xai-org · FalconOrtiz · Sep 8, 2026
diff --git a/docs/spacexai-ire-2026-09-08/ENG_GREEN.md b/docs/spacexai-ire-2026-09-08/ENG_GREEN.md
@@ -0,0 +1,24 @@
+# ENG GREEN — SpaceXAI merge-ready pack
+**Reviewer:** Head of Engineering (IRE) · **2026-09-08**  
+**Base:** `902a06f` · **Patch:** `0001-spacexai-f1-f5-a-d.patch`  
+**External send:** still **HOLD** until Falcon go
+
+## Verdict
+**ENG GREEN** to land / propose. Acceptance criteria met.
+
+| # | Criterion | Result |
+|---|-----------|--------|
+| 1 | PR-A FS only; defaults 4/0/20 | PASS — comment + T1–T3 arms; defaults unchanged |
+| 2 | PR-F1 prefer-original FS-gated default false | PASS — first-wins off; Following parity on |
+| 3 | PR-F5 Enable=false / K=1; IN-only; source_key RT→quoted→tweet_id; ≤1/slate; after PreviouslyServed; no Age/OON/VF | PASS |
+| 4 | PR-D classify hooks only | PASS — `serve_card_classify.rs` |
+| 5 | Out list untouched | PASS |
+| 6 | Tests green + EXP± | PASS — standalone **22/22** (re-run by HoE); Python smoke **22/22** |
+
+## Non-blockers (treatment / monorepo follow-ups)
+1. **F5 v1 scopes to PreviouslyServed only** — `PreviouslySeen*` still drops impressed related ids. Correct blast radius for landable default-off; treatment arms may need a measured seen-policy later (not open F2b).  
+2. **D hooks fire on F5 keep path** — full serve-mix baseline when F5 off still needs product Under-the-Hood / monorepo counter wire.  
+3. **OSS cannot run full home-mixer cargo** — private `xai_*`; standalone crate is the required green bar here.
+
+## Go
+Falcon go → send from GitHub `falconortiz` with patch + mixed cover. Until then: HOLD.
diff --git a/docs/spacexai-ire-2026-09-08/PATCHES.md b/docs/spacexai-ire-2026-09-08/PATCHES.md
@@ -0,0 +1,43 @@
+# Touched files — SpaceXAI pack @ 902a06f
+
+## Modified
+
+| File | Pack |
+|------|------|
+| `home-mixer/params/param.rs` | A (comments), F1 params, F5 params |
+| `home-mixer/filters/retweet_deduplication_filter.rs` | F1 |
+| `home-mixer/filters/previously_served_posts_filter.rs` | F5 (defer when armed) |
+| `home-mixer/filters/mod.rs` | F5 module export |
+| `home-mixer/candidate_pipeline/phoenix_candidate_pipeline.rs` | F5 wire |
+| `home-mixer/util/candidates_util.rs` | F5 `source_key` |
+| `home-mixer/util/mod.rs` | D module export |
+
+## Added
+
+| File | Pack |
+|------|------|
+| `home-mixer/filters/controlled_source_reexposure_filter.rs` | F5 |
+| `home-mixer/util/serve_card_classify.rs` | D |
+
+## Standalone test crate (cache, not in x-algorithm)
+
+| Path | Role |
+|------|------|
+| `x-algo-cache/merge-ready/spacexai-pack-tests/` | Pure F1/F5/D `cargo test` |
+| `x-algo-cache/merge-ready/STATUS.md` | This status |
+| `x-algo-cache/merge-ready/PATCHES.md` | This list |
+| `x-algo-cache/merge-ready/0001-spacexai-f1-f5-a-d.patch` | `git diff` export |
+
+## Params added (defaults = today's behavior)
+
+- `EnablePreferOriginalRetweetDedup`: bool = **false**
+- `EnableControlledSourceReexposure`: bool = **false**
+- `ControlledSourceReexposureMaxServesK`: u32 = **1**
+- `ControlledSourceReexposureMinRequestGap`: u32 = **0**
+- `ControlledSourceReexposureInNetworkOnly`: bool = **true**
+
+## Params NOT changed (PR-A)
+
+- `FollowAuthorWeight` = 4.0  
+- `ProfileClickWeight` = 0.0  
+- `ShareViaCopyLinkWeight` = 20.0  
diff --git a/docs/spacexai-ire-2026-09-08/RESULTS.md b/docs/spacexai-ire-2026-09-08/RESULTS.md
@@ -0,0 +1,25 @@
+# e2e_smoke_f5 RESULTS — 2026-09-08
+**Command:** `python3 smoke_controlled_reexposure.py`  
+**Outcome:** **22 PASS / 0 FAIL**
+
+## EXP+ (niche re-exposure)
+| Check | Result |
+|-------|--------|
+| Current: first serve original | PASS |
+| Current: 2nd request burned (related/source) | PASS |
+| Current: same-slate 3 RTs → 1 | PASS |
+| F5: 2nd/3rd IN nudge within K=3 | PASS |
+| F5: 4th blocked at K | PASS |
+| Attribution = original source_id | PASS |
+
+## EXP− (snowball)
+| Check | Result |
+|-------|--------|
+| Current 20 RTs → 1 | PASS |
+| Current 20 quotes survive RT-dedup | PASS (risk → need quote cap in F5) |
+| F5 ≤1/slate + K=3 across requests | PASS |
+| F5 rejects OON RT path | PASS |
+| Following prefer-native / For You RT-first gap (F1) | PASS |
+
+## Viability
+F5 **not in tree** — implement as FS A/B on top of existing filters. Building blocks all present @ `902a06f`.
diff --git a/docs/spacexai-ire-2026-09-08/SEND_GO.md b/docs/spacexai-ire-2026-09-08/SEND_GO.md
@@ -0,0 +1,5 @@
+# SEND GO
+**Falcon GO:** 2026-09-08 Europe/Madrid
+**Eng:** GREEN
+**Account:** falconortiz
+**Base:** 902a06f
diff --git a/docs/spacexai-ire-2026-09-08/SPACEXAI_COVER_MIXED_FINAL.md b/docs/spacexai-ire-2026-09-08/SPACEXAI_COVER_MIXED_FINAL.md
@@ -0,0 +1,33 @@
+# SpaceXAI mixed request — creator reach equilibrium (A + D + F1 + F5)
+**From:** Falcon Ortiz (@falconortiz) / IRE · GitHub send: **falconortiz**  
+**Code:** https://github.com/xai-org/x-algorithm @ `902a06f`  
+**Pattern:** `docs/BIDIRECTIONAL_BOOST_CHANGE.md` (FS → A/B → ship)  
+**SpaceXAI owns merge.** IRE proposes only.
+
+## Plain-language ask
+For You maximizes `Σ w·P(action)` with FollowAuthor=4 vs CopyLink=20 and ProfileClick=0. Hard rules (48h age, RT XOR original, source burn, quotes bypass RT-dedup) cut prolonged niche re-exposure. Creator evidence: follow/imp **0.00636%** vs Batch B **0.0644%**; OCR HT **49.1K/500K** at η **3.78%**.
+
+We request a **mixed program**:
+1. **Study/telemetry (PR-D)** — serve mix original vs RT vs quote; re-serve≥2×; negatives if available  
+2. **Immediate A/Bs** — **PR-A** weight rebalance; **PR-F1** prefer-original same-slate (Following parity)  
+3. **PR-F5 ControlledSourceReexposure** — implement behind FS **default off** (K=2–3, IN-only, gap, ≤1/slate, **quotes share source cap with RTs**); enable after D baseline  
+
+**Not in this ask:** open burn-off (F2b), post-48h Age rewrite (F4), OCR ranking head (E), VF SpamHighRecall loosen.
+
+## Evidence (empirical, not SLOs)
+| Signal | Value |
+|--------|------:|
+| Cohort follow/imp | 0.00636% (79/1.24M) |
+| Batch B | 0.0644% |
+| High-amp vs low-amp strata | 0.0678% vs 0.0053% |
+| OCR | 180/500 VF · 49.1K/500K HT |
+| F5 smoke | 22/22 PASS (sim @ filters) |
+
+## Success
+↑ follow/imp on non-viral OON; ↑ original serve share after amplify; F5 arms show re-serve 2–3 with no report/mute spike; no quote/RT snowball.
+
+## Attachments for eng
+- `SPACEXAI_PARAM_SKETCH_PR_A.md`  
+- `SPACEXAI_PR_D_TELEMETRY_NOTE.md`  
+- `SPACEXAI_F5_IMPLEMENTATION.md`  
+- `SPACEXAI_MIXED_PR_F5.md` · `e2e_smoke_f5/`  
diff --git a/docs/spacexai-ire-2026-09-08/SPACEXAI_F5_IMPLEMENTATION.md b/docs/spacexai-ire-2026-09-08/SPACEXAI_F5_IMPLEMENTATION.md
@@ -0,0 +1,43 @@
+# PR-F5 — ControlledSourceReexposure implementation note
+**Status:** Landable with mixed PR as FS + filter sketch · treatment default **off**  
+**Depends on:** PR-D counters before enabling K>1 arms  
+**Companions:** `SPACEXAI_MIXED_PR_F5.md`, `e2e_smoke_f5/` (22/22 PASS sim), HoE quote-cap harden  
+**Repo:** x-algorithm @ `902a06f` · SpaceXAI owns merge
+
+## Algorithm (per viewer request)
+```
+slate = candidates after hydration
+slate = RetweetDedup + prefer-original when F1 on
+slate = drop OON RTs (existing OONRetweetReplyFilter — unchanged)
+for each candidate:
+  source = source_key(candidate)
+  if source in served_or_impressed_related:
+    if not EnableControlledSourceReexposure: drop   # today
+    elif serves[source] >= K: drop
+    elif requests_since_last[source] < Gap: drop
+    elif InNetworkOnly and amplifier not IN: drop
+    else: keep as re-exposure card
+  else:
+    keep  # first exposure
+# Prefer original over amp when both re-admitted for same source
+enforce ≤1 card per source in slate
+log serve(source, card_type ∈ {original, rt, quote})
+```
+
+## source_key (quotes REQUIRED)
+```
+if retweeted_tweet_id: return retweeted_tweet_id
+if quoted_tweet_id:    return quoted_tweet_id
+return tweet_id
+```
+RTs and quotes of the same original share one K bucket. Do not ship a false arm on IncludeQuotes.
+
+## Rollout
+1. Land params + logging (D) with **Enable=false** and/or **K=1** (behavior ≡ today)  
+2. Shadow: compute would-allow rate from serve mix  
+3. After baseline: A/B **K=2** vs **K=3**, small %; InNetworkOnly=true  
+4. Guardrails: report / mute / block; duplicate-fatigue proxies  
+5. Ship winning K or keep off  
+
+## Out of scope
+AgeFilter change · OON RT enable · VF / SpamHighRecall loosen · OCR ranking head · open F2b burn-off · F4 post-48h prolong
diff --git a/docs/spacexai-ire-2026-09-08/SPACEXAI_MIXED_PR_F5.md b/docs/spacexai-ire-2026-09-08/SPACEXAI_MIXED_PR_F5.md
@@ -0,0 +1,109 @@
+# Mixed SpaceXAI request — study + solutions + implementable F5
+**Account evidence:** @falconortiz Studio + API 2026-09-08  
+**Code:** https://github.com/xai-org/x-algorithm @ `902a06f`  
+**Sender when approved:** GitHub **falconortiz** · SpaceXAI owns merge  
+**Pattern:** `docs/BIDIRECTIONAL_BOOST_CHANGE.md` (FS → A/B → ship)  
+**Status:** HOLD until Falcon go  
+**Eng:** Head of Engineering (IRE) — hardened packaging
+
+---
+
+## 1. POV — is a mixed PR the right vehicle?
+
+**Yes — as one coherent cover / program with layered diffs, not one mega-commit.**
+
+| Layer | Role | Land now? |
+|-------|------|-----------|
+| **Study (PR-D)** | Serve mix original vs RT vs quote; re-serve≥2×; negatives if exposed | **Yes** — instrumentation |
+| **Solutions (PR-A + PR-F1)** | FollowAuthor/ProfileClick A/B; prefer-original For You dedup | **Yes** — FS A/B on |
+| **Structure (PR-F5)** | ControlledSourceReexposure — full FS + filter sketch | **Yes code / default off** — treatment arms only after D baseline |
+
+**Not in this ask as substitutes:** open F2b (burn off) or F4 (post-48h age prolong). F5 is the controlled alternative.  
+**Not pertinent:** study-only with no F5 sketch · F5-only without A/D/F1 · enabling F5 treatment before serve-mix baseline.
+
+**Projection:** Land D + A/F1 A/B + F5 params/filter **flag off** in one program. Keep F5 treatment off for ~1–2 experiment cycles until serve-mix baseline exists, then turn K=2/3 arms. Same “creator reach equilibrium” story — three layers, one cover.
+
+---
+
+## 2. Problem statement (metrics-backed)
+
+1. **Conversion:** cohort follow/imp **0.00636%** vs Batch B **0.0644%** (~10×); FollowAuthor=4 vs CopyLink=20; ProfileClick=0.  
+2. **OCR:** 180/500 VF · 49.1K/500K HT; rolling floor ~**5,556** verified HT/day at η=3.78% or NEVER.  
+3. **Hard TL:** Age 48h · RT XOR original · source burn after serve · quotes bypass RT-dedup.  
+4. **Proxy:** high-amp strata follow/imp **0.0678%** vs low-amp megaviral **0.0053%**; serve % RT-vs-original **unmeasurable** today.  
+5. **Smoke:** current burn blocks niche re-see; F5 restores K=2–3 IN nudges without 20-RT flood (`e2e_smoke_f5/` 22/22 PASS sim).
+
+Hypothesis floors only — not global SLOs.
+
+---
+
+## 3. F5 structure (implementable, default off)
+
+### 3.1 Placement
+Policy **override on seen/served source burn** for a source_id when F5 allows re-serve — **not** an AgeFilter rewrite and **not** OON RT enable.
+
+Order of intent:
+1. Same-slate ≤1 (existing RT dedup + **F1** prefer-original).  
+2. Seen/served burn as today when F5 off (K≡1).  
+3. When F5 on: re-admit source for a later request if K/gap/IN rules pass. Prefer **original** candidate if present; else IN amplifier card of that source.
+
+### 3.2 Feature Store params (proposed)
+
+| Param | Type | Default (ship) | Treatment arms |
+|-------|------|----------------|----------------|
+| `EnableControlledSourceReexposure` | bool | **false** | true |
+| `ControlledSourceReexposureK` | u32 | **1** (≡ today’s no re-see) | **2**, **3** |
+| `ControlledSourceReexposureGapRequests` | u32 | 1 | 1–3 |
+| `ControlledSourceReexposureInNetworkOnly` | bool | **true** | keep true for first ship |
+| `ControlledSourceReexposureIncludeQuotes` | bool | **true** | **fixed true** — no false arm (HoE) |
+
+### 3.3 `source_key` (quotes required)
+
+```
+if retweeted_tweet_id.is_some() -> retweeted_tweet_id
+else if quoted_tweet_id.is_some() -> quoted_tweet_id   # REQUIRED — smoke EXP−
+else -> tweet_id
+```
+
+Quotes and RTs of the same original **share one cap bucket**.
+
+### 3.4 Acceptance
+- Same request: ≤1 card per source (original / RT / quote).  
+- Across requests: ≤K serves per source / viewer / window.  
+- OON RT never used as F5 nudge.  
+- Attribution / logging: `get_original_tweet_id` + card_type ∈ {original, rt, quote}.  
+- No material ↑ report / mute / block vs control (needs PR-D).
+
+### 3.5 Files (sketch — SpaceXAI owns merge)
+- Filter / policy near `previously_seen_*` / `previously_served_*` (re-admission)  
+- Params: `home-mixer/params/param.rs`  
+- Counters: Under the Hood / side effects (PR-D)  
+- Tests: mirror `e2e_smoke_f5/` EXP±  
+
+Detail: `SPACEXAI_F5_IMPLEMENTATION.md`
+
+---
+
+## 4. Mixed PR table of contents (external)
+1. Cover one-pager (soft misalignment + hard rules + F5 north star)  
+2. **PR-D** study / telemetry  
+3. **PR-A** param sketch  
+4. **PR-F1** prefer-original (Following parity)  
+5. **PR-F5** implementation note — flag default **off**  
+6. Explicit non-asks: open F2b / F4 / E · no OCR invent · no merge claim · no VF SpamHighRecall loosen  
+
+Companions already drafted: `SPACEXAI_COVER_TRANCHE1.md`, `SPACEXAI_PARAM_SKETCH_PR_A.md`, `SPACEXAI_PR_D_TELEMETRY_NOTE.md`, `SPACEXAI_RT_QUOTE_REQUEST.md`
+
+---
+
+## 5. Risk / blast radius
+| Risk | Mitigation |
+|------|------------|
+| Quote wall | Same source_key as RT; IncludeQuotes fixed true |
+| Snowball | K + gap + ≤1/slate |
+| Blind A/B | F5 treatment off until D serve-mix baseline |
+| Safety VF | Do not loosen SpamHighRecall |
+| Creator floors as SLOs | Label empirical only |
+
+## 6. Success projection (hypothesis, not SLO)
+If η moves toward ~10% and follow/imp toward Batch B on non-viral mix, OCR HT path becomes more reachable; F5 targets **quality re-exposure**, not raw viral. Celebrity strata stay excluded from baseline reporting.