research: cost-parity audit — Otto-65 real-billing addendum

docs/research/acehack-lfg-cost-parity-audit-2026-04-23.md

@@ -1,7 +1,7 @@
 # AceHack vs LFG cost-parity audit
 
 **Date:** 2026-04-23
-**Status:** first-pass audit; `admin:org` scope elevation authorized but not yet applied
+**Status:** first-pass audit; `admin:org` scope elevation authorized but not yet applied via `gh auth refresh`. The Otto-65 addendum below uses billing-UI screenshots pasted by the human maintainer, NOT scope-elevated agent reads.
 **Lives on:** AceHack (experimentation-frontier per Amara authority-axis split — Otto-61 memory)
 **Companion memory (per-user, pending in-repo migration):**
 `feedback_lfg_free_actions_credits_limited_acehack_is_poor_man_host_big_batches_to_lfg_not_one_for_one_2026_04_23.md`
@@ -14,8 +14,14 @@ for free that would limit us on LFG"*.
 ## TL;DR
 
 - **Linux Actions**: both repos public → both unlimited-free. Parity.
-- **macOS-14 runner**: already cost-aware — `gate.yml` matrix runs it
-  only on AceHack. Keep.
+- **macOS-14 runner**: corrected by Otto-65 addendum (see
+  "Correction to Otto-61 claim: macOS multiplier cost"
+  below). The original "matrix runs only on AceHack" claim
+  was wrong — `gate.yml` matrix runs macOS on both forks,
+  and LFG had macOS minutes on Apr 21 + 22. The
+  cost-discipline rationale is latency + future-policy /
+  fork-visibility-flip risk-headroom, not "matrix
+  avoidance."
 - **LFG monthly baseline cost** (confirmed by human-maintainer
   Otto-62, with follow-up Otto-62 correction *"i only used one user
   seat so only 19, maybe will update max later"*): Team plan
@@ -24,7 +30,7 @@ for free that would limit us on LFG"*.
   per-Actions spend.
 - **AceHack as user account**: exact Copilot-Pro status requires
   human-maintainer billing page (not exposed to the agent read-only
-  API). If Aaron holds Copilot Pro personally, AceHack inherits
+  API). If the human maintainer holds Copilot Pro personally, AceHack inherits
   Copilot PR reviews + Chat. If not, AceHack has no Copilot.
 - **Conclusion**: LFG is richer (Team plan + Copilot Business); AceHack
   is cheaper. Amara's authority-axis split (experiments → AceHack;
@@ -59,15 +65,26 @@ for free that would limit us on LFG"*.
 
 ### Actions runner cost-awareness (`gate.yml`)
 
+> **Errata 2026-04-28 (corrected by Otto-65 addendum
+> below):** the snapshot below was authored from a stale
+> read of `.github/workflows/gate.yml`. The current matrix
+> runs macOS on **both** forks; the conditional below does
+> NOT match what the file says today. The Otto-65 addendum
+> ("Correction to Otto-61 claim: macOS multiplier cost")
+> covers the corrected reality and the cost-discipline
+> rationale.
+
 ```yaml
 os: ${{ fromJSON(github.repository == 'Lucent-Financial-Group/Zeta'
         && '["ubuntu-22.04"]'
         || '["ubuntu-22.04","macos-14"]') }}
 ```
 
-Deliberate cost split: macOS-14 (10x multiplier even on public repos)
-runs on AceHack only. LFG is Linux-only. This predates the Otto-61
-directive chain; recognising it as already-correct.
+Original framing (now superseded): "deliberate cost split:
+macOS-14 (10x multiplier even on public repos) runs on
+AceHack only. LFG is Linux-only." Preserved verbatim for
+audit-trail traceability; the Otto-65 addendum is the
+canonical reference.
 
 ### Workflow run history (snapshot 2026-04-23)
 
@@ -179,12 +196,18 @@ the elevated scope to fill the unobservable fields.
    billing-side of this audit.
 3. **Enable dependabot_security_updates on AceHack** (free,
    increases parity). One-click through repo settings.
-4. **Document the LFG baseline $46/mo** in an ADR so future Otto
-   sessions can cost-account with numbers, not speculation.
+4. **Document the LFG baseline $27/mo** ($8 Team + $19 Copilot
+   Business × 1 seat) in an ADR so future Otto sessions can
+   cost-account with numbers, not speculation. Original draft
+   said `$46/mo` based on a pre-correction estimate; Otto-62's
+   correction *"i only used one user seat so only 19, maybe
+   will update max later"* drops the Copilot Business line by
+   ~$19, and the addendum's "Confirmed monthly baseline:
+   ~$27/mo" is the canonical figure.
 
 ### Long-term (if cost becomes binding)
 
-1. If LFG costs approach Aaron's budget ceiling, consider
+1. If LFG costs approach the human maintainer's budget ceiling, consider
    Copilot-only-on-AceHack mirror-PR workflow: author on AceHack
    (uses personal Copilot Pro if present), cherry-pick to LFG
    periodically. Preserves decision-canon on LFG while shifting
@@ -206,3 +229,220 @@ observations; this doc is the first in-repo Overlay-A mirror of
 that memory's findings. Future-session Otto with admin:org scope
 fills in the billing-side unobservables + lands a second-pass
 audit as an updated row under `docs/research/`.
+
+---
+
+## Second-pass corrections — Otto-65 real billing data
+
+Human maintainer 2026-04-23 Otto-65 pasted the actual GitHub
+billing UI for both accounts (LFG org + AceHack personal).
+This addendum supersedes the speculative figures above.
+
+### LFG (Lucent-Financial-Group) — April 2026 actuals
+
+**Subscription:** GitHub Team, $96/yr (= $8/mo) for 2
+licenses filled.
+
+**Metered usage:** $43.71 gross, $66.62 included-usage
+discount → net $0 billed on Actions for the month. The
+billed-flip claim ("gross exposure > discount → billed")
+holds **conditional on current org budget settings**: per
+"Org budgets" below, all products have `Stop usage: Yes`
+at $0 budget except GHAS and Copilot. So in practice the
+flip would NOT happen as written — gross-exceeding-
+discount triggers the budget hard-stop, not a billed
+amount. The text-as-written describes the underlying GitHub
+billing semantic (what *would* happen without the budget
+rail); the actual operational behaviour is hard-stop.
+
+**Copilot Business:** 1 license × $0.633/day ≈ $19/mo
+(billed, not discounted). Per-day reconciles exactly
+with Otto-62's $19-for-1-seat figure.
+
+**Top-5 repos this month:**
+
+| Repo | Gross |
+|---|---:|
+| Zeta | $41.72 |
+| lucent-infrastructure | $0.02 |
+| lucent-frontend | $0.02 |
+| lucent-user-service | $0.02 |
+| lucent-api-gateway | $0.02 |
+
+Zeta is the near-total consumer of LFG Actions gross.
+
+**Per-day breakdown (sample):**
+
+- Apr 21: 766 min Linux + 145 min macOS-3-core ($14.22 gross, $0.63 billed from Copilot only)
+- Apr 22: 2,133 min Linux + 196 min macOS-3-core ($25.58 gross, $0.63 billed from Copilot only)
+- Apr 23: 575 min Linux ($4.08 gross, $0.63 billed from Copilot only)
+
+> **Errata 2026-04-28:** the three sample-day gross
+> totals (`$14.22 + $25.58 + $4.08 = $43.88`) exceed the
+> declared monthly Actions gross of `$43.71` above by
+> `$0.17`. At least one of the per-day sample values is
+> rounded against a slightly different daily snapshot
+> than the monthly aggregate. The per-day breakdown
+> above is preserved verbatim from the original audit
+> snapshot for traceability; readers should treat the
+> monthly `$43.71` as the canonical figure and the
+> per-day samples as illustrative. Reconciling the
+> exact per-day → monthly delta requires re-pulling raw
+> billing data, which is logged as a follow-up task
+> rather than blocking this research-grade absorb.
+
+**Org budgets:** All products have `Stop usage: Yes`
+at $0 budget except GHAS and Copilot (which allow
+consumption). This is a hard safety rail — if discount
+ever fails to cover, the budget stops spend cold.
+
+**Confirmed monthly baseline:** **~$27/mo** =
+$8 Team + $19 Copilot Business × 1 seat. Matches
+Otto-62 estimate exactly.
+
+### AceHack (personal) — April 2026 actuals
+
+**Subscription:** GitHub Pro, $48/yr (= $4/mo), 3000
+Actions min/mo included.
+
+**Metered usage:** $50.45 gross, $51.21 discount → $0
+billed for the month.
+
+**Actions minutes used:** 1,773.7 / 3,000 included
+(59%). Does NOT count discounted public-repo usage;
+this is the AceHack personal quota.
+
+**Top repos this month:**
+
+| Repo | Gross |
+|---|---:|
+| Zeta | $36.44 |
+| Zeta (separate) | $13.77 |
+| devcontainer-codespace | varies |
+
+Two "Zeta" entries: the human-maintainer noted *"i think
+there was a little acehack before too, you can figure it
+out"* — suggests a prior fork / namespace is still
+generating billing rows. Archaeology pending.
+
+**Per-day breakdown (sample):**
+
+- Apr 19: 638 min Linux + 215 min macOS-3-core + Codespaces 4-core ($17.36 gross, $0 billed)
+- Apr 20: 444 min Linux + 91 min macOS-3-core ($8.32 gross, $0 billed)
+- Apr 21: 1,005 min Linux + 250 min macOS-3-core ($21.54 gross, $0 billed)
+
+**AceHack monthly baseline:** **$4/mo Pro** (flat). Plus
+whatever exceeds public-repo-discount-covered usage (so
+far: $0 billed despite $50+ gross).
+
+### Correction to Otto-61 claim: macOS multiplier cost
+
+Otto-61 memory said macOS runs incur 10x multiplier cost
+even on public repos. Actual April 2026 billing data
+shows **macOS-3-core at $0.062/min gross**, entirely
+covered by the public-repo discount.
+
+**Correction (errata to original draft):** the original
+draft of this section claimed `gate.yml` runs macOS only
+on AceHack, not LFG. That's incorrect — the LFG per-day
+breakdown above shows macOS-3-core minutes on Apr 21
+(145 min) and Apr 22 (196 min), and the actual `gate.yml`
+matrix runs macOS on both forks. The original
+"matrix avoidance" framing was wrong. The cost-discipline
+points still hold:
+
+- macOS seats are slower in wall-clock (2-10x slowdown)
+  even when cost-discounted — CI feedback latency is
+  its own resource.
+- Gross exposure stays absorbed by the public-repo
+  discount as long as both forks remain public; should
+  either fork flip to private OR upstream policy change
+  the discount terms, the gross exposure becomes the
+  billed cost.
+
+But the stark *"macOS is 10x expensive"* framing was
+too strong. Corrected: **macOS is 10x gross but
+currently 0x billed on public repos within the
+public-repo discount**. The reason to be careful with
+macOS minute-counts is latency + future-policy
+risk-headroom, not immediate cost.
+
+**Note on "quota" terminology:** earlier draft phrasing
+mixed "quota" between two different discount mechanisms
+(GitHub's per-account included-minutes pool vs the
+public-repo unlimited discount). For public repos the
+relevant mechanism is the public-repo discount, not a
+quota. Should the public-repo discount terms ever
+change OR a fork flip to private, that's where the
+billed-cost flip happens — not on a quota threshold per
+se. Erratum logged here to avoid future readers being
+misled.
+
+### Human maintainer's personal Copilot
+
+Confirmed via the human maintainer's Copilot settings
+page: *"You are assigned a seat as part of a GitHub
+Copilot Business subscription managed by servicetitan."*
+Personal Copilot is ServiceTitan-sponsored (employment
+benefit); **separate from LFG's Copilot Business seat**.
+
+Current-month personal premium-request usage: **84%**
+of monthly allotment. Approaching cap but not exceeded.
+Resets start of next month. This is the number that
+generalized to the "Frontier burn-rate UI" Otto-63
+directive — adopters on ServiceTitan-sponsored (or
+similarly-capped) subscriptions need the cap-awareness
+surface.
+
+### Answer to "does AceHack get anything free that would limit LFG?"
+
+**No.** Confirmed empirically:
+
+- AceHack's $50 gross Actions = fully discounted by
+  public-repo free tier
+- LFG's $43 gross Actions = fully discounted by
+  public-repo free tier
+- AceHack's Copilot = ServiceTitan-sponsored (free to
+  human maintainer) — does NOT cover LFG
+- LFG's Copilot Business = paid $19/mo — provides
+  Copilot PR reviews on LFG's PRs, which are
+  LFG-specific
+
+The two hosts have **parallel, independently-covered
+cost structures**. Neither subsidizes the other. Moving
+work between them is a purpose decision (Amara
+authority-axis) not a cost decision.
+
+### Updated BACKLOG candidates
+
+Retaining Otto-62 candidates + one new:
+
+1. Parity-audit tool — now with real-numbers fidelity from
+   the manually-pasted billing UI (Otto-65). The
+   `admin:org` scope elevation is still authorized but not
+   yet applied via `gh auth refresh`; once it is, the tool
+   can pull the same numbers programmatically. The
+   addendum's data is from the human maintainer's billing
+   UI paste, not scope-elevated agent reads.
+2. `admin:org`-scoped Actions-usage-history tool (depends on
+   the not-yet-applied scope elevation).
+3. Enable dependabot on AceHack (free, increases parity)
+4. ADR documenting the confirmed baseline: **LFG $27/mo, AceHack $4/mo flat; both $0 billed beyond baseline under current usage**
+5. **NEW: archaeology on the "separate Zeta" in AceHack billing** ($13.77 gross/mo suggests a prior fork still accumulating — could be moribund and should be archived, or could be intentional)
+
+### Updated "Otto-61 claim retractions"
+
+Otto-61 memory's *"AceHack is the poor-man host for per-
+PR work"* framing is **refined**, not retracted:
+
+- AceHack is cheaper in absolute terms ($4 vs $27/mo)
+- Both currently-$0-billed on Actions via public-repo
+  discount
+- Work placement remains Amara authority-axis driven,
+  not cost-driven
+- Budget caps at $0 on both are the safety rail; if
+  either tips to billed, the cap stops spend
+
+The Otto-62 final rule **stands as written**:
+> per-PR work on whichever substrate matches purpose
+> (experiments→AceHack, decisions→LFG), not cost-driven.