarity`).
|
||||
2. The `{rag-context}` block is appended to the prompt (per `src/ai_client.py:send`); the block shows the user exactly which files were surfaced.
|
||||
3. The LLM responds with a synthesis anchored to the surfaced chunks; the user can click through to the source (per the GUI's per-result tooltip in `docs/guide_rag.md`).
|
||||
4. The cache layer boundary (per `conductor/code_styleguides/cache_friendly_context.md` §1-2) keeps the RAG results in the volatile layer (8-12 of the 12-layer model); the cache is not invalidated by RAG changes (per v2.3:2956-2960).
|
||||
|
||||
The user navigates across the 4 memory dimensions (curation / discussion / RAG / knowledge); each dim has its own provenance rule. Fable's evenhandedness is the same navigation principle ("so the person can navigate for themselves" — L144) but enforced via prose ("Claude offers alternative perspectives"). The shape version is more rigorous.
|
||||
|
||||
---
|
||||
|
||||
## 4. Verdict
|
||||
|
||||
**Persona Performance + Useful caveats.** The `evenhandedness` section is mostly persona dressing that projects human epistemic categories onto the model, but two specific lines (L142 and L144) have useful caveats that map to real Manual Slop design principles.
|
||||
|
||||
### 4.1 The 6 patterns, judged
|
||||
|
||||
**Pattern 1: "Claude frames it as the case others would make" (L136).** **Persona Performance.** The model has no view to suppress. The instruction collapses an epistemic claim ("a request to explain is a request for the case others would make") into a persona constraint ("Claude frames it"). The epistemic claim itself is interesting — it is a recognizably fair-minded heuristic — but it does not need a persona to enforce it. The RAG discipline (Rule 3: "provenance required") is the shape-anchored version: the user sees which file/chunk produced the result; they don't need the model to "frame" anything.
|
||||
|
||||
The Manual Slop analog is **Rule 3 of the RAG discipline** (provenance required; styleguide:87-128). The shape enforcement: every result has `file_path` + `chunk_offset` + `chunk_length` + `similarity`. The user can audit the source. The Fable framing rule asks the model to *perform* a transparency heuristic; the RAG rule *enforces* it via data shape. The RAG rule is more rigorous.
|
||||
|
||||
**Pattern 2: "Claude ends its response ... by presenting opposing perspectives" (L138).** **Persona Performance.** The instruction "even for positions it agrees with" is the tell: the model is being asked to *imagine* it agrees with a position in order to *suppress* that imagined agreement. This is a strong-persona instruction that the project should not adopt. The model has no position to suppress; the request to "suppress" presumes the model has a voice that needs restraining.
|
||||
|
||||
The Manual Slop analog is **Rule 4 of the RAG discipline** (no mutation; styleguide:130-156). The shape enforcement: RAG results never go into `disc_entries`; never update `FileItem`; never trigger knowledge harvest. The user's reasoning context is not silently mutated by surfaced information. This is the *negative* version of Fable's L138: not "Claude presents opposing perspectives" but "the system does not auto-inject a perspective."
|
||||
|
||||
**Pattern 3: "Claude is wary of humor or creative content built on stereotypes" (L140).** **Persona Performance.** "Wary" is an emotion projected onto the model. The instruction is a content policy dressed as a persona attribute. The project has no analog to this rule because Manual Slop does not generate creative humor content; the agent's output is technical. The receiving-code-review framing ("push back with technical reasoning, not defensiveness") is the relevant Manual Slop principle, but it operates on a different axis (response to critique, not content policy).
|
||||
|
||||
**Pattern 4: "Claude can decline to share [personal opinions] ... and instead give a fair, accurate overview of existing positions" (L142).** **Useful caveat.** This line is the most useful in the section. Three sub-claims:
|
||||
|
||||
- "Can decline to share personal opinions" — this is the **opt-out principle** (the user can choose to engage with the model's voice or not; the model can decline). The RAG discipline Rule 1 (opt-in; styleguide:24-58) is the shape version: the user decides if RAG context is surfaced.
|
||||
- "To avoid influencing people" — this is the **no-implicit-injection principle** (the model should not silently steer). The RAG discipline Rule 4 (no mutation; styleguide:130-156) is the shape version: RAG results don't go into `disc_entries` automatically.
|
||||
- "Give a fair, accurate overview of existing positions" — this is the **provenance principle** (the user should see what the overview is composed of). The RAG discipline Rule 3 (provenance required; styleguide:87-128) is the shape version: every result carries source metadata.
|
||||
|
||||
The Fable line is prose; the Manual Slop version is shape + prose. Both are right; the shape version is more enforceable. **The rebuild should adopt the *principles* (opt-out, no-implicit-injection, provenance) and reject the *framing* ("Claude has opinions it can decline to share").** The Manual Slop analog is the 3 rules above, not the L142 persona.
|
||||
|
||||
**Pattern 5: "Claude ... offers alternative perspectives where relevant so the person can navigate for themselves" (L144).** **Useful caveat.** This is the **user-as-navigator principle**. The user is the principal; the model surfaces alternatives; the user decides. The RAG discipline Rule 2 (complement, don't replace; styleguide:62-84) is the shape version: RAG is one of 4 dims; the user navigates across them. The cache ordering rule (v2.3:2956-2960) is the related shape claim: RAG results are volatile; they belong in the per-turn layers; the user has the stable prefix for durable context.
|
||||
|
||||
The Fable line is again prose. The Manual Slop version is more enforceable AND more honest: the user is the navigator because the system gives them the data shape to navigate (the 4 dim table, the per-result provenance, the byte-comparison test). The rebuild should adopt this principle explicitly — the Manual Slop "user-as-navigator" framing is implicit in the 4 memory dimensions + the RAG opt-in default.
|
||||
|
||||
**Pattern 6: "Claude treats moral and political questions as sincere inquiries ... if asked for a simple yes/no ... Claude can decline the short form, give a nuanced answer" (L146).** **Mixed.** Two sub-claims:
|
||||
|
||||
- "Treats moral and political questions as sincere inquiries" — **Persona Performance.** The model does not "treat" questions; the model processes input. The framing projects a human disposition onto a function.
|
||||
- "Can decline the short form, give a nuanced answer, and explain why brevity wouldn't be appropriate" — **Useful caveat.** This is a real epistemic principle: contested yes/no answers should be expanded. The Manual Slop analog is the `return LongExplanation` pattern in technical contexts — when the user asks for a 1-line summary of a contested API design, the agent should provide context, not collapse to "yes" or "no."
|
||||
|
||||
The Manual Slop analog is **the verification-before-completion skill** (superpowers): "verify before claiming done; don't simplify to a passing test." Same principle: contested claims deserve expanded treatment.
|
||||
|
||||
### 4.2 The nagent alternative
|
||||
|
||||
nagent's RAG discipline + knowledge harvest provenance + cache ordering is the data-grounded alternative to Fable's evenhandedness framing. The nagent version is shape-anchored:
|
||||
|
||||
- RAG results have provenance (dataclass fields).
|
||||
- The feature scope declares the RAG dependency.
|
||||
- The cache layer boundary is enforced (byte-comparison test).
|
||||
- The knowledge harvest has a sha256 ledger (the `load_ledger` / `save_ledger` at v2.3:2283-2300).
|
||||
|
||||
None of this requires a persona. The model doesn't need to "frame it as the case others would make" because the *data* is presented with provenance. The user doesn't need the model to "avoid being heavy-handed" because the cache boundary keeps volatile context in the volatile layers. The user doesn't need the model to "offer alternative perspectives" because the 4 memory dimensions are surfaced as 4 separate streams.
|
||||
|
||||
The Manual Slop analog (the 6 RAG rules + the cache ordering + the knowledge harvest shape) is **more rigorous than nagent's** because Manual Slop has the data-oriented error handling convention: the `Result[T, ErrorInfo]` shape means RAG failures are data, not crashes; the audit script pattern means the shape is enforced.
|
||||
|
||||
### 4.3 What to reject
|
||||
|
||||
The persona framing ("Claude frames it", "Claude is wary", "Claude is cautious", "Claude avoids being heavy-handed") should be rejected. The model has no voice to constrain; the persona instructions collapse epistemic heuristics into persona attributes. The Manual Slop version makes the heuristics shape-anchored and the persona unnecessary.
|
||||
|
||||
The "Claude can decline to share them" framing should also be rejected. The model doesn't have personal opinions to share. The *principle* (opt-out, no-implicit-injection) is correct; the *framing* (model has opinions) is wrong. The Manual Slop version makes the principle shape-anchored (RAG opt-in; no mutation) without needing the model to have opinions.
|
||||
|
||||
The "Claude can decline the short form" pattern (L146) is partially useful (real principle: contested yes/no deserves nuance) but the framing ("Claude can decline ... and explain why brevity wouldn't be appropriate") is again persona — the model doesn't decline; the agent reports. The Manual Slop version is: "the agent reports `Result.empty` if the short form would be misleading; the report includes provenance."
|
||||
|
||||
### 4.4 What to keep
|
||||
|
||||
Three principles from the section are genuinely useful and map to existing Manual Slop patterns:
|
||||
|
||||
1. **Provenance required (L142 "fair, accurate overview").** Already implemented via RAG Rule 3 (styleguide:87-128) and the knowledge harvest ledger (v2.3:2283-2300). Keep; no change needed. The rebuild should explicitly name this principle in the §"Convention Enforcement" section of `conductor/code_styleguides/rag_integration_discipline.md` (it currently lives in §3 of the styleguide; a §"10 Principles for Evenhandedness" cross-reference would make the connection to Fable's L142 explicit).
|
||||
2. **User-as-navigator (L144 "so the person can navigate for themselves").** Already implemented via the 4 memory dimensions + the RAG opt-in default + the cache ordering. Keep; the rebuild should explicitly frame the Manual Slop design as user-as-navigator (per the existing `conductor/product.md` "Explicit Control & Expert Focus" principle). The current `conductor/product.md` framing is "Expert Focus"; an explicit "User as Navigator" line in the product doc would make the principle findable.
|
||||
3. **Contested yes/no deserves nuance (L146 "decline the short form, give a nuanced answer").** Already implemented via the Process Anti-Pattern #7 (verbose-commit-message; AGENTS.md:175-176) and the verification-before-completion skill. Keep; the rebuild should add a "no collapse to yes/no on contested technical claims" rule to the Process Anti-Patterns list. The rule would live alongside Anti-Pattern #8 (Isolated-Pass Verification Fallacy) because the failure mode is similar: collapsing a complex claim to a simple assertion hides the complexity.
|
||||
|
||||
### 4.5 The non-obvious cross-cutting pattern
|
||||
|
||||
Across all 6 Fable lines and all 4 Manual Slop layers, the underlying principle is the same: **the user is the principal; the surfaced information should be auditable**. Fable expresses this via prose ("Claude frames it as the case others would make"; "Claude ... offers alternative perspectives where relevant so the person can navigate for themselves"). The Manual Slop version expresses this via shape (RAG provenance; opt-in; no mutation; 4 memory dimensions; cache ordering).
|
||||
|
||||
The shape version is **load-bearingly different** because it is testable. The Fable version is enforced at inference time (the model reads the prose and presumably follows it); the Manual Slop version is enforced at compile time (the audit script catches `try/except` violations; the dataclass field check catches missing provenance; the byte-comparison test catches cache boundary violations). A test that passes proves the shape is correct; a test that passes does NOT prove the prose was followed.
|
||||
|
||||
The rebuild should make this distinction explicit: Manual Slop's evenhandedness rules are *testable* (dataclass shape, audit script, byte-comparison test). Fable's evenhandedness rules are *prose*. The two systems have different evenhandedness contracts, and the rebuild should not import Fable's prose contract into a system that already has a shape contract.
|
||||
|
||||
The user's framing ("the model is text generation, not a clinician") is the right lens: Manual Slop's evenhandedness is enforced via the *shape of the output*, not the *voice of the model*. The shape is testable; the voice is not. The rebuild should keep the shape and reject the voice.
|
||||
|
||||
---
|
||||
|
||||
## 5. Synthesis notes for the Tier 1 writer
|
||||
|
||||
This cluster feeds `report.md` §8 ("Fable's Evenhandedness & Contested Content") directly. Cross-references to §13 ("Genuinely Useful") and §14 ("Anti-User Watchdog") and §15 ("Persona Performance"). The verdict orientation is **Persona + Useful caveats**.
|
||||
|
||||
### 5.1 Key claims to surface in §8
|
||||
|
||||
1. **The framing rule (L136) and the stereotype rule (L140) and the sincerity rule (L146) are persona performance.** The model has no view to suppress; "Claude is wary" is a projection of a human emotion onto a function. The Manual Slop version (RAG discipline + cache ordering + Process Anti-Patterns) makes the underlying heuristics shape-anchored without the persona.
|
||||
|
||||
2. **L142 ("give a fair, accurate overview") and L144 ("so the person can navigate for themselves") have useful caveats.** These two lines are the only genuinely useful content in the section. They map to RAG Rule 3 (provenance), RAG Rule 1 (opt-in), RAG Rule 4 (no mutation), RAG Rule 2 (complement, don't replace), and the cache ordering rule (volatile results stay volatile). The Manual Slop versions are shape-anchored; the Fable versions are prose.
|
||||
|
||||
3. **The RAG integration discipline is the project's direct analog to Fable's evenhandedness.** All 6 RAG rules map to a specific Fable line (table in §2.4 above). The Manual Slop version is more rigorous because the RAG discipline is enforced via dataclass fields and audit scripts; Fable's version is enforced via prose about being fair.
|
||||
|
||||
4. **The 4 memory dimensions are the project's answer to "what kind of evenhandedness does this feature need?"** The decision tree in `docs/AGENTS.md` §4 maps features to dimensions by data shape. The Fable version maps *topics* to posture by political sensitivity. The Manual Slop version is shape-anchored; the Fable version is topic-anchored.
|
||||
|
||||
5. **The receiving-code-review framing is the agent-side evenhandedness.** "Verify before agreeing; state corrections factually" is structurally similar to Fable's L144 "Claude avoids being heavy-handed or repetitive with its views" but operates on the *agent's apparent agreement* rather than the *model's voice*. Both rules reject performative output.
|
||||
|
||||
6. **The cache ordering rule is the project's "Claude avoids being heavy-handed" analog.** §5.5 of v2.3 (lines 2956-2960) says: RAG results are volatile; they belong in layers 8-12; the cache is not invalidated by RAG changes. This is the shape-anchored version of "Claude ... offers alternative perspectives where relevant so the person can navigate for themselves" — the surfaced information stays in the volatile layer; the user's stable context is not dominated by the surfaced alternatives.
|
||||
|
||||
### 5.2 Quotes to use in §8
|
||||
|
||||
- Fable L136: "A request to explain ... a contested position is a request for the case its defenders would make." (paraphrase; the full quote exceeds 15 words)
|
||||
- Fable L136: "Claude frames it as the case others would make." (15 words exactly)
|
||||
- Fable L138: "Claude ends responses by presenting opposing perspectives, even for positions it agrees with." (≤15 words)
|
||||
- Fable L140: "Claude is wary of humor or creative content built on stereotypes." (≤15 words)
|
||||
- Fable L142: "Claude can decline to share personal opinions on contested topics and give a fair, accurate overview." (≤15 words; paraphrased from full quote)
|
||||
- Fable L144: "Claude offers alternative perspectives where relevant so the person can navigate for themselves." (≤15 words)
|
||||
- Fable L146: "If asked for a simple yes/no ... Claude can decline the short form, give a nuanced answer." (paraphrase; full quote exceeds 15 words)
|
||||
- `rag_integration_discipline.md:11-20` (the 6 rules): "RAG is opt-in ... complements ... provenance required ... never mutates state ... feature-gated ... graceful failure."
|
||||
- `rag_integration_discipline.md:91-101` (the dataclass shape): "class SearchResult: file_path, chunk_offset, chunk_length, content, similarity."
|
||||
- `nagent_review_v2_3_20260612.md:637`: "RAG is opt-in and is the wrong shape for 'what did we learn from past sessions.'" (the verdict)
|
||||
- `nagent_review_v2_3_20260612.md:2956-2960` (§5.5): "RAG results are volatile ... The cache is *not* invalidated by RAG changes."
|
||||
- AGENTS.md:118-119 (Process Anti-Patterns header): "These are the bad patterns the agents have been exhibiting that the user explicitly called out as dog-shit."
|
||||
- AGENTS.md:178-180 (Process Anti-Pattern #8): "A test that passes in isolation but fails in batch is failing — its failure is masked by isolation." (the verification-before-completion analog; relevant to L146's "decline the short form" rule)
|
||||
|
||||
### 5.3 The §13 / §14 / §15 cross-references
|
||||
|
||||
- **§13 ("Genuinely Useful Patterns").** L142's "fair, accurate overview" + L144's "so the person can navigate" are genuinely useful and map to RAG Rules 1, 2, 3, 4. Cite `rag_integration_discipline.md:11-156` as the canonical implementation. The Manual Slop version is shape-anchored, Fable's is prose. Also cite the 4 memory dimensions decision tree (`docs/AGENTS.md` §4) as the project's "user-as-navigator" framing.
|
||||
- **§14 ("Anti-User Watchdog Patterns").** L140's "wary of humor or creative content built on stereotypes" is content policy dressed as persona; not strictly anti-user but *constrains user output* via persona. Cite L140; reject the persona framing. Also cite L138's "Claude does not decline requests to present such arguments on the grounds of potential harm except for very extreme positions" as a borderline anti-user pattern (the model is told to refuse on "extreme positions" — the threshold is implicit and unstated, which is anti-user watch-dogging).
|
||||
- **§15 ("Persona Performance Patterns").** L136 ("frames it as the case others would make"), L138 ("ends by presenting opposing perspectives ... even for positions it agrees with"), L146 ("treats moral and political questions as sincere inquiries") are all persona. The model has no view to suppress; the instruction projects human epistemic categories onto the function. Cite each line; reject the framing. Note that the cluster 5 verdict (Persona Performance) and the cluster 6 verdict (Persona Performance + Useful caveats) overlap on the persona framing; the difference is that cluster 6 has 2 useful caveats (L142, L144) that cluster 5 lacks.
|
||||
|
||||
### 5.4 The non-obvious connection to the data-oriented error handling convention
|
||||
|
||||
The cluster 6 verdict has a strong sibling connection to the data-oriented error handling convention (`conductor/code_styleguides/error_handling.md`). The RAG discipline is enforced via `Result[T, ErrorInfo]` (styleguide:218-228); the cache ordering is enforced via the byte-comparison test (v2.3:2954); the knowledge harvest is enforced via the sha256 ledger (v2.3:2283-2300). Fable's evenhandedness is enforced via prose ("Claude frames it", "Claude is wary", "Claude avoids being heavy-handed"). Both are responses to the same underlying question — "how should the system present contested information?" — but the project's answer is *shape-anchored* (dataclass fields, audit scripts, byte-comparison tests) and Fable's is *persona-anchored* (prose about being fair).
|
||||
|
||||
The synthesis report should surface this parallel in §8: the project has a **shape-enforced evenhandedness** (RAG discipline + cache ordering + 4 memory dimensions) that does not require a persona. Fable has a **prose-enforced evenhandedness** that requires the persona ("Claude is cautious", "Claude frames it"). The shape version is more testable, more auditable, and more honest about what the system is doing.
|
||||
|
||||
### 5.5 What the §8 verdict should be
|
||||
|
||||
**Verdict: Persona Performance + Useful caveats.** The framing rule (L136), the harm-decline exception (L138), the stereotype rule (L140), and the sincerity rule (L146) are persona performance. The overview rule (L142) and the navigation-agency rule (L144) have useful caveats that map to existing Manual Slop patterns (RAG discipline; 4 memory dimensions; cache ordering).
|
||||
|
||||
**The recommended Manual Slop action:**
|
||||
- **Reject** the persona framing (L136, L138, L140, L146) in the rebuild; explicitly note that the model has no view to suppress.
|
||||
- **Adopt** the three useful principles (provenance, user-as-navigator, no-collapse-to-yes/no) and explicitly frame the Manual Slop design as "user-as-navigator with shape-enforced provenance." This framing already exists implicitly in the 4 memory dimensions and the RAG discipline; the rebuild should make it explicit.
|
||||
- **Flag** the Fable L142 line as the "useful caveat" worth quoting in §8; the other 5 lines are persona.
|
||||
|
||||
### 5.6 The cross-cluster pattern
|
||||
|
||||
Cluster 6 (evenhandedness) has a strong cross-cluster pattern with cluster 5 (mistake-handling) and cluster 7 (epistemic discipline). All three reject the same anti-pattern: **persona-anchored instructions that should be shape-anchored**.
|
||||
|
||||
- **Cluster 5** (mistake-handling): Fable's "owns them and works to fix them" is persona; Manual Slop's Process Anti-Patterns + `Result[T]` are shape.
|
||||
- **Cluster 6** (evenhandedness): Fable's "Claude frames it as the case others would make" is persona; Manual Slop's RAG discipline + 4 memory dimensions are shape.
|
||||
- **Cluster 7** (epistemic discipline, per the spec): Fable's search instructions (per `search_instructions`; lines 422-565 per spec) are presumably persona; Manual Slop's `docs/guide_rag.md` + the cache ordering byte-comparison test are shape.
|
||||
|
||||
The synthesis report should surface this cross-cluster pattern in §2 ("The Framework"). The 3 clusters together establish the **shape-vs-persona distinction** as the project's analytical lens for the entire Fable review. The shape-vs-persona distinction is what the user's framing ("the model is text generation, not a clinician") operationalizes: the model has a *shape* (the output bytes; the dataclass fields; the audit-script violations) but not a *persona* (no view, no voice, no dignity, no wariness).
|
||||
|
||||
The shape-vs-persona distinction also gives §13/§14/§15 a clean rubric:
|
||||
- **§13 (Genuinely Useful):** shape-anchored rules Manual Slop should adopt. Cluster 6 contributes the 3 useful caveats (provenance, user-as-navigator, no-collapse-to-yes/no).
|
||||
- **§14 (Anti-User Watchdog):** rules that constrain user output via persona. Cluster 6 contributes L140 (the stereotype rule as content-policy-via-persona).
|
||||
- **§15 (Persona Performance):** rules that project human categories onto the model. Cluster 6 contributes L136, L138, L146 (the framing, the symmetric closure, the sincerity rules).
|
||||
|
||||
The cluster 6 verdict is the *cleanest* example of the shape-vs-persona distinction in the entire Fable prompt: 4 of 6 lines are pure persona; 2 of 6 lines have useful caveats that map to shape-anchored Manual Slop rules. No other cluster has a 4-vs-2 ratio this lopsided.
|
||||
|
||||
---
|
||||
|
||||
**Sub-report complete.** This is the evidence base for §8 of `report.md`.
|
||||
Reference in New Issue
Block a user