feat: #24 HIDDEN_COLLAPSE_INTERMITTENT_CLIFF — close estimation lane

MAP.md

@@ -94,7 +94,7 @@ Run the relevant generator after adding content, then verify with its `--check`.
 |---|---|---|
 | 165 | questions | `org_frontier/questions/q*_*/` |
 | 134 | probes | `org_frontier/probes/probe_*.py` |
-| 85 | studies | `org_frontier/studies/*/` |
+| 86 | studies | `org_frontier/studies/*/` |
 | 21 | essays | `org_frontier/essays/*.md` |
 | 9 | foundations experiments | `foundations/*/` with a `FINDINGS.md` |
 | 10 | literature reviews | `org_frontier/reviews/*/` with a `FINDINGS.md` |

README.md

@@ -152,7 +152,7 @@ Cross-program essays and reviews — the best places to start window-shopping.
 - [PR #739 package — arcs on `cursor/triad-template-census-6ac4`](org_frontier/PR739_PACKAGE.md) — PR #739 merge package — closed arcs, instrument, human gates
 - [Omit-atom arc — working picture](org_frontier/OMIT_ATOM_ARC.md) — omit/discrete-Φ arc — motif-ruled atoms, scale morphs
 - [Construct × scale ladder arc — working picture](org_frontier/CONSTRUCT_LADDER_ARC.md) — construct×ladder arc — COMMIT_READ, Φ=n−1 flip, gate seal
-- [Estimation lane (#21–#23, #25) — working picture](org_frontier/ESTIMATION_ARC.md) — estimation lane — topology bottleneck; closable
+- [Estimation lane (#21–#25) — working picture](org_frontier/ESTIMATION_ARC.md) — estimation lane — topology bottleneck; closable
 - [Stoch–temporal arc — working picture](org_frontier/STOCH_TEMPORAL_ARC.md) — stoch–temporal arc — noise, delay, coexistence, CT proxy; closable
 - [Formal theory arc — #47–#50 working picture](org_frontier/FORMAL_THEORY_ARC.md) — formal theory arc — #47–#50; closable
 - [AI / multi-agent arc — #37–#41 synthesis](org_frontier/AI_MULTIAGENT_ARC.md) — AI/multi-agent arc — membership cuts; closable
@@ -344,6 +344,7 @@ Multi-experiment batteries on one theme.
 - **[Oscillatory Φ scaling (agenda #11)](org_frontier/studies/oscillatory_scaling/README.md)** — traveling wave, period=n) carries a **constant Φ=2.0** law over
 - **[Parity radix blind spot (agenda #4)](org_frontier/studies/parity_radix_blindspot/README.md)** — flat full-system Φ band (≈0.5) at k=2,3,4, while min-commit is high
 - **[Parity vs conjunctive under noise (agenda #8)](org_frontier/studies/parity_vs_conjunctive_noise/README.md)** — conjunctive hubs lose the triadic verdict at the **same** p*=0.5 under
+- **[Partial observation / estimability (agenda #24)](org_frontier/studies/partial_observation_screen/README.md)** — strict-mediation n=3 family, mean pairwise MI under full observation
 - **[Party vs mediator noise (agenda #7)](org_frontier/studies/party_vs_mediator_noise/README.md)** — mediator-column flip-noise collapse the triad at the **same** p*=0.5
 - **[Φ-ascent adaptive mediator (agenda #14)](org_frontier/studies/phi_ascent_mediator/README.md)** — Φ does **not** converge to the conjunctive hub. Under fixed party
 - **[Random coupling ensemble at fixed n (agenda #18)](org_frontier/studies/random_coupling_ensemble/README.md)** — n=4 (ER / fixed-k / WS / BA-style), **every** core Φ lands on the discrete

ci/reproduce.json

@@ -4945,6 +4945,20 @@
       "timeout": 240,
       "source": "org_frontier/studies/active_label_acquisition/FINDINGS.md"
     },
+    {
+      "name": "partial-observation-screen",
+      "cmd": "python org_frontier/studies/partial_observation_screen/analyze_partial_obs.py",
+      "expect": [
+        "faithful triad: triadic \u03a6=2.000000  PASS",
+        "H1 (hidden party collapses):     SUPPORTED",
+        "H2 (intermittent smooth in \u03b4):   REFUTED",
+        "H3 (topology/family mediates):   SUPPORTED",
+        "verdict: HIDDEN_COLLAPSE_INTERMITTENT_CLIFF",
+        "full_auc=0.922  hidden_party_auc=0.547  hidden_mediator_auc=0.934"
+      ],
+      "timeout": 120,
+      "source": "org_frontier/studies/partial_observation_screen/FINDINGS.md"
+    },
     {
       "name": "commit-noise-phase",
       "cmd": "python org_frontier/studies/commit_noise_phase/analyze_phase.py",

org_frontier/ESTIMATION_ARC.md

@@ -1,19 +1,20 @@
-# Estimation lane (#21–#23, #25) — working picture
+# Estimation lane (#21–#25) — working picture
 
-A short synthesis of the cheap-screen / surrogate estimation arm on
-PR #739. Exact binary IIT-4.0 Φ; in-silico. Sibling arcs:
+A short synthesis of the cheap-screen / surrogate estimation arm.
+Exact binary IIT-4.0 Φ; in-silico. Sibling arcs:
 [`CONSTRUCT_LADDER_ARC.md`](CONSTRUCT_LADDER_ARC.md) (closed),
 [`OMIT_ATOM_ARC.md`](OMIT_ATOM_ARC.md),
 [`ROLE_TARGET_GRAIN.md`](ROLE_TARGET_GRAIN.md). No construct/omit/
 ladder/indeg reopen.
 
 ## Verdict in one line
 
-**The generalization bottleneck is topology, not sample length,
-GNN-style structure, or label order.** Within-family MI is fast;
-cross-family coupling inverts; structure-aware RF adds little under
-holdout; transfer-operator spectral gap is a partial cross-family
-lever; active acquisition does not beat random on this panel.
+**The generalization bottleneck is topology; under partial observation
+the failure is role-gated (missing party), not smooth in duty cycle.**
+Within-family MI is fast; cross-family coupling inverts; structure-aware
+RF and active acquisition add little; spectral gap is partial; a hidden
+party collapses the cheap screen while intermittent logging (δ≥0.10)
+does not.
 
 ## Arc
 
@@ -22,6 +23,8 @@ lever; active acquisition does not beat random on this panel.
 | 1 | [`structure_aware_surrogate/`](studies/structure_aware_surrogate/) | **NO_STRUCTURE_GAIN** (#22) | cm+function RF ≈ coupling under LOFO (lift +0.049) |
 | 2 | [`sample_complexity_screen/`](studies/sample_complexity_screen/) | **FAST_WITHIN_FAMILY** (#23) | family n=3 MI AUC≥0.97 at T*=125; longer T ≠ cross-topo |
 | 3 | [`spectral_invariant/`](studies/spectral_invariant/) | **SPECTRAL_PARTIAL** (#21) | `P_spectral_gap` AUC=0.893 vs inverted MI 0.775; lift +0.118 < 0.15 |
+| 4 | [`active_label_acquisition/`](studies/active_label_acquisition/) | **AL_NO_GAIN** (#25) | uncertainty Δ=+0.015 vs random; topo_balance hurts LOFO |
+| 5 | [`partial_observation_screen/`](studies/partial_observation_screen/) | **HIDDEN_COLLAPSE_INTERMITTENT_CLIFF** (#24) | hide party 0.922→0.547; δ≥0.10 holds, cliffs at δ=0; hide mediator ≈ full |
 
 Cited priors (not reopened): #122 within-family MI; #123/#129 OOD
 surrogate failure; #134 coupling inversion; cascade/residual arm in
@@ -47,9 +50,17 @@ surrogate failure; #134 coupling inversion; cascade/residual arm in
    misses the pre-registered lift bar (#21). Useful baseline, not a
    closed topology-invariant instrument.
 
-5. **Practice.** Use family-matched cheap screens (or cascade selective
+5. **Partial observation is role-gated.** Hiding a party collapses the
+   within-family MI screen; hiding the mediator does not. Intermittent
+   party observation stays informative down to δ=0.10 and cliffs only
+   at total absence (#24). Missingness fraction is the wrong control;
+   which role is logged is the right one.
+
+6. **Practice.** Use family-matched cheap screens (or cascade selective
    exact Φ). Do not expect one coupling or GNN-style surrogate to travel
-   across topology without labels from that class.
+   across topology without labels from that class. Treat an unobserved
+   party as a hard fail for coupling screens; sparse party logs are
+   much safer than silence.
 
 ## Active learning (#25)
 
@@ -60,9 +71,6 @@ Label order is secondary; it does not repair the topology bottleneck.
 
 ## Best next
 
-**Estimation lane closable.** #21–#23 plus #25 fix the picture:
-within-family MI is cheap and fast; cross-topo needs family-matched
-labels or selective exact Φ; structure-aware, spectral, and active
-acquisition are partial or null levers, not a closed invariant.
-Optional later (outside this lane): agenda **#24** partial observation.
+**Estimation lane closed** (#21–#25). Optional later outside this lane:
+**M3** overlay subset-Φ fidelity, or a new agenda gap.
 Construct/omit/ladder stay closed.

org_frontier/PR739_PACKAGE.md

@@ -25,7 +25,7 @@ In-silico; validation gap stated in each FINDINGS. No Hegel/Substack.
 | Omit-atom | [`OMIT_ATOM_ARC.md`](OMIT_ATOM_ARC.md) | Discrete motif-ruled Φ; size morphs rule form (**SCALE_MORPHS**), not continuum |
 | Role-target grain | [`ROLE_TARGET_GRAIN.md`](ROLE_TARGET_GRAIN.md) | ROLE_TARGETS / Z_TARGETS / JOINT_TZ purify MIX indegs |
 | Construct × ladder | [`CONSTRUCT_LADDER_ARC.md`](CONSTRUCT_LADDER_ARC.md) | COMMIT_READ + monotone Φ=n−1 flip; gate regimes seal (**GATE_SEAL_HOLDS**) |
-| Estimation | [`ESTIMATION_ARC.md`](ESTIMATION_ARC.md) | Topology bottleneck; MI fast within family; AL_NO_GAIN; lane closable |
+| Estimation | [`ESTIMATION_ARC.md`](ESTIMATION_ARC.md) | Topology bottleneck; MI fast within family; AL_NO_GAIN; **#24 HIDDEN_COLLAPSE_INTERMITTENT_CLIFF**; lane closed |
 | Stoch–temporal | [`STOCH_TEMPORAL_ARC.md`](STOCH_TEMPORAL_ARC.md) | p\*=0.5 smooth decay; construction decides time factoring; lane picture closed |
 | Formal theory | [`FORMAL_THEORY_ARC.md`](FORMAL_THEORY_ARC.md) | #47–#50 done (**PARTIAL_PROOFS** / **NORMALIZED_CUT** / **NOT_UNIQUE** / **LATTICE**); lane closable |
 | AI / multi-agent | [`AI_MULTIAGENT_ARC.md`](AI_MULTIAGENT_ARC.md) | Membership cuts transfer; displacement sharp; structure≠learnability; lane closable |
@@ -35,7 +35,7 @@ In-silico; validation gap stated in each FINDINGS. No Hegel/Substack.
 
 ## Agenda map (RESEARCH_AGENDA_50_V2 on this PR)
 
-Closed science cells include A #1–#4, B #5–#14 (stoch–temporal), topology #15–#20, estimation #21–#23+#25, PE #29–#36, AI #37–#41, omit/construct scale #42, CV #43–#46, formal #47–#50. Open curiosities (not blockers): estimation #24 partial observation; optional M3 subset-Φ fidelity; optional deep MARL.
+Closed science cells include A #1–#4, B #5–#14 (stoch–temporal), topology #15–#20, estimation #21–#25, PE #29–#36, AI #37–#41, omit/construct scale #42, CV #43–#46, formal #47–#50. Open curiosities (not blockers): optional M3 subset-Φ fidelity; optional deep MARL.
 
 ## Residual risks (not science reopen)
 
@@ -64,4 +64,4 @@ Spot-check closed arcs via their spine reproduce blocks (each `*_ARC.md`).
 | Merge style | squash vs merge commit | Squash collapses ~160 commits; merge preserves history |
 | Approving review | required (`REVIEW_REQUIRED`) | One approving review per branch protection |
 | CI | wait for reproduce green after 180m budget | update-directory already green historically |
-| Follow-ups | M3 overlay; #24; empirical/survey | Outside this PR’s science scope |
+| Follow-ups | M3 overlay; empirical/survey | Outside this PR’s science scope; #24 closed on follow-up branch |

org_frontier/RESEARCH_AGENDA_50_V2.md

@@ -125,12 +125,16 @@ runnable design yet. Themes group them; numbering is 1–50.
     See `studies/sample_complexity_screen/`.
 24. How fast does estimability degrade under partial observation — a hidden node, or a party observed
     only intermittently?
+    **Answered — HIDDEN_COLLAPSE_INTERMITTENT_CLIFF.** Hide party:
+    MI AUC 0.922→0.547; intermittent δ≥0.10 holds, cliffs at δ=0;
+    hide mediator ≈ full. See `studies/partial_observation_screen/`;
+    lane note `ESTIMATION_ARC.md` (closed).
 25. Which forms are most informative to label first when training a surrogate (active learning over the
     corpus)?
     **Answered — AL_NO_GAIN.** Uncertainty Δ=+0.015 vs random on pooled
     (H1/H2 refuted; H3 supported). Topo-balance hurts under LOFO.
     See `studies/active_label_acquisition/`; lane note
-    `ESTIMATION_ARC.md` (closable).
+    `ESTIMATION_ARC.md` (closed).
 
 ## F. The holistic residual
 

org_frontier/studies/partial_observation_screen/FINDINGS.md

@@ -0,0 +1,88 @@
+# Partial observation / estimability — findings
+
+**Verdict: HIDDEN_COLLAPSE_INTERMITTENT_CLIFF.** On the #122/#23
+strict-mediation n=3 family, mean pairwise MI under full observation
+ranks the verdict at AUC **0.922**. Hiding one party collapses the
+screen to **0.547** (Δ=0.375). Intermittent observation of that party
+does **not** degrade smoothly: AUC stays ≥0.92 for every duty cycle
+δ≥0.10, then cliffs to chance at δ=0 (max consecutive drop **0.422**).
+Hiding the mediator barely moves AUC (**0.934**). Role, not mere
+missingness volume, decides the damage.
+
+In-silico; exact IIT-4.0 labels; complete-case mean MI under the
+observation mask; T=2000, noise=0.08. Hypotheses fixed in
+`hypotheses.md`. Pointers: `sample_complexity_screen/` (#23),
+`spectral_invariant/` (#21), `ESTIMATION_ARC.md`. Construct/omit/
+ladder/indeg closed.
+
+## Curves
+
+### Family n=3 (24 tri / 24 dya) — primary
+
+| regime | MI AUC |
+|---|---:|
+| full | **0.922** |
+| hidden party C | **0.547** |
+| hidden party W | 0.524 |
+| hidden mediator S | **0.934** |
+
+### Intermittent party C (duty cycle δ)
+
+| δ | MI AUC |
+|---:|---:|
+| 1.00 | 0.941 |
+| 0.75 | 0.929 |
+| 0.50 | 0.924 |
+| 0.25 | 0.931 |
+| 0.10 | 0.943 |
+| 0.00 | **0.521** |
+
+Intermittent mediator S stays ≥0.92 at every δ, including 0.00.
+
+### Multifamily secondary (N=23 designed forms)
+
+Per-family AUC often undefined (single-class families). Where both
+labels exist, `single_hub` drops full→hidden 1.000→0.500 (Δ=0.500);
+`chain` / `majority` hold at 1.000. Pooled oriented AUC does not
+reproduce the within-family collapse (cross-topo #134 already limits
+the screen). Role contrast on `family_n3` is the clean H3 witness.
+
+## Hypotheses
+
+| hypothesis | result |
+|---|---|
+| H1 hidden party collapses (Δ≥0.20 or AUC<0.70) | **SUPPORTED** |
+| H2 intermittent degrades smoothly (max drop≤0.25) | **REFUTED** (cliff at δ=0) |
+| H3 topology/family mediates | **SUPPORTED** (mediator vs party \|Δ\|=0.387; family Δ spread=0.500) |
+
+## Reading
+
+Estimability under partial observation is **role-gated**, not a smooth
+function of observation fraction. A sparse party channel (δ≥0.10) is
+enough to keep the #122 MI screen alive on this family. Total absence
+of a party kills it. The mediator can drop out of the log without
+hurting mean-MI ranking — the cheap screen rides party–party coupling
+that survives S-masking on these trajectories. Practice: treat a
+missing party as a hard fail for coupling screens; intermittent logging
+of that party is far less dangerous than silence. Cascade / exact Φ
+remains the backstop when a role is unobserved.
+
+## Limits
+
+One draw per form; designed multifamily N=23 with several single-class
+families (per-family AUC often nan). Oriented AUC. No field logs. No
+claim about real organizations.
+
+## Best next experiment
+
+**#24 closes the estimation lane** with #21–#23+#25. Optional later:
+**M3** overlay subset-Φ fidelity (`BEYOND_BINARY_ARC.md` /
+`PR739_PACKAGE.md`), or a new agenda gap outside estimation. Do not
+reopen construct/omit/ladder. Skip deep MARL unless separately scoped.
+
+## Reproduce
+
+```
+python org_frontier/studies/partial_observation_screen/analyze_partial_obs.py
+```
+(~8 s)

org_frontier/studies/partial_observation_screen/README.md

@@ -0,0 +1,24 @@
+# Partial observation / estimability (agenda #24)
+
+How fast does the cheap mean-MI screen degrade when a node is hidden or
+a party is observed only intermittently?
+
+## Run
+
+```
+python org_frontier/studies/partial_observation_screen/analyze_partial_obs.py
+```
+
+## Result in one line
+
+**HIDDEN_COLLAPSE_INTERMITTENT_CLIFF** — hide party: AUC 0.922→0.547;
+δ≥0.10 holds, cliffs at δ=0; hide mediator ≈ full.
+
+## Hypotheses
+
+Fixed before computing in [`hypotheses.md`](hypotheses.md).
+
+## Priors
+
+`sample_complexity_screen/` (#23), `spectral_invariant/` (#21),
+`ESTIMATION_ARC.md`. Exact binary IIT-4.0 Φ; in-silico.