pwn 0.5.647 → 0.5.650

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: b6227e3d61730e19d8ec946eaf02247f6222412204cf1b1977fdc473c9bb83e3
-  data.tar.gz: e2fbd14507544beeecc70a7894ba0fcb26263acca09e892ddce16699ce892ce7
+  metadata.gz: 394ee3adf7ad1cbf7d9259288a928bfd210af98deb50e77d08337e2068b1f3ec
+  data.tar.gz: 69668afba8bc9e82043e7497a997494e31c7fbc07b5b77f4e32dfa974cc1c0cf
 SHA512:
-  metadata.gz: cd7cd23daf97bc556f97d7ed58323268d414e1a4ef119b4f6b562f0d324d808dcc7cb1be4fb2708fa957332cc42e5daaa4b410a4756afb40f6520b97e85fd56e
-  data.tar.gz: 7e1576438befe49d01f189aeebd71dac1940f95273c9933c7f51feb0fb36423ee372bfed02e43e5df9558ed971b84606eed1d3d5f82c534cc656ed307618700c
+  metadata.gz: 2979343849a38d63ba659b54d606faca9e0de51e855584910a58e63c9282cdc02456f3315941a8fceb479564fb2410fba0a7fefab9672eb3e10b2134389599ee
+  data.tar.gz: 7896ea04844a847a9c17f370ed9536be045a92707f952efd6759791f868a7e87ce9e3e761fa7784d1e156ae3f03847e4ee1bf4762e405479dae6afd01f7d2839

data/documentation/Agent-Tool-Registry.md

@@ -6,7 +6,7 @@ toolsets; the JSON-Schema for each tool is what the model actually sees.
 
 ![Tool registry](diagrams/agent-tool-registry.svg)
 
-## Toolsets → Tools  (10 toolsets · 71 tools)
+## Toolsets → Tools  (12 toolsets · 78 tools)
 
 | Toolset | Tools | Backed by |
 |---|---|---|
@@ -15,7 +15,9 @@ toolsets; the JSON-Schema for each tool is what the model actually sees.
 | `memory` | `memory_remember` · `memory_recall` · `memory_forget` · `memory_clear` | `PWN::Memory` → `~/.pwn/memory.json` |
 | `skills` | `skill_list` · `skill_view` · `skill_create` · `skill_add_reference` · `skill_delete` · `skill_migrate_legacy` | `~/.pwn/skills/<name>/SKILL.md` (**[agentskills.io](https://agentskills.io) spec**; legacy flat `*.md` auto-migrated) |
 | `sessions` | `sessions_list` · `sessions_view` · `sessions_current` · `sessions_delete` · `sessions_stats` | `PWN::Sessions` → `~/.pwn/sessions/` |
-| `learning` | `learning_note_outcome` · `learning_reflect` · `learning_distill_skill` · `learning_stats` · `learning_outcomes` · `learning_consolidate` · `learning_reset` · `learning_auto_introspect_toggle` · **`mistakes_list`** · **`mistakes_record`** · **`mistakes_resolve`** · **`mistakes_reset`** · **`reward_judge`** · **`reward_prm`** · **`reward_sentinel`** · **`reward_preferences`** · **`reward_export_dpo`** · **`curriculum_practice`** · **`curriculum_train`** · **`curriculum_hindsight`** · **`learning_purge_noise`** | `PWN::AI::Agent::Learning` + `Mistakes` + `Reward` + `Curriculum` → `~/.pwn/learning.jsonl` + `~/.pwn/mistakes.json` + `~/.pwn/preferences.jsonl` + `~/.pwn/curriculum/` + `~/.pwn/finetune/` |
+| `learning` | `learning_note_outcome` · `learning_reflect` · `learning_distill_skill` · `learning_stats` · `learning_outcomes` · `learning_consolidate` · `learning_reset` · `learning_auto_introspect_toggle` · **`mistakes_list`** · **`mistakes_record`** · **`mistakes_resolve`** · **`mistakes_reset`** · **`reward_judge`** · **`reward_prm`** · **`reward_sentinel`** · **`reward_preferences`** · **`reward_export_dpo`** · **`reward_warm_sentinel`** · **`reward_scrub_preferences`** · **`reward_preference_balance`** · **`curriculum_practice`** · **`curriculum_train`** · **`curriculum_hindsight`** · **`curriculum_offline_judge`** · **`curriculum_preference_balance`** · **`learning_purge_noise`** | `PWN::AI::Agent::Learning` + `Mistakes` + `Reward` + `Curriculum` → `~/.pwn/learning.jsonl` + `~/.pwn/mistakes.json` + `~/.pwn/preferences.jsonl` + `~/.pwn/curriculum/` + `~/.pwn/finetune/` |
+| `reward` | **`reward_generator_mix`** | `PWN::AI::Agent::Reward.generator_mix` → W1 online source-mix controller (`preferences.jsonl`) |
+| `curriculum` | **`curriculum_practice_kpi`** | `PWN::AI::Agent::Curriculum.practice_kpi` → `~/.pwn/curriculum_kpi.jsonl` |
 | `metrics` | `metrics_summary` · `metrics_reset` | `PWN::AI::Agent::Metrics` → `~/.pwn/metrics.json` |
 | `extrospection` | `extro_snapshot` · `extro_drift` · `extro_observe` · `extro_observations` · `extro_intel` · **`extro_watch`** · **`extro_verify`** · **`extro_rf_tune`** · **`extro_osint`** · **`extro_serial`** · **`extro_telecomm`** · **`extro_packet`** · **`extro_vision`** · **`extro_voice`** · `extro_correlate` · `extro_stats` · `extro_reset` · `extro_auto_toggle` | `PWN::AI::Agent::Extrospection` (+ Serial/Packet/OCR/Voice/BareSIP/TransparentBrowser/GQRX) → `~/.pwn/extrospection.json` |
 | `cron` | `cron_list` · `cron_create` · `cron_run` · `cron_enable` · `cron_disable` · `cron_remove` | `PWN::Cron` → `~/.pwn/cron/jobs.yml` |
@@ -24,7 +26,10 @@ toolsets; the JSON-Schema for each tool is what the model actually sees.
 The `learning` toolset is deliberately fat: **Mistakes** (negative feedback),
 **Reward** (ORM/PRM/sentinel/DPO ledger) and **Curriculum** (self-play, HER,
 regression-gated LoRA) are all facets of the same self-improvement loop -
-see [Reinforcement Learning](Reinforcement-Learning.md).
+see [Reinforcement Learning](Reinforcement-Learning.md). The thin `reward` and
+`curriculum` toolsets expose controller/KPI surfaces (`reward_generator_mix`,
+`curriculum_practice_kpi`) so personas can grant just those without the full
+learning surface.
 
 ## Dynamic tool-set slimming (`ai.agent.tool_router`)
 
@@ -45,8 +50,8 @@ CORE_TOOLS  = shell · pwn_eval · memory_remember · memory_recall
 ```ruby
 PWN::AI::Agent::Registry.definitions(relevance: 'nmap sweep 10.0.0.0/8', top_k: 10)
 PWN::AI::Agent::Registry.rank(query: 'run a shell command')   # inspect ranking
-PWN::AI::Agent::Registry.toolsets                              # → the 10 names above
-PWN::AI::Agent::Registry.all.count                             # → 71
+PWN::AI::Agent::Registry.toolsets                              # → the 12 names above
+PWN::AI::Agent::Registry.all.count                             # → 78
 ```
 
 Frontier engines leave `tool_router` off and receive the full set.

data/documentation/Home.md

@@ -36,7 +36,7 @@
 | | |
 |---|---|
 | [AI / LLM Integration](AI-Integration.md) | OpenAI · Anthropic · Grok (OAuth) · Gemini · Ollama |
-| [Agent Tool Registry](Agent-Tool-Registry.md) | 10 toolsets · **71** LLM-callable tools |
+| [Agent Tool Registry](Agent-Tool-Registry.md) | 12 toolsets · **78** LLM-callable tools |
 | [Memory · Skills · Learning](Skills-Memory-Learning.md) | Introspection - the self-improvement loop |
 | [Mistakes](Mistakes.md) | **Negative feedback** - fingerprint failures · do-NOT-repeat · `[REPEATING]`/`[REGRESSED]` · inline self-correction |
 | [Reinforcement Learning](Reinforcement-Learning.md) | **`Reward` + `Curriculum`** - ORM/PRM · P9/P14 trajectory DPO · P15 scrub · P16 warm · P17 budget skill · P18 PRM→rank · P19 diet gate · export-ready LoRA |

data/documentation/How-PWN-Works.md

@@ -28,7 +28,7 @@ hardware).
 | Module | Role |
 |---|---|
 | `Loop` | plan → dispatch tool_calls → observe → repeat until final answer |
-| `Registry` | JSON-Schema function definitions grouped into 10 **toolsets** · **71 tools** |
+| `Registry` | JSON-Schema function definitions grouped into 12 **toolsets** · **78 tools** |
 | `Dispatch` / `Result` | execute a tool, capture stdout/value/error/duration |
 | `PromptBuilder` | inject MEMORY / SKILLS / LEARNING / **KNOWN MISTAKES + FIXES** / METRICS / EXTROSPECTION blocks |
 | `Metrics` · `Learning` · `Reflect` | **introspection** - how well am I doing? |

data/documentation/Reinforcement-Learning.md

@@ -175,11 +175,46 @@ and source diversity, not missing tiers:
 3. P15 scrub + export geometry filter purge historical prose flood before DPO/LoRA.
 4. R3/W3 warm via `offline_judge` + `warm_sentinel` so controllers are not permanently cold on ollama hosts (P10/P16).
 5. SFT is filtered as hard as DPO; gate v2 needs smoke + mean, not count alone; P19 also requires trajectory diet.
-6. Budget-exhaustion is a first-class curriculum target (P17/P23); PRM step_reward biases `Registry.rank` (P18); critic is text-only when budget-hot (P24).
+6. Budget-exhaustion is a first-class curriculum target (P17/P23); PRM step_reward biases `Registry.rank` (P18); critic is text-only when budget-hot (P24); last-iter force-final + no-CF-when-hot + exhaust-path Learning close the bare-string hole.
 7. ORM judge blends into Metrics/UCB when distrust is high (P20); W3 calibration lights from plan_first `p(success)=` (P22).
 8. Resolve/record_preference are trajectory-only at write time (P21/P25); write-time source quota still caps traj monoculture.
 9. Without unsloth/axolotl **or** a clean preference diet, `train_and_gate` stays `weight_loop: :export_ready` (honest).
 
+## Design-priority STATUS (post P14–P25)
+
+Collapse of the review-driven priority order into a living table. Stop minting new P-numbers for chores already covered; track these outcomes instead.
+
+| Pri | ID | Control | Module(s) | Success criterion |
+|-----|----|---------|-----------|-------------------|
+| **P0** | W1 generator diversity | `Reward::TARGET_SOURCE_MIX` + `generator_mix` + mix-urgent force on critic/counterfactual | `reward.rb`, `curriculum.rb` | `generator_mix.healthy` OR `recommendation` not stuck on `suppress:mistakes_resolve`; trajectory_fraction ≥ 0.5 |
+| **P0** | Introspect budget | `Learning::INTROSPECT_SOFT_MS` / `HARD_MS`; stage skip under soft/hard / `budget_exhaustion_hot?` | `learning.rb` | `auto_introspect` returns `stages_skipped` when over soft; post-answer path cannot re-thrash tool critic |
+| **P1** | Local judge calibration | Heuristic score shrinkage + `confidence`; `Metrics.effective_rate` scales distrust by `judge_confidence` | `reward.rb`, `metrics.rb` | distrust×heuristic no longer fully replaces proxy; local no-trace highs capped |
+| **P1** | Practice outer KPI | `Curriculum.practice_kpi` / `repeating_trend` → `~/.pwn/curriculum_kpi.jsonl` | `curriculum.rb` | week-over-week `delta_repeating` ≤ 0 on budget fingerprints after practice nights |
+| **P2** | PRM sample efficiency | `PRM_MIN_N=5`, shrinkage to `PRM_FULL_N=20`, fleet coverage gate in `Registry.rank` | `metrics.rb`, `registry.rb` | `prm_advantage=0` until n≥5; rank delta=0 until ≥3 tools ready |
+| **P2** | STATUS over flag archaeology | This table | docs | New work cites Pri/ID here, not fresh P26+ comments for the same theme |
+| **ops** | Nightly diet close | `offline_judge` → `scrub_preferences` + `generator_mix` + `practice_kpi` | `curriculum.rb` | Cron path returns `scrub`/`generator_mix`/`practice_kpi`; raw resolve prose does not survive the night |
+| **ops** | Shape backfill | `Reward.infer_shape` + scrub rewrite | `reward.rb` | Legacy shapeless rows get `winning_trace`/`revised_answer` when content warrants; traj_f measurable |
+| **ops** | Mix in prompt | `Metrics.to_context` emits `W1 MIX:` when unhealthy | `metrics.rb` | Unhealthy diet visible every turn without a tool call |
+| **P0** | Budget exhaust deepen | Last-iter force-final (tools=nil); skip CF when `budget_exhaustion_hot?`; tighter caps 8/12; exhaust path `append_session`+`auto_introspect` | `loop.rb` | Exhaust returns a judged final, not a bare string; CF cannot re-enter under hot; last iter cannot tool-call |
+
+
+### Config additions
+
+```yaml
+:ai:
+  :agent:
+    # P0 introspect budget (ms wall-clock inside auto_introspect)
+    # INTROSPECT_SOFT_MS / HARD_MS are constants; override only via code/reload today.
+    :critic: null            # also force-on when generator_mix.urgent includes critic
+    :counterfactual: null    # also force-on when generator_mix.urgent includes counterfactual
+```
+
+### Tools (design-priority)
+
+`reward_generator_mix` · `curriculum_practice_kpi` (plus existing reward/curriculum set)
+
+---
+
 **See also:** [Skills, Memory & Learning](Skills-Memory-Learning.md) ·
 [Mistakes](Mistakes.md) · [Cron](Cron.md) · [pwn-ai Agent](pwn-ai-Agent.md)
 

data/documentation/What-is-PWN.md

@@ -22,7 +22,7 @@ tool-calling AI agent** on top of it.
 | `PWN::FFI::*` | **8** | Native DSP/RF backends: Volk · Liquid · FFTW · RTLSdr · HackRF · AdalmPluto · SoapySDR · Stdio |
 | `PWN::AI::*` | **5** engines | OpenAI, Anthropic, Grok (OAuth device-flow), Gemini, Ollama |
 | `bin/pwn_*` | **53** | Headless CLI drivers for CI/CD |
-| Agent toolsets | **10** · **71 tools** | terminal · pwn · memory · skills · sessions · learning · metrics · extrospection · cron · swarm |
+| Agent toolsets | **12** · **78 tools** | terminal · pwn · memory · skills · sessions · learning · metrics · extrospection · cron · swarm |
 
 ## The three ways to use it
 

data/documentation/pwn-ai-Agent.md

@@ -81,7 +81,7 @@ $ pwn --ai "run bin/pwn_sast against ./src and push findings to DefectDojo"
 
 ## What the agent can call
 
-10 toolsets · **71 tools** - full table at
+12 toolsets · **78 tools** - full table at
 [Agent Tool Registry](Agent-Tool-Registry.md).
 
 The two that matter most:

data/lib/pwn/ai/agent/curriculum.rb

@@ -193,12 +193,15 @@ module PWN
           end
           save_cooldown(cooldown: cool)
           log(event: :practice, data: results)
+          kpi = practice_kpi(results: results)
           {
             practiced: results.length,
             resolved: results.count { |r| r[:resolved] },
             skipped_cooldown: cool.count { |_, v| v[:fail_nights].to_i >= COOLDOWN_FAIL_NIGHTS },
             results: results,
-            dry_run: dry_run
+            dry_run: dry_run,
+            kpi: kpi,
+            generator_mix: (defined?(Reward) && Reward.respond_to?(:generator_mix) ? Reward.generator_mix : nil)
           }
         rescue StandardError => e
           { error: "#{e.class}: #{e.message}" }
@@ -304,11 +307,22 @@ module PWN
           mean = scored.empty? ? nil : (scored.sum { |r| r[:score].to_f } / scored.length).round(3)
           # P10 — keep R3 window warm so proxy_distrust can engage on local hosts
           warm = (Reward.warm_sentinel(limit: 120) if commit && defined?(Reward) && Reward.respond_to?(:warm_sentinel))
+          # P0 ops — nightly ledger hygiene so generator_mix success criteria can
+          # fire: drop prose flood, backfill missing shapes, then snapshot mix+KPI.
+          scrub = nil
+          mix = nil
+          kpi = nil
+          if commit && defined?(Reward)
+            scrub = Reward.scrub_preferences(dry_run: false) if Reward.respond_to?(:scrub_preferences)
+            mix = Reward.generator_mix if Reward.respond_to?(:generator_mix)
+          end
+          kpi = practice_kpi(results: []) if commit && respond_to?(:practice_kpi)
           out = {
             scored: scored.length, mean: mean, since_hours: since_h,
-            results: scored.first(10), sentinel_warm: warm
+            results: scored.first(10), sentinel_warm: warm,
+            scrub: scrub, generator_mix: mix, practice_kpi: kpi
           }
-          log(event: :offline_judge, data: out)
+          log(event: :offline_judge, data: out.except(:results))
           out
         rescue StandardError => e
           { error: "#{e.class}: #{e.message}" }
@@ -349,7 +363,16 @@ module PWN
         end
 
         public_class_method def self.counterfactual(opts = {})
-          return nil unless enabled?(key: :counterfactual)
+          # P0 — when generator_mix marks counterfactual underfilled, run even
+          # if the auto-flag would keep it off (remote-default still respected
+          # only when mix is healthy). Still refuse recursion.
+          mix_need = begin
+            m = defined?(Reward) && Reward.respond_to?(:generator_mix) ? Reward.generator_mix : {}
+            Array(m[:urgent]).include?('counterfactual')
+          rescue StandardError
+            false
+          end
+          return nil unless enabled?(key: :counterfactual) || mix_need || opts[:force]
           return nil if in_curriculum?
 
           request = opts[:request].to_s
@@ -416,7 +439,13 @@ module PWN
         # self-correction becomes DPO signal.
 
         public_class_method def self.critic(opts = {})
-          return { verdict: :pass, source: :disabled } unless enabled?(key: :critic) || opts[:text_only]
+          mix_need = begin
+            m = defined?(Reward) && Reward.respond_to?(:generator_mix) ? Reward.generator_mix : {}
+            Array(m[:urgent]).include?('critic')
+          rescue StandardError
+            false
+          end
+          return { verdict: :pass, source: :disabled } unless enabled?(key: :critic) || opts[:text_only] || mix_need || opts[:force]
           return { verdict: :pass, source: :recursion } if in_curriculum?
 
           # P24 — text_only: single Reflect shot, no tool-armed persona swarm.
@@ -473,6 +502,17 @@ module PWN
           return nil unless enabled?(key: :red_team_plan)
           return nil if in_curriculum?
 
+          # P17 — never nest a red-team persona loop when budget_exhaustion
+          # fingerprints dominate open mistakes (amplifier of agent_loop ×N).
+          begin
+            if defined?(Loop) && Loop.respond_to?(:budget_exhaustion_hot?, true) &&
+               Loop.send(:budget_exhaustion_hot?)
+              return nil
+            end
+          rescue StandardError
+            # fall through
+          end
+
           ensure_persona(name: RED_TEAM_NAME, role: 'You are pwn-ai\'s adversarial plan reviewer. Given a numbered tool plan and telemetry from THIS host (tool success rates, known mistakes, environment drift), identify the ONE step most likely to fail and say why in ≤2 lines. Cite the metric/mistake/drift. If the plan is sound reply: SOUND.')
           telemetry = build_telemetry
           reply = with_curriculum_guard do
@@ -602,6 +642,86 @@ module PWN
         # Supported Method Parameters::
         # PWN::AI::Agent::Curriculum.calibrate(predicted:, actual:, engine:)
 
+        # ----------------------------------------------------------------
+        # P1 — Outer curriculum KPI: does practice cut live [REPEATING]?
+        # ----------------------------------------------------------------
+        # Snapshot unresolved repeating counts before/after practice nights
+        # into ~/.pwn/curriculum_kpi.jsonl so week-over-week delta is visible
+        # without scraping Mistakes by hand. practice() always appends a row.
+
+        KPI_FILE = File.join(Dir.home, '.pwn', 'curriculum_kpi.jsonl')
+
+        public_class_method def self.practice_kpi(opts = {})
+          results = Array(opts[:results])
+          top = defined?(Mistakes) ? Mistakes.top(limit: 50, unresolved_only: true) : []
+          repeating = top.select { |m| m[:count].to_i >= 3 }
+          budgetish = repeating.count do |m|
+            t = m[:tool].to_s
+            e = m[:error].to_s.downcase
+            t == 'agent_loop' || t == 'assistant_answer' ||
+              e.include?('budget') || e.include?('iteration budget')
+          end
+          row = {
+            at: Time.now.utc.iso8601,
+            unresolved_total: top.length,
+            repeating_n: repeating.length,
+            repeating_sum_count: repeating.sum { |m| m[:count].to_i },
+            budget_repeating_n: budgetish,
+            practiced: results.length,
+            resolved_tonight: results.count { |r| r[:resolved] },
+            mean_holdout: if results.empty?
+                            nil
+                          else
+                            (results.sum { |r| r[:mean_score].to_f } / results.length).round(3)
+                          end
+          }
+          begin
+            FileUtils.mkdir_p(File.dirname(KPI_FILE))
+            File.open(KPI_FILE, 'a') { |f| f.puts(JSON.generate(row)) }
+          rescue StandardError
+            nil
+          end
+          trend = repeating_trend
+          row.merge(trend: trend)
+        rescue StandardError => e
+          { error: "#{e.class}: #{e.message}" }
+        end
+
+        # Week-over-week (or last-N snapshots) delta on repeating_n.
+        # Positive delta_repeating = getting worse; negative = practice working.
+        public_class_method def self.repeating_trend(opts = {})
+          limit = (opts[:limit] || 14).to_i
+          return { samples: 0, delta_repeating: nil, status: :no_data } unless File.exist?(KPI_FILE)
+
+          rows = File.readlines(KPI_FILE).last(limit).filter_map do |l|
+            JSON.parse(l, symbolize_names: true)
+          rescue StandardError
+            nil
+          end
+          return { samples: 0, delta_repeating: nil, status: :no_data } if rows.empty?
+          return { samples: rows.length, delta_repeating: 0, status: :baseline, latest: rows.last } if rows.length < 2
+
+          first = rows.first
+          last = rows.last
+          d_rep = last[:repeating_n].to_i - first[:repeating_n].to_i
+          d_budget = last[:budget_repeating_n].to_i - first[:budget_repeating_n].to_i
+          status = if d_rep <= -2 then :improving
+                   elsif d_rep >= 2 then :regressing
+                   else :flat
+                   end
+          {
+            samples: rows.length,
+            from: first[:at],
+            to: last[:at],
+            delta_repeating: d_rep,
+            delta_budget_repeating: d_budget,
+            latest_repeating_n: last[:repeating_n],
+            status: status
+          }
+        rescue StandardError => e
+          { samples: 0, status: :error, error: "#{e.class}: #{e.message}" }
+        end
+
         public_class_method def self.calibrate(opts = {})
           p = opts[:predicted].to_f.clamp(0.0, 1.0)
           a = opts[:actual].to_f.clamp(0.0, 1.0)

data/lib/pwn/ai/agent/extrospection.rb

@@ -1722,14 +1722,24 @@ module PWN
           c = opts[:claim].to_s
           return :doc     if opts[:url] || c =~ URI::DEFAULT_PARSER.make_regexp(%w[http https])
           return :cve     if c =~ /CVE-\d{4}-\d{4,}/i
-          return :version if c =~ /\bv?\d+\.\d+(?:\.\d+)?\b/
+          # P26 — full semver (x.y.z) OR "latest X is [v]N.N" product claims.
+          # Bare two-part floats alone stay :generic (metrics like "cap 0.2").
+          return :version if c =~ /\b[A-Za-z][\w.+-]{2,}\s+v?\d+\.\d+\.\d+\b/
+          return :version if c =~ /\blatest\b.+\bis\b\s+v?\d+\.\d+/i
 
           :generic
         end
 
         private_class_method def self.commit_verdict(opts = {})
-          claim = opts[:claim]
+          claim = opts[:claim].to_s
           ev    = Array(opts[:evidence]).first
+          # P26 — never pollute learning.jsonl with :unknown on metric crumbs
+          # or sub-8-char blobs; those are not human-reviewable world claims.
+          junk_unknown = opts[:verdict].to_s == 'unknown' && (
+            claim.length < 12 ||
+            claim.match?(/\A(?:cap|share|proxy|judge|success|only|now|clears|gap|score|rate|mean|brier|overconf|distrust|trajectory_fraction|handler|orm|prm|delta|limit|window|pct|percent|ms|iter|budget|conf|confidence|ruby|python|linux|kernel|host|cwd|e\.g)\b/i) ||
+            (claim.match?(/\d+\.\d+/) && !claim.match?(/CVE-|\d+\.\d+\.\d+/i))
+          )
           case opts[:verdict]
           when :refuted
             Mistakes.record(tool: 'assumption', error: "REFUTED (extro_verify #{opts[:kind]}, conf=#{opts[:confidence].round(2)}): #{claim}", args: ev&.dig(:final_url), source: :model) if defined?(Mistakes)
@@ -1739,8 +1749,12 @@ module PWN
             observe(source: 'extro_verify', category: :intel, target: ev&.dig(:final_url), data: claim, tags: %w[verify confirmed], ttl: 30 * 24 * 3600)
             :extro_observe
           else
-            Learning.note_outcome(task: "extro_verify: #{claim[0, 120]}", success: false, details: 'verdict :unknown - needs human review', tags: %w[needs_human extro_verify]) if defined?(Learning) && Learning.respond_to?(:note_outcome)
-            :learning_note
+            if junk_unknown
+              :skipped_junk
+            else
+              Learning.note_outcome(task: "extro_verify: #{claim[0, 120]}", success: false, details: 'verdict :unknown - needs human review', tags: %w[needs_human extro_verify]) if defined?(Learning) && Learning.respond_to?(:note_outcome)
+              :learning_note
+            end
           end
         rescue StandardError
           nil