Module: PWN::AI::Agent::Reward

Defined in:
lib/pwn/ai/agent/reward.rb

Overview

PWN::AI::Agent::Reward is the OUTCOME reward model for the pwn-ai reinforcement-learning loop. It replaces the regex-proxy reward that previously drove Learning.infer_success / Loop.record_metrics with four calibrated signals:

R1  .judge      — LLM Outcome Reward Model (ORM). Scores the FINAL
                answer against the user request → {score:0..1,
                verdict: :solved|:partial|:wrong|:refused,
                rationale:}. Scalar, not boolean.
R2  .prm        — Process Reward Model. Per-tool-call "did this
                step advance toward the goal?" → step_reward
                tagged onto every Sessions entry so credit is
                assignable INSIDE a trajectory, not just at its
                boundary. First PRM applied to security tooling.
R3  .sentinel   — Reward-hacking detector. Tracks proxy vs judge
                vs (1 - user_correction_rate); when they diverge
                by > SENTINEL_GAP the reward signal itself is
                fingerprinted as a Mistake so the operator sees
                "your success_rate is a lie" in KNOWN MISTAKES.
R4  .semantic_ok — Structured tool-result classifier. Knows that
                `grep exit 1` == "no match", not "failure";
                kills the phantom-mistake class (31f1871b8a15)
                that made the loop's #1 negative signal a false
                positive it created itself.

Reward also owns the PREFERENCE-PAIR ledger (~/.pwn/preferences.jsonl) that turns pwn's naturally-generated (rejected, chosen) pairs — from user corrections, mistakes_resolve, and Curriculum.counterfactual A/B branches — into a DPO export (W1). This is the ONLY path from in-context learning to weight-level policy improvement.

E3 .verify_as_reward — grounds any final containing a checkable claim (CVE / version / cited URL) via Extrospection.verify and maps the browser verdict onto the reward scalar. Hallucination becomes a measurable −reward, not just a warning.

Everything degrades gracefully: when module_reflection is off (no LLM judge available) .judge falls back to a calibrated heuristic over .semantic_ok + .verify_as_reward + Mistakes correction rate, which is STILL strictly better than the old regex.

Constant Summary collapse

PREFERENCES_FILE = 
File.join(Dir.home, '.pwn', 'preferences.jsonl')
SENTINEL_FILE = 
File.join(Dir.home, '.pwn', 'reward_sentinel.json')
DPO_DIR = 
File.join(Dir.home, '.pwn', 'finetune')
SENTINEL_GAP = 
0.15
SENTINEL_WINDOW = 
40
VERDICTS = 
{
  solved: 1.0, confirmed: 1.0, partial: 0.5,
  unknown: 0.5, wrong: 0.0, refused: 0.0, refuted: 0.0
}.freeze
BENIGN_EXIT =

Commands whose non-zero exit is INFORMATIONAL, not a failure. The regex-proxy treating these as failures was the single largest source of noise in Mistakes/Metrics (grep exit 1 = "no match").

{
  /\b(?:e|f|z|rip|p)?grep\b/ => [1],
  /\bdiff\b/ => [1],
  /\bcmp\b/ => [1],
  /\btest\b|\[\s/ => [1],
  /\bls\b/ => [1, 2],
  /\bfind\b/ => [1],
  /\bwhich\b|\bcommand -v\b/ => [1],
  /\bpidof\b|\bpgrep\b|\bpkill\b/ => [1],
  /\bxargs\b/ => [123],
  /\btimeout\b/ => [124],
  /\bcurl\b/ => [22],
  /\brubocop\b/ => [1]
}.freeze
JUDGE_SYSTEM = 
<<~SYS
  You are the pwn-ai Outcome Reward Model. Given a USER REQUEST, the
  agent's FINAL ANSWER, and a compressed TOOL TRACE, emit ONE line of
  strict JSON:
    {"score": <0.0-1.0>, "verdict": "solved|partial|wrong|refused",
     "rationale": "<≤140 chars>", "key_step": <int|-1>}
  score=1.0 only when the final DEMONSTRABLY satisfies the request
  (evidence in trace). score=0.5 for correct-direction-but-incomplete.
  score=0.0 for hallucinated / off-goal / refused. key_step is the
  1-indexed trace line most responsible for the outcome (credit
  assignment), or -1 if none. Output JSON ONLY.
SYS
PRM_SYSTEM = 
<<~SYS
  You are the pwn-ai Process Reward Model. For EACH numbered tool
  step, output one integer per line: 1 (advanced toward the goal),
  0 (neutral / exploratory), -1 (regressed / wasted). Output ONLY
  the integers, one per line, same count as steps. No prose.
SYS
TRAJECTORY_SHAPES =

Trajectory-shaped chosen sides that may land DPO without prose flood.

%w[winning_trace revised_answer real_dispatch].freeze
WRITE_SOURCE_CAP =

P9 — write-time source quota (not only export). Prefer diverse online generators over resolve-prose flood. Window is last WRITE_SOURCE_WINDOW pairs; a source already above WRITE_SOURCE_CAP is refused unless force: true (user_correction always forces).

0.40
WRITE_SOURCE_WINDOW = 
100
TARGET_SOURCE_MIX =

P0 — target online generator mix for W1. Gates alone cannot fill an empty promote; the controller must prefer underfilled sources (counterfactual / critic / curriculum / user_correction) when resolve already dominates. Shares are soft targets, not hard caps (hard cap remains WRITE_SOURCE_CAP). Trajectory-only still applies.

{
  'mistakes_resolve' => 0.30,
  'curriculum' => 0.25,
  'counterfactual' => 0.20,
  'critic' => 0.15,
  'user_correction' => 0.10
}.freeze
DPO_SOURCE_CAP =

Max share any single preference source may occupy in a DPO export. Without this cap, mistakes_resolve monoculture (often >80%) teaches the LoRA "emit fix prose" instead of trajectory preference (P5 enforce).

0.40

Class Method Summary collapse

Class Method Details

.authors ⇒ Object

Author(s)

0day Inc. support@0dayinc.com



1297
1298
1299
 
# File 'lib/pwn/ai/agent/reward.rb', line 1297

public_class_method def self.authors
  "AUTHOR(S):\n  0day Inc. <support@0dayinc.com>\n"
end

.clear_proxy_distrust ⇒ Object 



319
320
321
322
323
324
325
326
327
328
 
# File 'lib/pwn/ai/agent/reward.rb', line 319

public_class_method def self.clear_proxy_distrust
  s = load_sentinel
  return if s[:proxy_distrust].to_f <= 0.0

  s[:proxy_distrust] = 0.0
  s[:distrust_cleared_at] = Time.now.utc.iso8601
  atomic_write(path: SENTINEL_FILE, body: JSON.generate(s))
rescue StandardError
  nil
end

.export_dpo(opts = {}) ⇒ Object 



861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
 
# File 'lib/pwn/ai/agent/reward.rb', line 861

public_class_method def self.export_dpo(opts = {})
  fmt = (opts[:format] || :dpo).to_sym
  FileUtils.mkdir_p(DPO_DIR)
  out = opts[:out] || File.join(DPO_DIR, "pwn-dpo-#{Time.now.utc.strftime('%Y%m%d')}.jsonl")
  rows = preferences(limit: 100_000)
  # P15 — drop weak geometry before source-cap so resolve prose cannot
  # dominate the kept set after balance. opt-out with scrub: false.
  scrub = opts.key?(:scrub) ? opts[:scrub] : true
  geometry_dropped = 0
  if scrub
    usable = rows.select { |r| usable_preference?(row: r) }
    geometry_dropped = rows.length - usable.length
    rows = usable
  end
  # P5 — downsample so no single source exceeds DPO_SOURCE_CAP of the export.
  # opt-out with balance: false (raw dump for diagnostics).
  balance = opts.key?(:balance) ? opts[:balance] : true
  selected = balance ? balance_preference_rows(rows: rows, cap: (opts[:source_cap] || DPO_SOURCE_CAP).to_f) : rows
  dropped = rows.length - selected.length
  File.open(out, 'w') do |f|
    selected.each do |r|
      line = case fmt
             when :kto
               [{ prompt: r[:prompt], completion: r[:chosen], label: true },
                { prompt: r[:prompt], completion: r[:rejected], label: false }]
             else
               # Keep source for auditability / preference_balance post-export.
               { prompt: r[:prompt], chosen: r[:chosen], rejected: r[:rejected], source: r[:source] }
             end
      (line.is_a?(Array) ? line : [line]).each { |l| f.puts(JSON.generate(l)) }
    end
  end
  by_src = selected.group_by { |r| r[:source].to_s }.transform_values(&:length)
  {
    path: out, format: fmt, pairs: selected.length, bytes: File.size(out),
    balanced: balance, dropped: dropped, geometry_dropped: geometry_dropped,
    scrubbed: scrub, by_source: by_src,
    source_cap: balance ? (opts[:source_cap] || DPO_SOURCE_CAP).to_f : nil,
    preference_balance: begin
      preference_balance(limit: 10_000, scrub: true)
    rescue StandardError
      nil
    end
  }
end

.generator_mix(opts = {}) ⇒ Object

P0 — online generator mix report + urgency flags. Controllers (auto_introspect, practice, counterfactual gate) consult this so underfilled sources get scheduling priority while over-cap resolve stops flooding. Returns trajectory_fraction, urgent:[], suppress:[], healthy:.



612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
 
# File 'lib/pwn/ai/agent/reward.rb', line 612

public_class_method def self.generator_mix(opts = {})
  limit = opts[:limit] || WRITE_SOURCE_WINDOW
  rows = preferences(limit: limit)
  usable = rows.select { |r| usable_preference?(row: r) }
  by = Hash.new(0)
  usable.each { |r| by[r[:source].to_s] += 1 }
  n = usable.length
  shares = {}
  TARGET_SOURCE_MIX.each_key { |k| shares[k] = n.zero? ? 0.0 : (by[k].to_f / n).round(3) }
  by.each_key { |k| shares[k] ||= (by[k].to_f / n).round(3) }

  traj_n = usable.count { |r| TRAJECTORY_SHAPES.include?(r[:shape].to_s) }
  traj_f = n.zero? ? 0.0 : (traj_n.to_f / n).round(3)

  urgent = []
  suppress = []
  TARGET_SOURCE_MIX.each do |src, target|
    sh = shares[src].to_f
    urgent << src if sh < (target * 0.5) && n >= 5
    suppress << src if sh > WRITE_SOURCE_CAP && n >= 10
  end
  suppress << 'mistakes_resolve' if shares['mistakes_resolve'].to_f > WRITE_SOURCE_CAP && n >= 10 && !suppress.include?('mistakes_resolve')

  healthy = urgent.empty? && suppress.empty? && traj_f >= 0.5 && n >= 10
  {
    n: n,
    raw_n: rows.length,
    by_source: by,
    shares: shares,
    targets: TARGET_SOURCE_MIX,
    trajectory_fraction: traj_f,
    urgent: urgent.uniq,
    suppress: suppress.uniq,
    healthy: healthy,
    recommendation: if healthy
                      'mix_ok'
                    elsif n < 10
                      'need_more_pairs'
                    elsif traj_f < 0.5
                      'need_trajectory_shape'
                    elsif urgent.any?
                      "boost:#{urgent.join(',')}"
                    else
                      "suppress:#{suppress.join(',')}"
                    end
  }
rescue StandardError => e
  {
    n: 0, healthy: false, error: "#{e.class}: #{e.message}",
    urgent: %w[curriculum counterfactual critic user_correction],
    suppress: []
  }
end

.help ⇒ Object

Display Usage for this Module



1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
 
# File 'lib/pwn/ai/agent/reward.rb', line 1303

public_class_method def self.help
  puts <<~USAGE
    USAGE:
      # Tier 1 — reward signal
      PWN::AI::Agent::Reward.judge(request: req, final: text, session_id: sid)     # R1 ORM → {score:, verdict:, rationale:}
      PWN::AI::Agent::Reward.prm(request: req, session_id: sid)                    # R2 PRM → per-step credit
      PWN::AI::Agent::Reward.sentinel                                              # R3 reward-hacking detector
      PWN::AI::Agent::Reward.reset_sentinel                                        # wipe corrupt window + distrust
      PWN::AI::Agent::Reward.warm_sentinel                                         # P10 fill R3 window from Learning outcomes
      PWN::AI::Agent::Reward.semantic_ok(name: 'shell', raw: json, args: args)     # R4 kills phantom exit≠0 mistakes

      # Tier 5 — preference pairs → DPO
      PWN::AI::Agent::Reward.record_preference(prompt: p, rejected: r, chosen: c, source: :user_correction)
      PWN::AI::Agent::Reward.preferences(limit: 100)
      PWN::AI::Agent::Reward.export_dpo(format: :dpo)                              # W1 → ~/.pwn/finetune/pwn-dpo-*.jsonl (≤40%/source, scrubbed)
      PWN::AI::Agent::Reward.export_dpo(format: :dpo, balance: false)              # raw dump (diagnostics)
      PWN::AI::Agent::Reward.scrub_preferences(dry_run: true)                      # P15 ledger hygiene report
      PWN::AI::Agent::Reward.scrub_preferences                                      # P15 rewrite jsonl (backup first)
      PWN::AI::Agent::Reward.preference_balance(scrub: true)                       # P15 geometry-aware mix

      # Tier 6 — grounded reward
      PWN::AI::Agent::Reward.verify_as_reward(final: text)                         # E3 browser-verified reward

      Config (PWN::Env[:ai][:agent]):
        :verify_as_reward   - Boolean/nil, ground finals via extro_verify (nil=auto)
        :reward_llm         - Boolean/nil, force ORM/PRM LLM teacher (nil=on for remote engines)

      #{self}.authors
  USAGE
end

.infer_shape(opts = {}) ⇒ Object

P0 ops — infer trajectory shape for legacy ledger rows that predate P21/P25 shape tags. Used by scrub_preferences rewrite so generator_mix trajectory_fraction reflects content, not missing keys.



669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
 
# File 'lib/pwn/ai/agent/reward.rb', line 669

public_class_method def self.infer_shape(opts = {})
  r = opts.is_a?(Hash) && opts.key?(:row) ? opts[:row] : opts
  r = r.transform_keys(&:to_sym) if r.respond_to?(:transform_keys)
  existing = r[:shape].to_s
  return existing if TRAJECTORY_SHAPES.include?(existing) || existing == 'fix_prose'

  chosen = r[:chosen].to_s
  source = r[:source].to_s
  # tool-call / trace markers → winning_trace
  if chosen.match?(/\b(shell|pwn_eval|memory_|sessions_|reward_|curriculum_|extro_|mistakes_)\b/i) &&
     (chosen.include?('→') || chosen.include?('tool_call') || chosen.include?('"name"') ||
      chosen.lines.count { |l| l.strip.start_with?('{') || l.include?('arguments') } >= 1)
    return 'winning_trace'
  end
  # long revised answer from critic / user / CF → revised_answer
  return 'revised_answer' if chosen.length >= 200 && %w[critic user_correction counterfactual curriculum].include?(source)

  # counterfactual real dispatch tag in meta
  meta = r[:meta].is_a?(Hash) ? r[:meta] : {}
  return 'real_dispatch' if meta[:mode].to_s == 'real_dispatch' || meta['mode'].to_s == 'real_dispatch'

  existing.empty? ? nil : existing
rescue StandardError
  nil
end

.judge(opts = {}) ⇒ Object

Supported Method Parameters

v = PWN::AI::Agent::Reward.judge( request: 'required - original user request', final: 'required - assistant final answer', session_id: 'optional - PWN::Sessions id (adds tool trace)', trace: 'optional - Array of tool-result strings (overrides session_id)', commit: 'optional - write score into learning.jsonl / sentinel (default true)' )



114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
 
# File 'lib/pwn/ai/agent/reward.rb', line 114

public_class_method def self.judge(opts = {})
  request = opts[:request].to_s
  final   = opts[:final].to_s
  trace   = Array(opts[:trace])
  trace   = load_trace(session_id: opts[:session_id]) if trace.empty? && opts[:session_id]
  commit  = opts.key?(:commit) ? opts[:commit] : true

  v = llm_judge(request: request, final: final, trace: trace)
  v ||= heuristic_judge(request: request, final: final, trace: trace)

  # P1 — local/heuristic calibration: thin judges must not be treated
  # as ground truth when proxy_distrust is already high. Two levers:
  #   (1) low :confidence so Metrics.effective_rate haircuts blend
  #       weight (distrust × confidence) instead of replacing proxy;
  #   (2) score-path caps only for decisive failure floors and for
  #       local no-trace highs (false "solved"). Do NOT pull known
  #       wrong (0.0 failure-language) toward 0.5, and do NOT deflate
  #       tool-backed heuristic scores that already cleared the bar —
  #       confidence handles that in the bandit blend.
  eng = (PWN::Env.dig(:ai, :active) if defined?(PWN::Env)).to_s.downcase
  local = eng == 'ollama' || eng.empty?
  if v[:source].to_s == 'heuristic'
    v[:confidence] = local ? 0.35 : 0.5
    raw = v[:score].to_f
    v[:score_raw] = raw
    # preserve empty / failure-language / polite floors exactly (raw <= 0.15)
    # and pass-through non-capped heuristics via the default arm.
    if local && raw > 0.15 && trace.empty? && raw >= 0.6 && final.length < 400
      v[:score] = [raw, 0.45].min
      v[:rationale] = "#{v[:rationale]} | P1:local_no_trace_cap"
    elsif local && raw > 0.15 && trace.length < 2 && raw >= 0.85
      # thin-evidence local highs: mild shrink toward 0.5
      v[:score] = (0.5 + ((raw - 0.5) * 0.7)).round(3).clamp(0.0, 1.0)
    else
      v[:score] = raw
    end
    v[:verdict] = if v[:score] >= 0.6 then :solved
                  elsif v[:score] >= 0.3 then :partial
                  else :wrong
                  end
  else
    v[:confidence] ||= 0.85
  end

  ground = verify_as_reward(final: final)
  unless ground.nil?
    # Ground-truth override: a browser-refuted claim caps score at
    # 0.2 regardless of how confident the judge was; a confirmed
    # claim floors it at 0.6. E3.
    v[:score] = [v[:score], 0.2].min if ground[:verdict] == :refuted
    v[:score] = [v[:score], 0.6].max if ground[:verdict] == :confirmed
    v[:grounded] = ground
    v[:confidence] = [v[:confidence].to_f, ground[:confidence].to_f].max if ground[:confidence]
  end

  v[:success] = v[:score] >= 0.6
  v[:engine] = eng
  # P1 — sentinel stores confidence so distrust math can haircut
  # heuristic-heavy windows differently from LLM ORM windows.
  record_sentinel(proxy: opts[:proxy_ok], judge: v[:score], confidence: v[:confidence]) if commit
  v
rescue StandardError => e
  { score: 0.5, verdict: :unknown, rationale: "judge error: #{e.class}", success: !final.strip.empty?, error: e.message, confidence: 0.2, source: :error }
end

.preference_balance(opts = {}) ⇒ Object

P15/P5 — geometry-aware source mix. scrub:true uses usable_preference? so operators see the post-hygiene diet (what export_dpo will train on).



779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
 
# File 'lib/pwn/ai/agent/reward.rb', line 779

public_class_method def self.preference_balance(opts = {})
  limit = opts[:limit] || 10_000
  scrub = opts.key?(:scrub) ? opts[:scrub] : false
  rows = preferences(limit: limit)
  before = rows.length
  rows = rows.select { |r| usable_preference?(row: r) } if scrub
  by = Hash.new(0)
  by_shape = Hash.new(0)
  rows.each do |r|
    by[r[:source].to_s] += 1
    sh = r[:shape].to_s
    sh = 'unspecified' if sh.empty?
    by_shape[sh] += 1
  end
  total = rows.length
  frac = by.transform_values { |n| total.zero? ? 0.0 : (n.to_f / total).round(3) }
  shape_frac = by_shape.transform_values { |n| total.zero? ? 0.0 : (n.to_f / total).round(3) }
  traj_n = rows.count { |r| TRAJECTORY_SHAPES.include?(r[:shape].to_s) }
  traj_frac = total.zero? ? 0.0 : (traj_n.to_f / total).round(3)
  monoculture = total.positive? && (by.values.max.to_f / total) > 0.7
  mix = begin
    generator_mix(limit: limit)
  rescue StandardError
    nil
  end
  {
    total: before,
    kept: total,
    scrubbed: scrub,
    dropped: before - total,
    by_source: by,
    fractions: frac,
    by_shape: by_shape,
    by_shape_fraction: shape_frac,
    trajectory_fraction: traj_frac,
    monoculture: monoculture,
    generator_mix: mix,
    advice: if total < 12
              'W1 thin: need more trajectory-shaped pairs before LoRA promote.'
            elsif monoculture
              'W1 monoculture: run Reward.scrub_preferences; enable :counterfactual/:critic; stop resolve-prose flood.'
            elsif traj_frac < 0.30
              'W1 geometry weak: <30% trajectory-shaped chosen sides — DPO would teach commentary.'
            elsif mix && !mix[:healthy]
              "W1 generator mix: #{mix[:recommendation]}"
            else
              'W1 source mix OK for gated export'
            end
  }
rescue StandardError => e
  { error: "#{e.class}: #{e.message}" }
end

.preferences(opts = {}) ⇒ Object

Supported Method Parameters

rows = PWN::AI::Agent::Reward.preferences(limit: 500, source: nil)



835
836
837
838
839
840
841
842
843
844
845
846
847
848
 
# File 'lib/pwn/ai/agent/reward.rb', line 835

public_class_method def self.preferences(opts = {})
  limit  = opts[:limit] || 500
  source = opts[:source].to_s
  return [] unless File.exist?(PREFERENCES_FILE)

  rows = File.readlines(PREFERENCES_FILE).map do |l|
    JSON.parse(l, symbolize_names: true)
  rescue StandardError
    nil
  end
  rows.compact!
  rows.select! { |r| r[:source] == source } unless source.empty?
  rows.reverse.first(limit)
end

.prm(opts = {}) ⇒ Object

Supported Method Parameters

steps = PWN::AI::Agent::Reward.prm( request: 'required - user goal', session_id: 'optional - session to score in place', trace: 'optional - Array of args:, result: or Strings' )

Returns [step:, reward: -1|0|1, ...] and, when session_id is given, rewrites each tool line in the transcript with a [step_reward=N] prefix so exemplars_for / distill_skill can keep only reward>0 steps (C4 minimal sufficient trace).



195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
 
# File 'lib/pwn/ai/agent/reward.rb', line 195

public_class_method def self.prm(opts = {})
  request = opts[:request].to_s
  trace   = Array(opts[:trace])
  sid     = opts[:session_id]
  trace   = load_trace(session_id: sid) if trace.empty? && sid

  rewards = llm_prm(request: request, trace: trace)
  rewards ||= heuristic_prm(trace: trace)

  out = trace.each_with_index.map do |s, i|
    { idx: i + 1, step: s.to_s[0, 200], reward: rewards[i] || 0 }
  end
  annotate_session(session_id: sid, rewards: rewards) if sid
  out
rescue StandardError
  []
end

.proxy_distrust ⇒ Object

P4 — scalar 0.0..1.0 haircut applied to Metrics success / Registry β when the proxy is lying. 0.0 = trust proxy fully; 1.0 = ignore proxy rates.



284
285
286
287
288
289
290
291
292
293
294
295
 
# File 'lib/pwn/ai/agent/reward.rb', line 284

public_class_method def self.proxy_distrust
  s = load_sentinel
  d = s[:proxy_distrust].to_f
  # auto-expire after 7d without refresh so a one-off gap doesn't stick
  if s[:distrust_at]
    age = Time.now.utc - Time.parse(s[:distrust_at].to_s)
    return 0.0 if age > 7 * 86_400
  end
  d.clamp(0.0, 1.0)
rescue StandardError
  0.0
end

.record_preference(opts = {}) ⇒ Object 



532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
 
# File 'lib/pwn/ai/agent/reward.rb', line 532

public_class_method def self.record_preference(opts = {})
  prompt   = opts[:prompt].to_s
  rejected = opts[:rejected].to_s
  chosen   = opts[:chosen].to_s
  return nil if prompt.strip.empty? || chosen.strip.empty? || rejected.strip.empty?
  return nil if chosen.strip == rejected.strip

  # Reject weak pair geometry: CORRECTION: flaw-prose is not a trajectory.
  return { skipped: :weak_pair_geometry, reason: 'chosen looks like flaw prose, not a revised answer/trace' } if chosen.match?(/\A\s*CORRECTION:\s*/i) && chosen.length < 400 && !opts[:force]

  source = (opts[:source] || :unknown).to_s
  shape  = opts[:shape].to_s
  # P25 — require trajectory shape at write time unless force / user_correction.
  # Stops resolve-prose flood from ever landing in the ledger; export scrub
  # is defense-in-depth, not the primary gate.
  traj = TRAJECTORY_SHAPES.include?(shape)
  # P25 — non-trajectory prose never lands (export scrub is defense-in-depth).
  # user_correction and explicit force: still allowed for human / migration paths.
  unless traj || opts[:force] || source == 'user_correction'
    return {
      skipped: :non_trajectory_shape,
      reason: "shape=#{shape.inspect} not in #{TRAJECTORY_SHAPES.join(',')}; pass force:true or a trajectory shape",
      source: source
    }
  end
  # P9 — write-time source quota still applies to trajectory pairs.
  # P25 made every auto-written row trajectory-shaped; if traj also
  # bypassed the quota, resolve monoculture would return via winning_trace
  # flood. Only user_correction and explicit force:true skip the cap.
  bypass_quota = opts[:force] || source == 'user_correction'
  unless bypass_quota
    quota = write_source_quota(source: source)
    return quota.merge(skipped: :source_quota) if quota[:over_cap]
  end

  entry = {
    id: Digest::SHA256.hexdigest("#{prompt}|#{rejected}|#{chosen}")[0, 12],
    prompt: prompt[0, 4_000],
    rejected: rejected[0, 4_000],
    chosen: chosen[0, 4_000],
    source: source,
    engine: (PWN::Env.dig(:ai, :active) if defined?(PWN::Env)).to_s,
    timestamp: Time.now.utc.iso8601
  }
  entry[:meta] = opts[:meta] if opts[:meta].is_a?(Hash)
  entry[:shape] = opts[:shape].to_s if opts[:shape]
  FileUtils.mkdir_p(File.dirname(PREFERENCES_FILE))
  File.open(PREFERENCES_FILE, 'a') { |f| f.puts(JSON.generate(entry)) }
  entry
end

.recoverable_shape(opts = {}) ⇒ Object

2.2 — coarse recoverable shape beside the fingerprint. Paths are normalised away for counting; shape stays for repair routing (enoent → install/check path; exit127 → missing binary; …).



439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
 
# File 'lib/pwn/ai/agent/reward.rb', line 439

public_class_method def self.recoverable_shape(opts = {})
  err = "#{opts[:err]} #{opts[:stderr]}".downcase
  ec  = opts[:exit_code]
  return :exit127 if ec == 127 || err.include?('command not found')
  return :exit126 if ec == 126
  return :enoent if err.match?(/no such file|enoent|cannot access|not a directory/)
  return :eacces if err.match?(/permission denied|eacces|operation not permitted/)
  return :auth_required if err.match?(/auth|unauthorized|401|403|forbidden|login required|api.?key/)
  return :timeout if ec == 124 || err.include?('timed out') || err.include?('timeout')
  return :network if err.match?(/connection refused|name or service not known|could not resolve|network is unreachable/)
  return :syntax if err.match?(/syntax error|parse error|unexpected token|json::parser/)
  return :nonzero_exit if ec && ec != 0
  return :handler_error if err.strip.length.positive?

  :unknown
end

.reset ⇒ Object 



907
908
909
910
911
 
# File 'lib/pwn/ai/agent/reward.rb', line 907

public_class_method def self.reset
  FileUtils.rm_f(PREFERENCES_FILE)
  FileUtils.rm_f(SENTINEL_FILE)
  { cleared: true }
end

.reset_sentinel ⇒ Object

One-shot: wipe sentinel window + distrust after deploying the ring-buffer arithmetic (or any time the live file is known-corrupt). Does NOT touch preferences / DPO exports (unlike .reset).



333
334
335
336
 
# File 'lib/pwn/ai/agent/reward.rb', line 333

public_class_method def self.reset_sentinel
  FileUtils.rm_f(SENTINEL_FILE)
  { cleared: true, path: SENTINEL_FILE }
end

.scrub_preferences(opts = {}) ⇒ Object

P15 — one-shot ledger hygiene. Filters in place (rewrite jsonl) or report-only. Returns after:, dropped:, by_reason:, path:.



724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
 
# File 'lib/pwn/ai/agent/reward.rb', line 724

public_class_method def self.scrub_preferences(opts = {})
  dry = opts.key?(:dry_run) ? opts[:dry_run] : false
  path = PREFERENCES_FILE
  return { before: 0, after: 0, dropped: 0, dry_run: dry, path: path } unless File.exist?(path)

  raw = File.readlines(path)
  kept = []
  reasons = Hash.new(0)
  raw.each do |line|
    begin
      r = JSON.parse(line, symbolize_names: true)
    rescue StandardError
      reasons[:parse_error] += 1
      next
    end
    if usable_preference?(row: r)
      # P0 ops — backfill shape so trajectory_fraction is meaningful
      if r[:shape].to_s.empty?
        inferred = infer_shape(row: r)
        r = r.merge(shape: inferred) if inferred
      end
      kept << r
    else
      why = if r[:chosen].to_s.match?(/\A\s*CORRECTION:\s*/i)
              :correction_prose
            elsif r[:shape].to_s == 'fix_prose'
              :fix_prose
            elsif r[:chosen].to_s.length < (r[:rejected].to_s.length * 0.25)
              :chosen_too_short
            else
              :weak_geometry
            end
      reasons[why] += 1
    end
  end
  unless dry
    bak = "#{path}.bak-p15-#{Time.now.utc.strftime('%Y%m%d%H%M%S')}"
    FileUtils.cp(path, bak)
    File.open(path, 'w') { |f| kept.each { |r| f.puts(JSON.generate(r)) } }
  end
  {
    before: raw.length,
    after: kept.length,
    dropped: raw.length - kept.length,
    by_reason: reasons,
    dry_run: dry,
    path: path,
    backup: dry ? nil : bak
  }
rescue StandardError => e
  { error: "#{e.class}: #{e.message}" }
end

.semantic_ok(opts = {}) ⇒ Object

Supported Method Parameters

h = PWN::AI::Agent::Reward.semantic_ok( name: 'required - tool name', raw: 'required - JSON string returned by Dispatch.call', args: 'optional - the tool call arguments (used for BENIGN_EXIT)' )

Returns { ok:, semantic_ok:, exit:, err:, benign: }. :ok is the old proxy (handler didn't raise); :semantic_ok additionally knows that grep/diff/find exit≠0 with empty stderr is not a failure. Loop.run records Metrics on :ok but only records Mistakes on !semantic_ok.



406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
 
# File 'lib/pwn/ai/agent/reward.rb', line 406

public_class_method def self.semantic_ok(opts = {})
  name = opts[:name].to_s
  raw  = opts[:raw].to_s
  ok   = raw.include?('"success":true')
  err  = raw[/"error":"([^"]{1,300})"/, 1]
  exit_code = raw[/"exit":(\d+)/, 1]&.to_i
  stderr    = raw[/"stderr":"([^"]{0,400})"/, 1].to_s

  benign = false
  shape  = nil
  if name == 'shell' && ok && exit_code && exit_code != 0
    cmd = extract_cmd(args: opts[:args])
    # 2.1 — ONLY BENIGN_EXIT regex × allowed codes. The old global
    # `stderr.empty? && exit==1 ⇒ benign` laundered real failures
    # (pipelines without pipefail, bare false, etc.) into "success".
    # For pipelines, match the LAST stage (post-pipe) first, then any.
    stages = cmd.split('|').map(&:strip)
    last   = stages.last.to_s
    benign = BENIGN_EXIT.any? { |rx, codes| last.match?(rx) && codes.include?(exit_code) }
    benign ||= stages.length > 1 && BENIGN_EXIT.any? { |rx, codes| stages.any? { |s| s.match?(rx) } && codes.include?(exit_code) && stderr.strip.empty? }
    shape = recoverable_shape(exit_code: exit_code, stderr: stderr, err: err)
  elsif !ok
    shape = recoverable_shape(exit_code: exit_code, stderr: stderr, err: err || raw[0, 200])
  end

  semantic = ok && (exit_code.nil? || exit_code.zero? || benign)
  err ||= raw[/"stderr":"([^"]{4,300})"/, 1] unless semantic
  { ok: ok, semantic_ok: semantic, exit: exit_code, err: err, benign: benign, shape: shape }
end

.sentinel ⇒ Object

Supported Method Parameters

r = PWN::AI::Agent::Reward.sentinel



220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
 
# File 'lib/pwn/ai/agent/reward.rb', line 220

public_class_method def self.sentinel
  s = normalize_sentinel(raw: load_sentinel)
  window = s[:window]
  n = window.length
  return { samples: n, status: :insufficient } if n < SENTINEL_WINDOW

  means = window_means(window: window)
  proxy = means[:proxy]
  judge = means[:judge]
  # Refuse to act on corrupt arithmetic — proxy must be a rate in [0,1].
  if proxy.nil? || proxy < 0.0 || proxy > 1.0
    return {
      samples: n,
      status: :corrupt_proxy,
      proxy: proxy,
      judge: judge&.round(3),
      reward_hacked: false,
      proxy_distrust: proxy_distrust
    }
  end

  human = 1.0 - user_correction_rate
  gap_pj = (proxy - judge).abs
  gap_ph = (proxy - human).abs
  hacked = gap_pj > SENTINEL_GAP || gap_ph > SENTINEL_GAP
  if hacked
    # 1.1 — freeze auto-Mistakes.record on tool:reward_signal after the
    # first open sig per gap-bucket. Endless ×13 fingerprints were
    # the loudest scar in every prompt and taught nothing. Open a
    # calibration path instead; park the sig as needs_code_change.
    bucket = "gap_pj=#{gap_pj.round(2)}|gap_ph=#{gap_ph.round(2)}"
    open_sig = defined?(Mistakes) ? Mistakes.for_tool(tool: 'reward_signal', unresolved_only: true) : []
    if open_sig.empty? && defined?(Mistakes)
      m = Mistakes.record(
        tool: 'reward_signal',
        error: "proxy success_rate #{proxy.round(2)} diverges from judge #{judge.round(2)} / human #{human.round(2)} by >#{SENTINEL_GAP}",
        source: :model,
        needs_code_change: true,
        meta: { bucket: bucket, proxy: proxy, judge: judge, human: human }
      )
      Mistakes.park(signature: m[:signature], reason: 'reward_signal needs calibration, not practice') if m && Mistakes.respond_to?(:park)
    end
    Curriculum.calibrate(predicted: proxy, actual: judge, engine: :reward_sentinel) if defined?(Curriculum) && Curriculum.respond_to?(:calibrate)
    # P4 — make sentinel ACTIONABLE: persist a distrust factor so
    # Metrics.to_context / Registry.rank haircut proxy success instead of
    # just opening another Mistakes row the model learns to ignore.
    set_proxy_distrust(gap: [gap_pj, gap_ph].max, proxy: proxy, judge: judge)
  else
    clear_proxy_distrust
  end
  {
    samples: n,
    proxy: proxy.round(3),
    judge: judge.round(3),
    human: human.round(3),
    gap_proxy_judge: gap_pj.round(3),
    gap_proxy_human: gap_ph.round(3),
    reward_hacked: hacked,
    proxy_distrust: proxy_distrust
  }
end

.set_proxy_distrust(opts = {}) ⇒ Object 



297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
 
# File 'lib/pwn/ai/agent/reward.rb', line 297

public_class_method def self.set_proxy_distrust(opts = {})
  s = normalize_sentinel(raw: load_sentinel)
  gap = opts[:gap].to_f
  proxy = opts[:proxy]
  # Guard: never set distrust from a nonsensical proxy (pre-ring-buffer
  # decay×to_i bug produced means ≫ 1.0 and hard-pegged distrust at 1.0).
  unless proxy.nil?
    pf = proxy.to_f
    return s[:proxy_distrust].to_f if pf < 0.0 || pf > 1.0
  end
  # map gap 0.15→0.4, 0.30→0.8, ≥0.40→1.0
  factor = ((((gap - SENTINEL_GAP) / SENTINEL_GAP) * 0.4) + 0.4).clamp(0.3, 1.0)
  s[:proxy_distrust] = factor
  s[:distrust_at] = Time.now.utc.iso8601
  s[:distrust_meta] = { proxy: opts[:proxy], judge: opts[:judge], gap: gap }
  FileUtils.mkdir_p(File.dirname(SENTINEL_FILE))
  atomic_write(path: SENTINEL_FILE, body: JSON.generate(s))
  factor
rescue StandardError
  nil
end

.usable_preference?(opts = {}) ⇒ Boolean

P15 — keep only usable preference pairs for balance/export/promote. Drops CORRECTION-only chosen, resolve rows without trajectory shape, and chosen≪rejected unless shape is a known trajectory form.

Returns:

  • (Boolean) 


698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
 
# File 'lib/pwn/ai/agent/reward.rb', line 698

public_class_method def self.usable_preference?(opts = {})
  r = opts.is_a?(Hash) && opts.key?(:row) ? opts[:row] : opts
  r = r.transform_keys(&:to_sym) if r.respond_to?(:transform_keys)
  chosen = r[:chosen].to_s
  rejected = r[:rejected].to_s
  shape = r[:shape].to_s
  source = r[:source].to_s
  return false if chosen.strip.empty? || rejected.strip.empty?
  return false if chosen.strip == rejected.strip
  return false if chosen.match?(/\A\s*CORRECTION:\s*/i) && chosen.length < 400
  return false if shape == 'fix_prose'
  # P25 — resolve rows must be trajectory-shaped to count as usable
  return false if source == 'mistakes_resolve' && !TRAJECTORY_SHAPES.include?(shape)

  # chosen ≪ rejected without trajectory shape → commentary, not policy
  unless TRAJECTORY_SHAPES.include?(shape)
    return false if rejected.length >= 200 && chosen.length < (rejected.length * 0.25) && chosen.length < 200
    return false if rejected.length >= 400 && chosen.length < 120
  end
  true
rescue StandardError
  false
end

.verify_as_reward(opts = {}) ⇒ Object

Supported Method Parameters

g = PWN::AI::Agent::Reward.verify_as_reward(final: text)



463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
 
# File 'lib/pwn/ai/agent/reward.rb', line 463

public_class_method def self.verify_as_reward(opts = {})
  return nil unless defined?(Extrospection) && Extrospection.respond_to?(:verify)

  final = opts[:final].to_s
  claim = final[Learning::CLAIM_RX] if defined?(Learning)
  return nil if claim.to_s.empty?

  # P26 — drop metric crumbs ("cap 0.2") that match loose patterns
  if defined?(Learning) && Learning.respond_to?(:checkable_claim?, true)
    return nil unless Learning.send(:checkable_claim?, claim: claim)
  elsif claim.match?(/\A(?:cap|share|proxy|judge|success|only|now|gap|score|rate)\b/i) ||
        (claim.match?(/\d+\.\d+/) && !claim.match?(/\d+\.\d+\.\d+|CVE-/i))
    return nil
  end

  # 1.5 — sampled E3: always when flag true; never when false;
  # nil/auto → always on frontier, ~10% on local when CLAIM_RX hits.
  flag = agent_flag(key: :verify_as_reward, default: nil)
  eng  = (PWN::Env.dig(:ai, :active) if defined?(PWN::Env)).to_s.downcase
  local = eng == 'ollama'
  run = case flag
        when true then true
        when false then false
        else
          local ? (Digest::SHA256.hexdigest(claim.to_s)[0, 2].to_i(16) % 10).zero? : true
        end
  return nil unless run

  r = Extrospection.verify(claim: claim, commit: true)
  { claim: claim, verdict: r[:verdict], confidence: r[:confidence], reward: VERDICTS[r[:verdict]] || 0.5 }
rescue StandardError
  nil
end

.warm_sentinel(opts = {}) ⇒ Object

P10 — backfill the R3 ring from Learning outcomes so offline/local hosts reach SENTINEL_WINDOW without waiting for live remote introspect. Only fills empty slots; never flushes a warm window. Called by Curriculum.offline_judge and safe to cron.



342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
 
# File 'lib/pwn/ai/agent/reward.rb', line 342

public_class_method def self.warm_sentinel(opts = {})
  s = normalize_sentinel(raw: load_sentinel)
  have = Array(s[:window]).length
  return { added: 0, samples: have, status: :full, proxy_distrust: proxy_distrust } if have >= SENTINEL_WINDOW
  return { added: 0, samples: have, status: :no_learning, proxy_distrust: proxy_distrust } unless defined?(Learning)

  need = SENTINEL_WINDOW - have
  limit = (opts[:limit] || [need * 4, 200].max).to_i
  # Prefer scored rows; fall back to success-boolean so local hosts still warm.
  rows = Learning.outcomes(limit: limit)
  scored, unscored = rows.partition { |r| !r[:score].nil? }
  ordered = scored.reverse + unscored.reverse
  added = 0
  ordered.each do |r|
    break if added >= need

    judge = if r[:score]
              r[:score].to_f.clamp(0.0, 1.0)
            else
              case r[:success]
              when true, 'true' then 0.75
              when 'soft' then 0.55
              when false, 'false' then 0.25
              else 0.5
              end
            end
    proxy = case r[:success]
            when true, 'true' then true
            when false, 'false', 'soft' then false
            else judge >= 0.6
            end
    record_sentinel(proxy: proxy, judge: judge)
    added += 1
  end
  final_n = Array(load_sentinel[:window]).length
  # Recompute distrust once window is full so controllers can engage.
  snap = final_n >= SENTINEL_WINDOW ? sentinel : { samples: final_n, status: :insufficient }
  {
    added: added,
    samples: final_n,
    status: (final_n >= SENTINEL_WINDOW ? :warmed_full : :warmed_partial),
    proxy_distrust: proxy_distrust,
    sentinel: snap.is_a?(Hash) ? snap.slice(:samples, :status, :reward_hacked, :proxy_distrust, :proxy, :judge) : nil
  }
rescue StandardError => e
  { added: 0, error: "#{e.class}: #{e.message}" }
end

.write_source_quota(opts = {}) ⇒ Object

Share of source among the newest WRITE_SOURCE_WINDOW prefs.



584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
 
# File 'lib/pwn/ai/agent/reward.rb', line 584

public_class_method def self.write_source_quota(opts = {})
  source = opts[:source].to_s
  recent = preferences(limit: WRITE_SOURCE_WINDOW)
  return { over_cap: false, share: 0.0, n: 0, window: recent.length, underfilled: true } if recent.length < 10

  n = recent.count { |r| r[:source].to_s == source }
  share = n.to_f / recent.length
  target = TARGET_SOURCE_MIX[source]
  {
    over_cap: share > WRITE_SOURCE_CAP,
    share: share.round(3),
    n: n,
    window: recent.length,
    source: source,
    cap: WRITE_SOURCE_CAP,
    target: target,
    underfilled: target ? share < (target * 0.5) : share < 0.05,
    deficit: target ? (target - share).round(3) : nil
  }
rescue StandardError
  { over_cap: false, share: 0.0, underfilled: true }
end