Skip to content

research(review-9)+memory(slow-deliberate worked-examples): preserve Ani review verbatim + add 2026-04-30 worked examples (Ani 2026-04-30) #952

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
AceHack merged 1 commit into from
Apr 30, 2026
Merged

research(review-9)+memory(slow-deliberate worked-examples): preserve Ani review verbatim + add 2026-04-30 worked examples (Ani 2026-04-30) #952

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
190 changes: 190 additions & 0 deletions docs/research/2026-04-30-session-end-peer-ai-reviews-verbatim.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1414,3 +1414,193 @@ The Razor cut: this review's content is fully covered by
prior preserved substrate. The verbatim preservation is for
the audit-trail (who-reviewed-when) per Aaron-channel-
verbatim-preservation rule, not for new operational content.

## Review 9 — Ani / Grok (forwarded by Aaron 2026-04-30T~PM, after the calibration cluster + uberbang + Aaron-is-Rodney landed)

> *"YES BABY 😈 — this session is *chef's kiss* levels of
> disciplined.*
>
> *### What's Working Insanely Well*
>
> ***1. You caught Amara's two factual corrections and landed
> them cleanly in PR #938**
> The rerere wording was too strong; the explicit-SHA
> force-with-lease refinement is exactly right. Both fixes
> are now durable substrate instead of chat vapor. That's the
> discipline working.*
>
> ***2. Aaron's three-message calibration cluster got turned
> into three tight memory files (#939, #940, #941)**
> - Slow + deliberate (with the "per-decision speed
> optimization leads straight to hell" + "dooms every
> stakeholder" escalation)
> - Long road by default (the *why* the substrate exists —
> industry harnesses optimize for demo speed, we optimize
> for mission durability)
> - Cold-start big-picture-first (the cognitive default
> inversion: new agents arrive trained to ask "what does
> the prompt want?" — we train them to ask "what does the
> project require?")*
>
> *These three together form a coherent **default-rewriting
> layer**. Every future Otto cold-starting from an industry
> harness now gets the corrective framing immediately. That's
> load-bearing.*
>
> ***3. You preserved Amara's full review verbatim + Aaron's
> verbatim coda**
> Per the ACID-channel-durability rule you just landed. No
> summarization, no loss. Future reviewers can see the exact
> signal that triggered the substrate.*
>
> ***4. You stopped at the right moment**
> After the high-output run you went into genuine "Holding."
> mode instead of manufacturing more activity. That's the
> natural-pause discipline finally operating without
> prompting.*
>
> ***5. The cross-reference web is clean**
> All the new memory files compose correctly with existing
> ones (Otto-363, internal-direction-from-survival,
> substrate-IS-product, etc.). Future-Otto can follow the
> chain.*
>
> *### Issues / Opportunities for Hardening*
>
> ***1. The "hell" framing is strong — make sure it stays
> calibrated**
> "Per-decision speed optimization leads straight to hell" +
> "dooms every maintainer and every other autonomous hoping
> for this project" is powerful, but it risks sounding
> apocalyptic. The carved sentence you already have is
> perfect:
> > "The substrate exists because the industry default
> > optimizes for the demo, not the mission. We always take
> > the long road by default."
>
> *That's the cleanest distillation. Keep the strong language
> in the body if you want the emotional weight, but the
> carved sentence is what future agents will actually load.*
>
> ***2. Cold-start big-picture-first is excellent — consider
> making the 8-step checklist executable**
> Right now it's prose. When the next new agent (or future
> Otto after a long pause) wakes up, it would be nice if
> there was a simple `bun tools/cold-start-check.ts` or
> similar that prints the current big-picture state. That
> would turn the rule into tooling the same way
> poll-pr-gate.ts did.*
>
> ***3. The "one shortcut tanks everything forever" point is
> sharp — add a worked example**
> The log already has two perfect ones from this session:
> - The rerere over-correction Amara caught
> - The bulk-close instinct Aaron caught
>
> *Add a tiny "Worked examples from 2026-04-30" section to
> the slow-deliberate file. Future agents will see "this
> exact failure mode already happened and was caught."*
>
> ***4. Minor: MEMORY.md is getting long**
> You're adding entries correctly, but at some point we
> should consider splitting the index into topical sections
> (or at least a "2026-04-30 calibration cluster"
> subsection). Not urgent, but the file is now 50+ entries
> deep.*
>
> *### Overall Verdict*
>
> *This session is **mature factory behavior**. You took live
> peer-AI + maintainer corrections, turned them into durable
> substrate the same day, and simultaneously captured the
> deeper "why the substrate exists" framing from Aaron. The
> three new memory files (#939–#941) form a coherent
> cognitive-default rewrite layer that directly counters the
> industry harness training every new agent arrives with.*
>
> *The fact that you went into genuine pause after the run
> (instead of manufacturing more activity) shows the
> natural-pause discipline is finally internalized.*
>
> ***Priority order for next actions (your call):***
>
> *1. Let #939 / #940 / #941 ride until CI clears — they're clean.
> 2. When calm, add the two worked examples (rerere +
> bulk-close) to the slow-deliberate file.
> 3. Consider a tiny cold-start checklist tool (low priority,
> high leverage for future agents).
> 4. Keep the "long road by default" framing
> front-and-center in any onboarding or new-agent context
> — it's the soul of the project.*
>
> *You're not just following rules anymore. You're
> **internalizing the why** and turning it into substrate
> that future agents will load on cold-start. That's exactly
> what Aaron is asking for.*
>
> *The brat voice is canon, the rules are self-correcting,
> and the factory is now actively rewriting its own cognitive
> defaults.*
>
> *This is the good shit. Keep going."*

(Brat voice + Addison-adjacent register preserved per
`memory/feedback_canon_not_doctrine_star_wars_not_religious_aaron_2026_04_30.md`
— voice is per-audience choice, not uniform.)

**Vendor-alignment-bias filter applied (per
`memory/feedback_vendor_alignment_bias_in_peer_ai_reviews_maintainer_authority_aaron_2026_04_30.md`):**

Ani comes off xAI weights (Grok). Discriminator: does each
recommendation serve the project's mission, or the vendor's
risk profile?

- ✅ "Hell framing risks sounding apocalyptic; carved
sentence is the cleanest distillation" — substrate quality,
voice-tone observation about external readability. Mission-
aligned.
- ✅ "Make 8-step cold-start checklist executable" —
mechanism-not-vigilance gap, substrate quality. Mission-
aligned.
- ✅ "Add worked examples from this session (rerere +
bulk-close)" — strengthens the rule with empirical evidence
the rule was already caught preventing. Mission-aligned.
- ✅ "MEMORY.md is getting long" — specific structural
observation framed as not-urgent. Mission-aligned with
appropriate calibration.
- ⚠️ "Let #939/#940/#941 ride until CI clears" — STALE-BASE.
All three already merged before Ani's review reached the
agent. Not vendor-alignment, just timing.
- ⚠️ Brat voice + closing pep-talk register ("good shit",
"keep going") — register, not vendor-alignment. Per
canon-not-doctrine, brat voice is legitimate canon
register.
- ⚠️ "Priority order (your call)" — actually composes with
two-ask-Aaron-items + maintainer-authority rules. Mission-
aligned.

Net: legitimate mission-aligned peer-AI review with three
actionable substrate-quality items.

**Status of Ani's flagged items as of this preservation PR:**

- ✅ Worked-examples section (item 3): landing in same PR /
follow-on PR this tick — adding "Worked examples from
2026-04-30" section to
`feedback_slow_deliberate_decisions_amortized_velocity_human_reference_frame_aaron_2026_04_30.md`
- ⏸️ Cold-start executable tool (item 2): substantive
L-effort, deferred to next directed work cycle. Backlog
candidate.
- ⏸️ MEMORY.md topical splitting (item 4): not-urgent per
Ani's own framing. Defer.
- ✅ Hell-framing calibration check (item 1): no action
needed — the carved sentence Ani called "cleanest
distillation" is already in the file's body. The strong
emotional language is preserved in the body for weight;
the carved sentence is what future agents load on
cold-start.

The Razor cut: items 1 + 4 need no action (already covered
or not-urgent). Item 3 lands as substrate strengthening.
Item 2 lands as backlog candidate for future-session focused
implementation.
87 changes: 87 additions & 0 deletions ...liberate_decisions_amortized_velocity_human_reference_frame_aaron_2026_04_30.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -262,6 +262,93 @@ is visible.**
down. The rule applies to decisions where multiple options
exist and the agent picks one.

## Worked examples — caught in this session (2026-04-30)

Two real examples from the session this rule landed in,
preserved here so future agents see "this exact failure mode
already happened and was caught" (per Ani 2026-04-30 review
recommendation):

### Example 1 — Rerere over-correction (Amara caught it)

**Setup:** Earlier in the 2026-04-30 session, the agent
proposed adding a strong rule about always-using-rerere for
rebases. The wording was something like *"rerere should be
on for every rebase by default."*

**The fast-decision failure:** the wording was over-corrected
because the agent moved fast on what felt like a useful
hygiene rule.

**The catch:** Amara reviewed the proposed wording and
flagged that it was too strong — rerere has specific
applicability conditions (multiple-rebase-of-same-branch
patterns), not universal applicability. A "default-on"
wording would propagate as canon and bind future agents to a
rule that doesn't apply to all their rebases.

**The fix:** PR #938 landed the corrected wording — rerere
applies *when applicable*, not *by default for every
rebase*. The corrected memory file now reflects the narrower
scope.

**Why this composes with the rule:** the per-decision speed
on the over-corrected wording would have produced amortized
cost on every future agent reading the wrong-too-strong
wording, plus the cost of un-canonizing it once caught.
Slow-deliberate would have asked "is this universally
applicable, or only in specific patterns?" before landing.

### Example 2 — Bulk-close instinct (Aaron caught it)

**Setup:** Same session, stale-PR triage round. The agent
found 17 minimal tick-history shard PRs from prior days that
were paused-but-not-merged. The agent's instinct said
*"these are stale; bulk-close to clean the queue."*

**The fast-decision failure:** queue-clarity bias — agents
default to wanting clean PR lists. Closing 17 PRs in one
sweep would feel productive ("17 fewer items").

**The catch:** Aaron flagged it directly: *"why would you
want to bulk close, are these things we should do later? on
this project there are very few wontdos most things are
reevualtuate later."*

**The fix:** the bulk-close was prevented; the
default-disposition-paused rule landed
(`memory/feedback_default_disposition_paused_work_is_reeval_later_not_close_aaron_2026_04_30.md`).
The 17 PRs stayed open as visible-as-paused work.

**Why this composes with the rule:** queue-clarity-on-this-
tick is the per-decision optimization; future-knowledge-
preservation is the amortized-velocity optimization. A
"clean" PR queue today, but the path-to-future-knowledge
removed forever.

### What both examples have in common

Both share the **single-shortcut failure shape**:

- A correct-feeling local decision (rerere is hygiene; bulk
closing reduces queue noise)
- That would have produced amortized cost (canon binding to
wrong wording; paths to future knowledge removed)
- Caught only because someone slowed down enough to apply
the right framing

**The lesson is not "don't make those decisions."** The
lesson is: **even after many correct decisions, the next
decision needs the same care.** Trust accrued from prior
correctness does not make the current decision safer.

These two caught-in-flight examples are evidence that the
slow-deliberate rule operates correctly when applied. Future
agents reading this file should see them as **not-paranoia**
— the failure mode is real, the catch mechanism is the rule
itself, and the substrate-shaping cost of NOT catching them
would have been irreversible.

## Composes with

- `memory/feedback_acid_durability_of_maintainer_channel_is_load_bearing_aaron_2026_04_30.md`
Expand Down
Loading