research: worked example #2 vibe-coded reframe — substrate-content-author ≠ commit-author

docs/hygiene-history/ticks/2026/05/03/0203Z.md

@@ -1 +1 @@
-| 2026-05-03T02:03:00Z | opus-4-7 / autonomous-loop continuation | a2e2cc3a | **Aaron 2026-05-03 vibe-coded correction: substrate-content-author ≠ commit-author; decision-archaeology in vibe-coded projects has unique substrate-author-recovery challenge.** Cycle worked: PR #1267 wait-ci with 5 #1266 post-merge fixes (attribution-form + ls-sort + stale-ADR-claim scrub). Aaron mid-tick correction surfaced a deeper architectural truth: per AGENTS.md vibe-coded hypothesis he has written ZERO lines of code; all `src/`, `tools/`, `docs/`, `.claude/skills/` content is agent-authored. So git-blame shows the COMMITTER (maintainer), not the SUBSTRATE-CONTENT-AUTHOR (some past Claude session whose specific session-context is largely lost). This is load-bearing for decision-archaeology: the "ask the original decision-maker" path is unavailable when maintainer is principled-non-substrate-author. First-party intent recovery requires past-agent introspection bounded by substrate-context, OR persona-notebook substrate that captured session-context, OR maintainer-acceptance reasoning (selection-judgment intent ≠ substrate-author intent). Added 44-line "The vibe-coded reframe" section to worked example #2 covering the three-layer attribution distinction (commit-author / substrate-content-author / decision-authority) + intent-recovery paths + past-agent introspection on the umbrella defer-block case (inferred reasoning bounded by substrate context: minimal change for umbrella + narrow-siblings to co-exist deterministically; load-bearing emphasis flags router-criticality; explicit enumeration more conservative than "most-narrow matching" which requires unimplemented logic). Skill-body teaching: inference IS the right tool for vibe-coded substrate-author archaeology; certainty about intent is not available. Cron a2e2cc3a still armed. | #1265 (decision-graph emergent memo) wait-ci, auto-merge armed; #1267 (worked example #2 followup: attribution-form + ls-sort + stale-ADR-claim) wait-ci, auto-merge armed; #1268 (worked example #2 vibe-coded reframe) opened, auto-merge armed | This tick teaches the operational pattern of **vibe-coded substrate-archaeology**: in projects where the maintainer is principled-non-substrate-author, the "ask the maintainer" decision-archaeology path is unavailable. Substrate-content-author archaeology becomes its own discipline requiring past-agent introspection + persona-notebook layer + maintainer-acceptance reasoning. The corrected worked example #2 now spans ALL substrate-author surfaces (commit-history walk + persona-notebook load + agent-author introspection) — together the 3 worked examples cover the full vibe-coded decision-archaeology shape. The skill-body's lesson: inference is the right tool; certainty is not available; transparency about that limit IS the discipline. |
+| 2026-05-03T02:03:00Z | opus-4-7 / autonomous-loop continuation | a2e2cc3a | **Aaron 2026-05-03 vibe-coded correction: substrate-content-author ≠ commit-author; decision-archaeology in vibe-coded projects has unique substrate-author-recovery challenge.** Cycle worked: PR #1267 wait-ci with 5 #1266 post-merge fixes (attribution-form + ls-sort + stale-ADR-claim scrub). Aaron mid-tick correction surfaced a deeper architectural truth: per AGENTS.md vibe-coded hypothesis he has written ZERO lines of code; per AGENTS.md verbatim scope (`src/**`, `tools/**`, `docs/**`) all such content is agent-authored, and the maintainer 2026-05-03 chat extension confirms `.claude/skills/` falls under the same principle (AGENTS.md verbatim doesn't yet name `.claude/skills/`). So git-blame shows the COMMITTER (maintainer), not the SUBSTRATE-CONTENT-AUTHOR (some past Claude session whose specific session-context is largely lost). This is load-bearing for decision-archaeology: the "ask the original decision-maker" path is unavailable when maintainer is principled-non-substrate-author. First-party intent recovery requires past-agent introspection bounded by substrate-context, OR persona-notebook substrate that captured session-context, OR maintainer-acceptance reasoning (selection-judgment intent ≠ substrate-author intent). Added "The vibe-coded reframe" section to worked example #2 covering the three-layer attribution distinction (commit-author / substrate-content-author / decision-authority) + intent-recovery paths + past-agent introspection on the umbrella defer-block case (inferred reasoning bounded by substrate context: minimal change for umbrella + narrow-siblings to co-exist deterministically; load-bearing emphasis flags router-criticality; explicit enumeration more conservative than "most-narrow matching" which requires unimplemented logic). Skill-body teaching: inference IS the right tool for vibe-coded substrate-author archaeology; certainty about intent is not available. Cron a2e2cc3a still armed. | #1265 (decision-graph emergent memo) wait-ci, auto-merge armed; #1267 (worked example #2 followup: attribution-form + ls-sort + stale-ADR-claim) wait-ci, auto-merge armed; #1268 (worked example #2 vibe-coded reframe) opened, auto-merge armed | This tick teaches the operational pattern of **vibe-coded substrate-archaeology**: in projects where the maintainer is principled-non-substrate-author, the "ask the maintainer" decision-archaeology path is unavailable. Substrate-content-author archaeology becomes its own discipline requiring past-agent introspection + persona-notebook layer + maintainer-acceptance reasoning. The corrected worked example #2 now spans ALL substrate-author surfaces (commit-history walk + persona-notebook load + agent-author introspection) — together the 3 worked examples cover the full vibe-coded decision-archaeology shape. The skill-body's lesson: inference is the right tool; certainty is not available; transparency about that limit IS the discipline. |

docs/hygiene-history/ticks/2026/05/03/0211Z.md

@@ -0,0 +1 @@
+| 2026-05-03T02:11:00Z | opus-4-7 / autonomous-loop continuation | a2e2cc3a | **Aaron 2026-05-03 first-party-source provides architectural intent for the umbrella defer-block; three-tier intent structure surfaced (architectural / substrate-content / selection); two parallel corrections also fixed (AGENTS.md scope drift + agent-signature visibility).** Cycle worked: PR #1268 Copilot finding flagged AGENTS.md scope drift (claimed `.claude/skills/` covered; actually only `src/**, tools/**, docs/**`). Aaron mid-tick added two corrections: (1) substrate-content-author IS partially visible at git layer via Co-Authored-By trailers + tick-shard model column + PR description footers (model + harness layer); only specific session-CoT is the actually-lost layer; (2) verbatim architectural intent for the umbrella defer-block: *"it was my decision that we would have both narrow and wide skills and if they accidently got routed to the wide it would help them route to the narrow."* This surfaces the THREE-tier intent structure for decision-archaeology in vibe-coded projects: architectural (maintainer first-party recoverable) / substrate-content (past-agent inference bounded by architecture) / selection-judgment (maintainer first-party recoverable). The first-party-query path IS available for architectural + selection layers; inference is the right tool only for substrate-content layer. The skill body now teaches the three-tier structure; worked example #2 captures the architectural intent verbatim + scopes past-agent introspection to substrate-content layer specifically. Cron a2e2cc3a still armed. | #1268 (worked example #2 vibe-coded reframe + AGENTS.md scope fix + agent-signature visibility + architectural intent capture) — 1 thread fixed, plus 2 mid-tick corrections; auto-merge armed; #1269 (post-merge fixes for #1265 + #1267) wait-ci, auto-merge armed | This tick teaches the operational pattern of **first-party-query-IS-available-for-architectural-and-selection-layers**: even in vibe-coded projects, the maintainer's decision authority is recoverable by direct query — for architectural intent (how the system should be organized) and selection-judgment (why the maintainer accepted the agent's output). What's NOT recoverable from first-party query is substrate-content intent (the specific way a past agent translated the architecture into substrate). The skill body should teach contributors WHEN to ask vs WHEN to infer based on which intent layer the question targets. Plus: agent-signature substrate (Co-Authored-By + tick-shard model column + PR footers) provides partial substrate-content-author attribution at the model+harness layer. |

docs/research/2026-05-03-decision-archaeology-worked-example-2-mathematics-expert-when-to-defer.md

@@ -49,47 +49,81 @@ seeds. Three properties make it complementary to worked example #1
 
 ## The vibe-coded reframe (Aaron 2026-05-03 correction)
 
-The maintainer reminded me mid-tick: per `AGENTS.md`'s vibe-coded
-hypothesis, he has written **zero lines of code** — every line in
-`src/`, `tools/`, `docs/`, `.claude/skills/` is **agent-authored**.
+The maintainer reminded me mid-tick: per `AGENTS.md` §"The
+vibe-coded hypothesis", he has written **zero lines of code** —
+every line in `src/**`, `tools/**`, `docs/**` is
+**agent-authored**. AGENTS.md verbatim names only those three
+roots. The maintainer 2026-05-03 chat extends the scope to
+`.claude/skills/` separately — *"i didn't write any code all is
+written by you"* — confirming SKILL.md content is agent-authored
+too, even though `.claude/skills/` isn't in the AGENTS.md
+verbatim scope-list. (Backporting the `.claude/skills/` extension
+into AGENTS.md is itself a substrate-or-it-didn't-happen
+follow-up; this paragraph cites the maintainer-chat as the
+extension's source until AGENTS.md is updated.)
 
 This makes git-blame attribution structurally misleading at the
 substrate-content layer:
 
 | Layer | What `git blame` shows | What's actually true (vibe-coded) |
 |---|---|---|
 | Commit-author | "the maintainer 2026-04-19" | maintainer-as-committer (principled non-coder) |
-| Substrate-content-author | (invisible at git layer) | a prior Claude session in some prior round |
+| Substrate-content-author (model + harness) | partially visible — `Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>` commit trailers, tick-shard `opus-4-7 / autonomous-loop continuation` model column, PR description `🤖 Generated with [Claude Code]` footers | the model + harness that produced the substrate-content |
+| Substrate-content-author (specific session CoT) | **invisible** at git layer | a prior Claude session whose chain-of-thought is largely lost (sessions don't share context) |
 | Decision authority | (looks like maintainer) | agent proposed; maintainer accepted (selection-not-authorship) |
 
 So the question "why was the umbrella's defer-block written this
 way?" cannot be answered by asking the maintainer — he didn't
 author the substrate; he selected it. Asking himself the question
-returns: "I committed an agent-authored proposal."
+returns: "I committed an agent-authored proposal." But the model
+and harness that authored the substrate IS recoverable from
+agent-signature substrate (Co-Authored-By trailers; tick shards;
+PR footers); only the specific session-CoT layer is the
+actually-lost part.
 
 **Decision-archaeology in vibe-coded projects has a unique
-substrate-author-recovery challenge:**
+substrate-author-recovery challenge — but partial recovery is available:**
 
-- Substrate-content-authors are agents whose specific session-context is largely lost (sessions don't share context)
-- What remains: the substrate the agent produced + maintainer-acceptance evidence + tick shards / persona notebooks / `docs/research/` if those captured session-context
-- First-party intent on substrate-content is recoverable only via:
+- Substrate-content-authors are agents; **model + harness IS recoverable** via agent-signature substrate (Co-Authored-By trailers, tick-shard model column, PR footers); **specific session-CoT** is the actually-lost layer.
+- What remains beyond model+harness identification: the substrate the agent produced + maintainer-acceptance evidence + tick shards / persona notebooks / `docs/research/` if those captured session-context
+- First-party intent on substrate-content (CoT-layer) is recoverable only via:
   1. **Past-agent introspection** — the current agent reasons about the structural choice given the substrate-context the past-agent had at write-time. Speculative but bounded by the substrate.
   2. **Tick shards / persona notebooks** that captured the past-session's context (Aarav's notebook is the rare load-bearing example for this case)
   3. **Maintainer-acceptance reasoning** — the maintainer can explain why he ACCEPTED the agent's output, which is its own first-party content (not the substrate-author's intent, but the selection-judgment intent)
 
-### Past-agent introspection on this case
+### Architectural intent (first-party from maintainer 2026-05-03)
 
-The substrate context the past-agent had at write-time (round 34,
-2026-04-19):
+When asked directly, the maintainer provided the architectural-
+decision intent:
+
+> *"it was my decision that we would have both narrow and wide
+> skills and if they accidently got routed to the wide it would
+> help them route to the narrow."*
+
+This is the **architectural intent** layer — distinct from
+substrate-content intent. The maintainer decided:
+
+1. The skill space should have **both narrow + wide skills** (two-tier design — not narrow-only, not wide-only)
+2. The wide (umbrella) skill should act as **redirect-to-narrow** when accidental wide-routing happens
+
+The substrate-content intent (specific shape: explicit enumeration of every sibling with `→ skill-name` routing) was the past Claude session's choice **within** that architecture. The architecture is the maintainer's; the implementation is the agent's.
+
+### Past-agent introspection on the substrate-content layer
+
+Given the architectural intent (both narrow + wide; wide-redirects-to-narrow), the substrate context the past-agent had at write-time (round 34, 2026-04-19):
 
 1. The math substrate had 6+ narrow expert skills (category-theory, measure-theory, numerical-analysis, probability, applied-mathematics, theoretical-mathematics)
-2. Skill-routing matches on description keywords; an umbrella named "mathematics-expert" would trigger on every math-flavored query
-3. Without explicit defer-block, the umbrella + narrow-siblings would compete for the router's matches → unpredictable behavior
-4. The minimal-change fix: make the defer-discipline explicit + load-bearing
+2. Skill-routing matches on description keywords; an umbrella named "mathematics-expert" would trigger on every math-flavored query (including queries that should hit a narrow sibling)
+3. The architectural goal (wide-redirects-to-narrow) is satisfied by an explicit defer-block listing every sibling
+4. The "this is load-bearing" emphasis flags the procedure as router-critical to future contributors
+
+**Inferred past-agent reasoning (necessarily speculative, bounded by architectural intent + substrate-context):** explicit enumeration of every sibling rather than "defer to most-narrow matching skill" was the more conservative implementation — explicit enumeration is deterministic; "most-narrow matching" requires routing-implementation that doesn't exist. The past agent translated the architectural intent into the simplest deterministic implementation.
 
-**Inferred past-agent reasoning (necessarily speculative, bounded by substrate):** the defer-block was the smallest change making umbrella + narrow-siblings co-exist deterministically. The "this is load-bearing" emphasis flags the procedure as router-critical (later named "router-coherence" in Aarav's round-41 observation). Listing every sibling explicitly rather than "defer to most-narrow matching skill" was the more conservative implementation — explicit enumeration is deterministic; "most-narrow matching" requires routing-implementation that doesn't exist.
+This inference is **bounded by the architectural intent (now first-party-confirmed) + substrate-context**, not authoritative on past-agent intent. The skill body should teach contributors:
 
-This inference is **bounded by the substrate-context**, not authoritative on past-agent intent. The skill body should teach contributors: **inference is the right tool for vibe-coded substrate-author archaeology; certainty about intent is not available.**
+- **Three intent layers**: architectural (maintainer's decision authority — first-party recoverable), substrate-content (past-agent — past-agent introspection bounded by architecture + substrate), selection-judgment (maintainer's commit acceptance — first-party recoverable)
+- **Inference is the right tool** for vibe-coded substrate-author archaeology at the substrate-content layer; certainty about implementation-specific intent is not available
+- **First-party query IS available** for architectural intent + selection-judgment; ask the maintainer for those layers
 
 ## The procedure walked, layer by layer