24601 · 24601 · Sep 20, 2026 · Sep 20, 2026 · Sep 20, 2026
diff --git a/.agents/skills/augustus/SKILL.md b/.agents/skills/augustus/SKILL.md
diff --git a/.agents/skills/augustus/scripts/uniqueness_gate.py b/.agents/skills/augustus/scripts/uniqueness_gate.py
@@ -7,7 +7,11 @@
 Each lock must appear as one consecutive substring in every listed overlay.
 Fragments scattered across files do not count.
 
-Also: YAML-parse SKILL.md frontmatter; notes.md owns §114–§121;
+Revisit / since-last-look protocol (`notes.md` §122) is a consecutive
+substring in the skill + research files (not a 21-overlay dump wall).
+Hourly must treat revisit HIGH like novel HIGH. Star-noise is not a fold.
+
+Also: YAML-parse SKILL.md frontmatter; notes.md owns §114–§122;
 composition items 289–316, 322–329, 330–336, 337–352, 353–368, and 369–384 exist;
 findings batches #97–#104 exist. Items 317–321 stay unused.
 CHANGELOG.md must not hold uniqueness dump walls (dumps live in
@@ -18,6 +22,7 @@
 """
 
 from pathlib import Path
+import subprocess
 import sys
 import yaml
 
@@ -129,6 +134,24 @@
     'Hourly 1143 uniqueness lock: open recreation ≠ calibrated replica; Qwen3.5-4B ≠ Archer; It is an open re-creation of Jev; less calibrated; perch 164★ MIT HEAD ba775a9940b6 README SHA 7ad0403b; semantic lint is a sensor not a proof; oxlint-plugin-jev cutoff 0.8 still soft; nassim-arifette/jevgrep ≠ Bentlybro/jevgrep ≠ can1357/jegrep ≠ uehaj/jev-semgrep; patdown fuzzy linter; PanAchy/jevvy ≠ Atominac/jevvy; No orders, no advice; SmartMoney-Cub 25★ HEAD d93cf493853d; paired bootstrap CIs *theirs*; emretheus/jev-rag-benchmark ≠ erendikmenn/jev-rag-benchmark; +0.82 pts XQuAD-EN 95% CI +0.35 to +1.31; +7.62 pts SciFact CI +4.88 to +10.38; Same accuracy, 35x faster *theirs*; systems comparison ≠ semantic equivalence; BANKING77 500 Jev 81.0% GPT-OSS 82.8% Mercury 73.2% Gemini 85.4% *theirs*; frozen cascade missed its evaluation accuracy target 430/500 vs GPT-5 432/500; This is not demonstrated equal-quality savings; 24 invented tickets; Routing errors caught by the gate 0 of 3; sample too small to establish calibration; This is not TypeSafe Jev; No real API requests were made; wire-compat ≠ replica; KonghaYao/laya-jev 按官方接口写的客户端只改一个 base URL; gqgs/laya-onnx densify 496.8 MiB; tozp/laya-onnx ≠ Mattepiu/laya-onnx ≠ gqgs/laya-onnx; serving substrate ≠ calibrated replica; BeatAPI/awesome-jev ≠ 99hansling/awesome-jev ≠ Vishnurr2k01/awesome-jev ≠ robokrunch/awesome-jev ≠ rudy2steiner/awesome-jev-hub; All 125 projects; catalog ≠ endorsement; Pasblinn/jev-lab ≠ tanayvasishtha/jev-lab ≠ dairui1/jev-lab ≠ BrendanH18/jev-lab ≠ yibie/laya-jev-lab ≠ q93304989-bit/jev-lab; Independent project. Not affiliated with TypeSafe; Kevthetech143/super-jev densify experimental V0.2.0; permission ≠ confidence; allay-team/openjev ≠ piyush-infocusp/openjev ≠ TheoLeeCJ/openjev; 2022 Mineflayer Jevalent collision; kushalpatil/jevify-gemma4-e4b GGUF densify; static quants; This dataset and model are independent research artifacts, not reproductions of Jev or RLCD; pngwn demo accuracy 0.705 ECE 0.046 ~112 ms *theirs*; cutoff 0.8 still soft; soft scores ≠ hard gates; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35/#36/#37/#38/#39/#40/#41/#42/#43; notes.md §121'
 )
 
+REVISIT_LOCK = (
+    "Revisit / since-last-look lock: catalogued repos are not done; "
+    "store fingerprints default_sha, pushed_at, description_hash, release_tag; "
+    "material change is README/API/release/calibration claim/serving port/bench rewrite; "
+    "star-noise is stars/likes/forks alone; densify the prior notes section, "
+    "do not mint a sibling first sighting; do not invent equivalence; "
+    "SHA move is not a replica; treat revisit HIGH like novel HIGH for Augustus; "
+    "notes.md §122"
+)
+
+REVISIT_OVERLAYS = [
+    ".agents/skills/augustus/SKILL.md",
+    "research/notes.md",
+    "research/README.md",
+    "research/revisit-checklist.md",
+    "CONTRIBUTING.md",
+]
+
 OVERLAYS = [
     "research/notes.md",
     "research/changelog-hourly.md",
@@ -190,6 +213,14 @@ def main() -> int:
             failed.append(f"1049 lock missing as one substring: {rel}")
         if UNIQ_1143 not in body:
             failed.append(f"1143 lock missing as one substring: {rel}")
+    for rel in REVISIT_OVERLAYS:
+        path = ROOT / rel
+        if not path.is_file():
+            failed.append(f"missing revisit overlay {rel}")
+            continue
+        body = path.read_text(encoding="utf-8")
+        if REVISIT_LOCK not in body:
+            failed.append(f"revisit lock missing as one substring: {rel}")
     notes = (ROOT / "research/notes.md").read_text(encoding="utf-8")
     if "## 114. Hourly 0843 HIGH" not in notes:
         failed.append("notes.md missing §114 heading")
@@ -207,6 +238,8 @@ def main() -> int:
         failed.append("notes.md missing §120 heading")
     if "## 121. Hourly 1143 HIGH" not in notes:
         failed.append("notes.md missing §121 heading")
+    if "## 122. Revisit / since-last-look" not in notes:
+        failed.append("notes.md missing §122 heading")
     algebra = (ROOT / ".agents/skills/augustus/references/composition-algebra.md").read_text(
         encoding="utf-8"
     )
@@ -316,6 +349,10 @@ def main() -> int:
                 "paired bootstrap CIs *theirs*",
                 "Same accuracy, 35x faster *theirs*",
                 "hourly 1143 / notes.md §121",
+                "revisit HIGH / since-last-look",
+                "catalogued repo changed",
+                "star-noise vs material change",
+                "densify prior notes without inventing equivalence",
             ):
                 if frag not in proto_line:
                     failed.append(f"SKILL.md protocol missing {frag!r}")
@@ -350,6 +387,42 @@ def main() -> int:
     eco = (ROOT / "docs/ecosystem.md").read_text(encoding="utf-8")
     if "decision-circuits tutorial. `notes.md` §114." not in eco:
         failed.append("docs/ecosystem.md 0843 blurb must cite notes.md §114")
+    if REVISIT_LOCK in changelog:
+        failed.append(
+            "CHANGELOG.md holds revisit lock dump "
+            "(protocol lock lives in SKILL.md + research/)"
+        )
+    helper = ROOT / "research/revisit_fingerprints.py"
+    store = ROOT / "research/revisit_fingerprints.json"
+    if not helper.is_file():
+        failed.append("missing research/revisit_fingerprints.py")
+    if not store.is_file():
+        failed.append("missing research/revisit_fingerprints.json")
+    else:
+        try:
+            r = subprocess.run(
+                [sys.executable, str(helper), "--self-test"],
+                cwd=str(ROOT),
+                capture_output=True,
+                text=True,
+                check=False,
+            )
+        except OSError as e:
+            failed.append(f"revisit_fingerprints.py could not run: {e}")
+        else:
+            if r.returncode != 0:
+                failed.append(
+                    "revisit_fingerprints.py --self-test failed: "
+                    + (r.stdout + r.stderr).strip()
+                )
+    skill_headings = (ROOT / ".agents/skills/augustus/SKILL.md").read_text(
+        encoding="utf-8"
+    )
+    if "## REVISIT / since last look" not in skill_headings:
+        failed.append("SKILL.md missing REVISIT / since last look heading")
+    contributing = (ROOT / "CONTRIBUTING.md").read_text(encoding="utf-8")
+    if "revisit HIGH like novel HIGH" not in contributing:
+        failed.append("CONTRIBUTING.md missing revisit HIGH like novel HIGH")
     if failed:
         print("uniqueness-gate FAIL")
         for line in failed:
@@ -362,7 +435,9 @@ def main() -> int:
         f"0940 chars={len(UNIQ_0940)} 0947 chars={len(UNIQ_0947)} "
         f"1049 chars={len(UNIQ_1049)} "
         f"1143 chars={len(UNIQ_1143)} "
-        f"overlays={len(OVERLAYS)}"
+        f"revisit chars={len(REVISIT_LOCK)} "
+        f"overlays={len(OVERLAYS)} "
+        f"revisit_overlays={len(REVISIT_OVERLAYS)}"
     )
     return 0
 

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -16,6 +16,28 @@ folds: `research/notes.md`.
 
 ## [Unreleased]
 
+Revisit / since-last-look protocol (`research/notes.md` §122).
+Does **not** bump the 0.5.0 pin. Catalogued repos get a densify
+card when fingerprints move. Star-noise is not a fold. Treat
+revisit HIGH like novel HIGH. uniqueness_gate.py checks the
+protocol substring in the skill and research files (not a
+21-overlay dump). Do not reopen or amend PR #23–#44. Merged #44
+owns §121.
+
+### Added
+
+- **Revisit / since-last-look protocol (`notes.md` §122).** Store
+  fingerprints `default_sha`, `pushed_at`, `description_hash`,
+  `release_tag` so hourly can diff. Material change is README /
+  API / release / calibration claim / serving port / bench
+  rewrite. Star-noise is stars / likes / forks alone. Densify
+  the prior notes section; do not mint a sibling first
+  sighting; do not invent equivalence; SHA move is not a
+  replica. Treat revisit HIGH like novel HIGH for Augustus.
+  Helper: `research/revisit_fingerprints.py`. Checklist:
+  `research/revisit-checklist.md`. **HARD RULE:** do not reopen
+  or amend PR #23–#44. Does **not** bump 0.5.0.
+
 Hourly 1143 HIGH (`research/notes.md` §121 / composition items
 369–384 / findings batch #104). Does **not** bump the 0.5.0 pin.
 Uniqueness dumps live in

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -28,6 +28,7 @@ Run what you can locally:
 ```bash
 python3 .agents/skills/augustus/scripts/evaluate_decisions.py --self-test
 python3 .agents/skills/augustus/scripts/uniqueness_gate.py
+python3 research/revisit_fingerprints.py --self-test
 ```
 
 The Pages workflow must stay green. After #34 it greps `_site/index.html`
@@ -43,10 +44,21 @@ re-opened as "new." Before folding:
 - Read `research/notes.md` and the uniqueness fragments in
   `.agents/skills/augustus/SKILL.md`
 - Do not re-fold an already-landed section as a new beat
-- Do not reopen or amend a merged fold PR (#23–#43)
-- uniqueness_gate.py checks 0843 + 0915 + jcr + 0922 + 0940 + 0947 + 1049 + 1143 consecutive locks.
+- Do not reopen or amend a merged fold PR (#23–#44)
+- uniqueness_gate.py checks 0843 + 0915 + jcr + 0922 + 0940 + 0947 + 1049 + 1143 consecutive locks, plus the revisit / since-last-look protocol substring in the skill and research files.
 - Hourly uniqueness dump: `research/changelog-hourly.md` (archive,
   not release notes)
+- Treat **revisit HIGH like novel HIGH**. Catalogued repos are not
+  done. If fingerprints moved (`default_sha`, `pushed_at`,
+  `description_hash`, `release_tag`) or the README / API / release /
+  calibration claim / serving port / bench rewrote, densify the prior
+  notes card. Stars / likes / forks alone is star-noise, not a fold.
+  Do not mint a sibling first sighting. Do not invent equivalence.
+  SHA move is not a replica. Checklist:
+  `research/revisit-checklist.md`. Helper:
+  `research/revisit_fingerprints.py`. `notes.md` §122.
+
+Revisit / since-last-look lock: catalogued repos are not done; store fingerprints default_sha, pushed_at, description_hash, release_tag; material change is README/API/release/calibration claim/serving port/bench rewrite; star-noise is stars/likes/forks alone; densify the prior notes section, do not mint a sibling first sighting; do not invent equivalence; SHA move is not a replica; treat revisit HIGH like novel HIGH for Augustus; notes.md §122
 
 ## Secrets
 

diff --git a/README.md b/README.md
@@ -74,8 +74,9 @@ One line per file. The living catalog is in the reference cards and
 - `.agents/skills/augustus/references/optimizer-integration.md`: Jev inside Ax/DSPy optimizer loops
 - `.agents/skills/augustus/references/agent-self-assessment.md`: agent self-supervision gates (pre-action, done, stuck)
 - `.agents/skills/augustus/scripts/evaluate_decisions.py`: offline Brier / reliability / cost-threshold evaluator
-- `research/notes.md`: living hourly catalog (dense); §118 llm-to-jev conversion assistant (heuristic on-ramp, not a replica)
+- `research/notes.md`: living hourly catalog (dense); §118 llm-to-jev conversion assistant (heuristic on-ramp, not a replica); §122 revisit / since-last-look protocol
 - `research/README.md`: evidence archive index (sources, refresh log, hourly dumps)
+- `research/revisit-checklist.md`: revisit already-catalogued repos when fingerprints move (revisit HIGH like novel HIGH)
 - `research/changelog-hourly.md`: hourly uniqueness dumps after v0.3.0
 - Hourly 0743 uniqueness dump: [`research/changelog-hourly.md`](research/changelog-hourly.md) (`notes.md` §113 / batch #96).
 - Hourly 0843 uniqueness lock: A hunch is a probability with a policy attached; { enter: 0.8, exit: 0.6 } is hysteresis; replay a policy change without inference; Decision models are providers, not the product; huncho ≠ Kungie/gut ≠ carldaws/hunch ≠ tpellet/hunch; pretrained Qwen2.5 base ECE 0.030 (0.5B) / 0.040 (7B); instruct 0.302 / 0.269; 70.9% → 70.0% mean conf 74.1% → 96.7%; temperature scaling still matches it in-distribution; No Jev API was called; Qwen2.5 ≠ Archer; Qwen/Qwen3.8-27B ≠ Archer; 学習済みモデル v0.1 は準備中です; bool AUROC 0.523; 先頭だと0件、末尾だと250件; 温度を渡さない場合、確率は較正されていません; このリポジトリには Jev を呼ぶコードが存在しません; g0runmezadam/what-is-jev IS tunahansahin897/what-is-jev (same GitHub id 1378007307); 947 repos scored; A 273 · B 302 · C 372; LLM rubric ≠ benches; Data as of 2026-09-20; HEAD 895b9498; README SHA 3ae98c56; 13 focused checks and one mutually exclusive outcome; Probabilities are advisory, not calibrated guarantees; omni-/ask-jev ≠ pedroknigge/mcp_jev; pd.cut bins by equal width while jeval bins by quantile; ECE 0.113 and ECE 0.076; jeval drift is not implemented yet; rlaope/jeval ≠ dayhaysoos/jevals; calibration does not compose; ECE has exactly zero statistical power to detect the failure mode that kills trajectories; 25–60× headline withdrawn; P(all-correct): 0.0071 vs 0.0001; TCE / AMS; Qwen 3.8 sparring ≠ Archer; Deferred Crispification; light_cutoff_applied_to_combination 0; recorded run, kinematic animation; BANKING77 Accuracy BERT-Base 93.02 Jev 79.90; Analyse jev calibration (NLL, ECE) backlog; BERT figures are published supervised references, not zero-shot; 档位措辞效应 分数极差中位 0.50、最大 1.32; 修好后对照组是 0.01; 不是 benchmark; 概率没做 calibration; ~1,430 API calls, about $0.15; xiaohuaxi/jev-study ≠ wjdjdakf17/jev-study ≠ baekenough/jev-study; AND: product (independence assumed and recorded in the trace); chat model's stated confidence is not calibrated; circuit-vl-4b ≠ Archer; Bring your own API key; vamsikrishna2421/jev-usecases ≠ whyashthakker/awesome-jev-use-cases; catalog ≠ endorsement; SemIf 2237★ (+30 vs §111 2207); jevlike 1049★ (+6 vs 1043); TypeLLM/TypeLLM 16★; AnotiaWang 98★; yibie/awesome-jev 520★ (+14 vs 506); Laya likes 861 (was 822); tracker likes 66 (+2 vs 64) lastModified UNCHANGED; Blackwood likes 2 gated manual; Archer still promised_not_landed; Hub archerhume/4rcherhume HTTP 401; do not reopen or amend PR #23/#24/#25/#26/#27/#28/#29/#30/#31/#32/#33/#34/#35; notes.md §114

diff --git a/research/README.md b/research/README.md
@@ -13,6 +13,13 @@ refreshes diff against a known baseline instead of re-discovering the world.
 - `archive/curriculum/` — attached research briefs folded into the skill
   (formal-methods × System One, mental-models across domains, source
   list). Provenance; the skill cards are the doctrine.
+- `revisit-checklist.md` — since-last-look protocol. Treat revisit HIGH
+  like novel HIGH. Material change vs star-noise. How densify cards
+  update prior notes without inventing equivalence.
+- `revisit_fingerprints.json` — last-look snapshots (`default_sha`,
+  `pushed_at`, `description_hash`, `release_tag`) so hourly can diff.
+- `revisit_fingerprints.py` — offline classifier. `--self-test` does
+  not fetch the network. SHA move is not a replica.
 
 Method (2026-09-18, UTC): built-in web search + direct docs reads
 (docs.typesafe.ai via llms.txt, GitHub READMEs, launch coverage, X posts via
@@ -34,3 +41,14 @@ result pages, sort by updated/created, never by stars alone.
 Hourly dumps (2026-09-18T14): when a live X+GH scan is provided, archive it
 verbatim under `archive/hourly/` before distilling into notes.md. Do not
 treat X ECE/latency claims as Contract until reproduced.
+
+Revisit / since last look (2026-09-20 standing order): already-catalogued
+repos are not done. Hourly diffs fingerprints against the last look.
+Material change (README / API / release / calibration claim / serving
+port / bench rewrite) is revisit HIGH: densify the prior `notes.md`
+section, do not mint a sibling first sighting, do not invent
+equivalence. Star-noise (stars / likes / forks alone) is a pulse, not
+a fold. Hourly fold prompts must treat revisit HIGH like novel HIGH
+for Augustus. `notes.md` §122.
+
+Revisit / since-last-look lock: catalogued repos are not done; store fingerprints default_sha, pushed_at, description_hash, release_tag; material change is README/API/release/calibration claim/serving port/bench rewrite; star-noise is stars/likes/forks alone; densify the prior notes section, do not mint a sibling first sighting; do not invent equivalence; SHA move is not a replica; treat revisit HIGH like novel HIGH for Augustus; notes.md §122
diff --git a/research/changelog-hourly.md b/research/changelog-hourly.md
@@ -13,6 +13,23 @@ This is the uniqueness-lock archive after hourly folds (#2–#40 / notes
 
 ---
 
+## Revisit / since-last-look protocol (notes.md §122)
+
+- Fresh PR off latest `main` after merged #44 (1143 / §121). **HARD RULE:**
+  do not reopen or amend PR #23–#44. Does not bump 0.5.0. Skip Archer.
+  Quote *theirs*. No wrappers. `invented_signal: false`.
+- Treat revisit HIGH like novel HIGH. Fingerprints: `default_sha`,
+  `pushed_at`, `description_hash`, `release_tag`. Material change is
+  README / API / release / calibration claim / serving port / bench
+  rewrite. Star-noise is stars / likes / forks alone. Densify the prior
+  notes section. Do not mint a sibling first sighting. SHA move is not
+  a replica.
+- Helper: `research/revisit_fingerprints.py`. Checklist:
+  `research/revisit-checklist.md`. uniqueness_gate checks the protocol
+  substring in the skill and research files (not a 21-overlay dump).
+
+---
+
 ## Hourly 1143 HIGH (notes.md §121 / items 369–384 / batch #104)
 
 - Fresh PR off latest `main` after merged #43 (1049 / §120). **HARD RULE:**