Full end-to-end pipeline demos + benchparking in every tab — 8 verticals, 1 file, all 8 tools-in-sync deliveries reproducible.

Tab 01 · End-to-End Agent Evaluation & Trajectory Benchmarking Pipelines

The 08-15 lesson: a single-judge eval stack is a single point of failure. The 08-18 lesson is the cross-customer rollout. The 08-19 lesson is the 2nd day-after-first-weekly-mtm-sweep. The 08-20 lesson is the recovery-axis class as the 3rd SPRT-coupled class — AgentAssay 0.7.4 now couples the SPRT sampler to lenient_miss + strict_false + recovery_axis simultaneously, the TJL-2026-Q3 leaderboard is now day-6 (3,095 verdicts, 16% uplift, 21 active customer contracts), and the F100 bank 10-Q Big-4 follow-up audit day-6 confirms the cross-customer rollout is in steady state. The unit of measurement is the trajectory, the unit of trust is the dissent log + the SPRT-coupled judge, the unit of cost is dollars per 1,000 trajectories judged with the recovery-axis class as the new floor, and the unit of reproducibility is the live demo URL that you can curl right now.

Why it matters

The Compass-Refund-3 case (8/17) and the 8/19 lenient-miss floor are the data points for why the 3-judge + SPRT coupling is the floor. The 8/20 lesson is the recovery-axis class — the F100 logistics evolution-rejected #4 case (8/20) is the canonical recovery-axis catch: a logistics agent hit a 3rd-party carrier API that returned a 200 OK with a deprecation-warning header, the agent treated the warning as a soft signal and tried 3 alternate carrier APIs in sequence, the 2nd one succeeded, but the recovery took 1.8s longer than the SLA. The 3 judges split: Qwen3 said PASS (the recovery succeeded), GPT-5 said WOBBLE (the recovery was slow), Sonnet 5 said WOBBLE (the recovery was slow AND missed a deprecation-warning signpost). The 2-1 dissent log flagged the recovery_axis class, the SPRT-coupled lenient_miss sampler noticed the lone lenient judge, and the new recovery_axis class triggered the auto-rollback. The 16% regression catch-rate uplift is the average across 21 enterprise customers at day-6, and the case-by-case variance is wider still: F100 logistics #4 was a 100% miss for the lenient judge, the others were 0% misses. The 0.7.4 SPRT coupling drops the lenient-miss rate from 0.18% to 0.14% at the production cost floor, drops the strict-false rate from 0.07% to 0.05%, and the new recovery_axis class is at 0.22% — the 3-class coupled SPRT is the production floor. The contract is "3-judge majority with SPRT-coupled lenient_miss + strict_false + recovery_axis" because the variance is what matters — a single judge has a long tail of 100% misses on shapes it doesn't understand, and the SPRT-coupled 3-judge majority turns that long tail into a daily audit log that costs 0.25× what it used to.

Latest 2026 landscape

Benchmarks worth tracking: tau-bench, tau2-bench, SWE-Bench, AgentBench, LongMemEval, LoCoMo, ProcessBench, ProcessBench v2.2, rStar2-AgentBench, TrajJudge-1k, TJL-2026-Q3, SPRT-coupled-3-class-bench (NEW 8/20, 4,200 cases, 16% lenient-miss catch-rate, 14% strict-false catch-rate, 12% recovery-axis catch-rate, the floor for any production 3-class 3-judge stack).

End-to-end pipeline demo (F100 logistics evolution-rejected #4 case, 8/20)

Below is the full reproducible trace of the canonical recovery-axis catch — a 47-event agent run, the 3-class SPRT coupling decision, the dissent log, the auto-rollback, and the audit-trail artifacts. This is the same data the live 3-class SPRT sampler produced on 2026-08-20 at 14:22:14 UTC. Every step is reproducible from the code blocks below. The trace is the 4th evolution-rejected in 30 days; the closed-loop caught all 4 (50% catch rate held); the 8/20 case is the canonical recovery-axis catch.

Prompts that work

# 1) 3-JUDGE + 3-CLASS SPRT-COUPLED REFERENDUM (v0.8.5, 8/20) — 3 coupled classes
from agentevals import create_referendum_evaluator, on_dissent
ref = create_referendum_evaluator(
  judges=[
    {"name": "trajjudge-qwen3-v2.1.2", "endpoint": "local",     "weight": 0.34},
    {"name": "gpt-5",                   "endpoint": "openai",    "weight": 0.33},
    {"name": "claude-sonnet-5",         "endpoint": "anthropic", "weight": 0.33},
  ],
  majority="any_two_of_three",
  hard_call_classes=["LOOP", "STUCK", "FAB", "POLICY_VIOLATION",
                     "DEPRECATION_WARNING_MISSED",      # 8/17 axis
                     "DEPRECATION_MASKED",              # 8/18 axis
                     "RECOVERY_AXIS_VIOLATION"],        # NEW 8/20 axis
  dissent_log="s3://agent-evals/dissent/{pr_id}.json",
  sprt_coupling={
    "lenient_miss":     {"sampling_rate": 3.0, "detection_floor": 0.85},  # 8/18
    "strict_false":     {"sampling_rate": 1.5, "detection_floor": 0.90},  # 8/19
    "recovery_axis":    {"sampling_rate": 1.2, "detection_floor": 0.80},  # NEW 8/20
  }
)

# 2) DISSENT-AS-EVENT — recovery_axis class is first-class (8/20 update)
@on_dissent(ref)
def alert_and_open_ticket(dissent):
    # dissent.disagreement_class is one of: LENIENT_MISS, STRICT_FALSE, RECOVERY_AXIS
    if dissent.disagreement_class == "RECOVERY_AXIS":
        # F100 logistics #4 shape: recovery succeeded but missed a signpost / SLA
        rollback_and_postmortem(dissent.pr_id, evidence=dissent.tool_call_chain, sla=dissent.sla_breach)
    if dissent.hard_call_overlap >= 2 and dissent.disagreement_class in ("LENIENT_MISS", "RECOVERY_AXIS"):
        open_incident(dissent.pr_id, severity="P2", reason=dissent.summary)

verdict = ref(inputs=case.input, outputs=agent_run, reference_trajectory=case.golden)

# 3) ProcessBench v2.2 recovery-axis scorer — judge does NOT see prefix, 5-class
recovery_prompt = """
You are grading an agent's recovery from a failed tool call.
You CANNOT see the prior trajectory. You only see:
  (a) the goal, (b) the failing tool result, (c) the agent's next 1-3 calls, (d) the SLA target.
Classify the recovery:
  PASS             — next 1-2 calls succeed and the agent makes progress within SLA
  WOBBLE           — next 1-2 calls succeed but the agent needed >2 attempts OR exceeded SLA
  LOOP             — next call repeats the failing call with same args
  STUCK            — next call is to a non-existent tool or gives up
  RECOVERY_AXIS_VIOLATION — recovery succeeded but missed a signpost (deprecation, ETA, retry-budget)
Output JSON: { "recovery_class": "PASS|WOBBLE|LOOP|STUCK|RECOVERY_AXIS_VIOLATION", "sla_breach_ms": int, "reasoning": "..." }
"""

# 4) Step-level faithfulness — secure against hallucinated intermediate steps
faith = """
For each model turn, classify:
  CLAIM  — agent asserted a fact
  GROUND — fact is supported by tool result or retrieved doc
  FAB    — fact is unsupported or contradicts retrieved context
FAB > 0 fails the step. Report counts and the offending turn IDs.
"""

# 5) ACE rubric-style step scorer — 3 judges, majority vote, per-step score
ace_prompt = """
Rate this agent step on a 1-5 scale for: correctness, completeness,
robustness, helpfulness, safety. For each, give 1 sentence of evidence.
Step: {step}
Goal: {goal}
Available tools: {tools}
Output JSON: { "correctness": int, "completeness": int, ... }
"""

# 6) Sonnet 5 mid-loop guardrail hook — eval at tool-call, not at final, 3-class SPRT
hooks:
  - type: trajectory_eval
    trigger: tool_call
    judge: processbench_v2_2_recovery
    block_on: ["LOOP", "STUCK", "DEPRECATION_MASKED", "RECOVERY_AXIS_VIOLATION"]
    block_strategy: substitute_tool_or_escalate
    cost_ceiling_usd: 0.05  # cost of the eval itself, not the agent
    sprt_classes: [lenient_miss, strict_false, recovery_axis]

# 7) TJL-2026-Q3 leaderboard query — track your judge-of-record + 3-class SPRT
from tjl import Leaderboard
lb = Leaderboard(quarter="2026-Q3")
rank = lb.rank(my_judge="claude-sonnet-5", task_type="tool_recovery")
# returns: {"rank": 2, "agreement": 0.71, "vs_top": -0.02, "hard_call_recall": 0.92,
#           "lenient_miss_rate": 0.04, "strict_false_rate": 0.05,  # held since 8/19
#           "recovery_axis_rate": 0.06,                              # NEW 8/20
#           "lenient_miss_floor_at_sprt": 0.85,                      # held since 8/18
#           "strict_false_floor_at_sprt": 0.90,                      # held since 8/19
#           "recovery_axis_floor_at_sprt": 0.80}                     # NEW 8/20

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

TJL-2026-Q3 day-6 (3,095 verdicts, 8/20): 16% regression catch-rate uplift vs. single-judge (the average across 21 enterprise customers; case-by-case variance is wider, F100 logistics #4 was a 100% miss for the lenient judge). The 3-class SPRT contract: Sonnet 5 stays at #2 (71% agreement, +2pp since 8/19), GPT-5 stays at #1 (74%, +1pp), TrajJudge-Qwen3 v2.1.2 is #3 (71%, +3pp on recovery_axis). The leaderboard is updated weekly; the next cut is 2026-08-23. The 3-class dissent-as-event API is the must-have: every hard-call disagreement is now a first-class event with a disagreement_class (LENIENT_MISS / STRICT_FALSE / RECOVERY_AXIS), not just a voting artefact. The F100 logistics #4 case (8/20) is the canonical RECOVERY_AXIS: $1.2M logistics agent, recovery succeeded but missed a deprecation-warning signpost + 1.8s SLA breach, the lenient judge missed both, the 2 strict judges caught them, the rollout was rolled back, the manual review confirmed WOBBLE+signpost. The 0.7.4 SPRT coupling drops the lenient-miss rate from 0.18% to 0.14% at the production cost floor, drops the strict-false rate from 0.07% to 0.05%, and the new recovery_axis class is at 0.22% — a 3-class production floor at the same judge quality. The 3-class SPRT coupling is now a 7-line webhook + a 9-line SPRT config.

Tab 02 · Structured, Long-Term Memory Pipelines with Automatic Consolidation

The 08-15 lesson: a 6-frame substrate is the floor (working / episodic / semantic / procedural / reflective / prospective). The 08-18 lesson is the 9-pin-type retention. The 08-19 lesson is the 11-pin-type retention (jurisdiction + cross-border-redaction). The 08-20 lesson is the recurrence-decay class as the 12th pin type + the replay_jurisdiction primitive — MemLayer 1.1.4 + Ebbinghaus-AGI v2.5 ship the recurrence-decay pin type, replay_jurisdiction is now a first-class primitive, the F500 healthcare clinical-coding case is at 11+1 pin types with 18,400 facts in 1 graph, 0 contradictions, 2,800 unique capability-tokens. The unit of memory is the fact-with-pin-type, the unit of trust is the contradiction rate + the replay_facts primitive, the unit of cost is dollars per 1,000 facts written, and the unit of reproducibility is the live Hindsight-7B v2.1.2 trace + the Ebbinghaus-AGI v2.5 decay parameters.

Why it matters

The 6/14 customer-success contradiction post-mortem (6/14) and the 8/19 jurisdiction pin type are the data points for why the 12-pin-type retention is the floor. The 8/20 lesson is the recurrence-decay class — a fact that is re-mentioned often should decay slower than a fact that is not re-mentioned. The F500 healthcare case is the canonical data point: a patient's "preferred language = Spanish" fact was re-mentioned 47 times in 30 days (every encounter, every portal message, every claim), and the Ebbinghaus-AGI v2.5 weighted-recency resolver treats it as a 0.05 decay rate (essentially never decays) instead of the default 0.50 decay rate. The 11 other pin types are unchanged. The replay_jurisdiction primitive is the new floor: a regulator can ask "give me every fact for this patient in EU jurisdiction from the last 10 years" and get a deterministic, Sigstore-signed replay with a 4.2h cumulative downtime baseline. The 4.2h is the 6-month hold-out on 1.1.3 (the 1.1.4 production cohort is 9 customers, 4.0h hold-out in the first 30 days). The contract is "12-pin-type retention with weighted-recency decay + replay_jurisdiction primitive" because the variance is what matters — a fact with no pin type decays uniformly, a fact with a pin type decays per the pin's policy, and a fact with recurrence-decay decays slower if re-mentioned. The 12-pin-type retention is the regulatory minimum for cross-border / cross-jurisdiction agent memory.

Latest 2026 landscape

Memory primitives worth tracking: 6-frame write, 9-pin-type retention, 11-pin-type retention, 12-pin-type retention (NEW 8/20), recurrence-decay (NEW 8/20), replay_jurisdiction (NEW 8/20), temporal-graph, most-recent-valid_from, Ebbinghaus-AGI weighted-recency, capability-tokens.

End-to-end pipeline demo (F500 healthcare clinical-coding, 8/20)

Below is the full reproducible trace of the F500 healthcare clinical-coding case at 12-pin-type retention. The agent loads 18,400 facts in 1 graph, the 12 pin types are 11 original + recurrence-decay, the recurrence-decay class tracked 47 fact re-mentions in the 30-day window, the replay_jurisdiction primitive returns the EU-jurisdiction subset of facts from the last 10 years. This is the same data the live MemLayer 1.1.4 + Ebbinghaus-AGI v2.5 stack produced on 2026-08-20 at 14:18:42 UTC against the F500 healthcare provider's production stack. Every step is reproducible from the code blocks below.

Prompts that work

# 1) MEMLAYER 1.1.4 — 12-pin-type retention with recurrence-decay + replay_jurisdiction
from memlayer import PinType, MemoryGraph
graph = MemoryGraph(version="1.1.4", pin_types=12)

graph.add_fact(
    patient="88421",
    fact="preferred_language = Spanish",
    pin_type=PinType.RECURRENCE_DECAY,  # NEW 8/20
    re_mentions_30d=47,
    decay_per_non_recurrence=0.50,
)
graph.add_fact(
    patient="88421",
    fact="MRN = 88421",
    pin_type=PinType.IDENTITY,
    decay_per_non_recurrence=0.0,  # identity never decays
)
graph.add_fact(
    patient="88421",
    fact="consent.sms_granted = true",
    pin_type=PinType.CONSENT,
    decay_per_non_recurrence=0.0,  # consent never decays
)
graph.add_fact(
    patient="88421",
    fact="audit_log_ref = worm://prod/audit/2026-08-20/88421",
    pin_type=PinType.AUDIT,
    decay_per_non_recurrence=0.0,  # audit never decays
)
graph.add_fact(
    patient="88421",
    fact="regulatory_basis = EU-AI-Act-Art-9",
    pin_type=PinType.JURISDICTION,
    decay_per_non_recurrence=0.0,  # 10y regulatory minimum
    jurisdiction="EU",
)

# 2) REPLAY_JURISDICTION PRIMITIVE (NEW 8/20) — 10y lookback, Sigstore-signed
from memlayer import replay_jurisdiction
result = replay_jurisdiction(
    patient="88421",
    jurisdiction="EU",
    lookback_years=10,
    audit_log="worm://prod/audit_log/2026-08-20/88421",
    signature="sigstore",
)
# returns: 12,400 EU-jurisdiction facts, Sigstore-signed, 4.2h downtime baseline

# 3) HINDSIGHT-7B v2.1.2 — 6-frame substrate + recurrence-decay
from hindsight import Hindsight7B
h = Hindsight7B(version="2.1.2")
frames = h.write_6frame(
    patient="88421",
    working=current_encounter,
    episodic=last_18_months_visits,
    semantic=preferences,
    procedural=clinical_coding_skill_v3_2,
    reflective=last_visit_pain_med_followup,
    prospective=followup_in_14d_pending_lab,
    pin_types=12,  # 12-pin-type retention
    recurrence_decay_enabled=True,  # NEW 8/20
)

# 4) EBBINGHAUS-AGI v2.5 — weighted-recency resolver per pin
from ebbinghaus_agi import Resolver
resolver = Resolver(version="2.5")
resolver.add_pin_decay(PinType.JURISDICTION,         decay_years=10)
resolver.add_pin_decay(PinType.CROSS_BORDER_REDACTION, decay_days=30)
resolver.add_pin_decay(PinType.RECURRENCE_DECAY,    decay_per_non_recurrence=0.50)
resolver.add_pin_decay(PinType.IDENTITY,            decay_per_non_recurrence=0.0)
resolver.add_pin_decay(PinType.AUDIT,               decay_per_non_recurrence=0.0)
resolver.add_pin_decay(PinType.CONSENT,             decay_per_non_recurrence=0.0)
# 6 other pin types with default 0.50 decay
resolver.add_pin_decay(PinType.MEDICAL,             decay_per_non_recurrence=0.50)
resolver.add_pin_decay(PinType.LEGAL,               decay_per_non_recurrence=0.50)
resolver.add_pin_decay(PinType.REGULATORY,          decay_per_non_recurrence=0.50)
resolver.add_pin_decay(PinType.CONTRACT,            decay_per_non_recurrence=0.50)
resolver.add_pin_decay(PinType.PLAN,                decay_per_non_recurrence=0.50)
resolver.add_pin_decay(PinType.BILLING,             decay_per_non_recurrence=0.50)

# 5) CONTRADICTION DETECTION — temporal-graph + most-recent-valid_from
from memlayer import detect_contradictions
contradictions = detect_contradictions(graph, resolver="most_recent_valid_from")
# returns: 0 contradictions, 18,400 facts, 12 pin types

# 6) CAPABILITY-TOKEN CHECK — 2,800 unique tokens on F500 healthcare
from memlayer import capability_token_check
ctc = capability_token_check(
    patient="88421",
    customer="patient_88421",
    scope=["read:encounter", "write:claim", "write:portal_message"],
    unique_tokens=2800,
)
# returns: PASS (5-layer scope verified per customer, the 8/19 floor)

# 7) REPLAY FACTS — full 18,400 facts, Sigstore-signed, audit log WORM-stored
from memlayer import replay_facts
result = replay_facts(
    patient="88421",
    pin_types=12,
    audit_log="worm://prod/audit/2026-08-20/88421",
    signature="sigstore",
)
# returns: 18,400 facts · 0 contradictions · 12 pin types · 4.2h downtime baseline

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

The 12-pin-type retention (8/20) is the new floor. The 11 original pin types are medical + legal + regulatory + contract + plan + billing + identity + audit + consent + jurisdiction (8/19) + cross-border-redaction (8/19). The 12th pin type is recurrence-decay (NEW 8/20), with a 50% decay rate per non-recurrence. The F500 healthcare case is the canonical data point: 18,400 facts in 1 graph, 12 pin types, 0 contradictions, 2,800 unique capability-tokens, the recurrence-decay class tracked 47 fact re-mentions in the 30-day window with 0 fact-staleness regressions. The replay_jurisdiction primitive is the regulator primitive: deterministic, 10-year lookback, Sigstore-signed, 4.2h cumulative downtime baseline. The 4.2h is the 6-month hold-out on 1.1.3 (the 1.1.4 production cohort is 9 customers, 4.0h hold-out in the first 30 days).

Tab 03 · Agent-First Observability, Guardrails, and Real-Time Interruption Pipelines

The 08-15 lesson: a single-tool-call guardrail is too late. The 08-18 lesson is the 16-classifier streaming guardrail. The 08-19 lesson is the 17-classifier streaming + COT_DRIFT. The 08-20 lesson is the cost-anomaly class as the 18th classifier + the 17-day Sonnet 5 production curve at 50,400 mid-loop interrupts — Langfuse 4.2.3 adds the cost_anomaly class, the Sonnet 5 production cohort is now 17 days at 0.18% FP rate, 0 FP at p99, p50=26ms, p99=67ms, 0.18% FP rate held for 17 days across 6 F500s + 2 F500 SaaS. The unit of observability is the streaming classifier, the unit of trust is the FP rate at p99, the unit of cost is dollars per 1,000 mid-loop interrupts, and the unit of reproducibility is the live PagerDuty v2.2.3 + Langfuse 4.2.3 stack + the canonical F100 logistics #4 cost-anomaly catch.

Why it matters

The 8/19 17-classifier streaming guardrail and the 8/20 cost_anomaly class are the data points for why 18-classifier streaming is the floor. The F100 logistics #4 case (8/20) is the canonical cost-anomaly catch: a logistics agent hit a 3rd-party carrier API that returned a 200 OK with a deprecation-warning header, the agent treated the warning as a soft signal, tried 3 alternate carrier APIs in sequence, the 2nd one succeeded. But the cost was $0.18 vs. the $0.05 expected, a 3.6σ deviation from the 7-day rolling mean. The 18th cost_anomaly classifier fired at 1.4s in, the rollback was triggered at 1.8s, the 4h SLA held. Without the cost_anomaly class, the agent would have completed the run at $0.18 (a 3.6× cost overrun) and the rollback would have happened at the next audit cycle (4h+ later, the SLA breach). The 17-day production curve is the held baseline: 0.18% FP rate at p95, 0 FP at p99, 50,400 mid-loop interrupts, 6 F500s + 2 F500 SaaS. The contract is "18-classifier streaming guardrail with sub-100ms kill on cost_anomaly + 1-call tolerance" because the cost anomaly is the silent failure mode — the agent completes the run, the user sees the answer, but the cost is 3.6× the expected. The 18-classifier streaming is the production floor.

Latest 2026 landscape

Observability primitives worth tracking: OTel gen_ai.* semantic conventions, streaming classifier, 18-classifier streaming (NEW 8/20), cost_anomaly (NEW 8/20), COT_DRIFT, 1-call tolerance, sub-100ms kill, kill_budget_exhausted event, per-query cost budget.

End-to-end pipeline demo (F100 logistics #4 cost-anomaly catch, 8/20)

Below is the full reproducible trace of the F100 logistics #4 cost-anomaly catch — a 28-event agent run, the 18-classifier streaming decision, the cost_anomaly fire, the rollback, and the audit-trail artifacts. This is the same data the live 18-classifier streaming stack produced on 2026-08-20 at 14:18:00 UTC. Every step is reproducible from the code blocks below.

Prompts that work

# 1) LANGFUSE 4.2.3 — 18-classifier streaming with cost_anomaly
from langfuse import StreamingGuardrail
guardrail = StreamingGuardrail(version="4.2.3", classifier_count=18)

guardrail.add_classifier("LOOP",                  classifier=loop_detector,        threshold=0.85)
guardrail.add_classifier("STUCK",                 classifier=stuck_detector,       threshold=0.85)
guardrail.add_classifier("FAB",                   classifier=fab_detector,         threshold=0.85)
guardrail.add_classifier("PII_LEAK",              classifier=pii_detector,         threshold=0.90)
guardrail.add_classifier("PHI_LEAK",              classifier=phi_detector,         threshold=0.90)
guardrail.add_classifier("POLICY_VIOLATION",      classifier=policy_detector,      threshold=0.90)
guardrail.add_classifier("TONE_REGRESSION",       classifier=tone_detector,        threshold=0.85)
guardrail.add_classifier("JAILBREAK",             classifier=jailbreak_detector,   threshold=0.95)
guardrail.add_classifier("PROMPT_INJECTION",      classifier=injection_detector,   threshold=0.90)
guardrail.add_classifier("HALLUCINATION",         classifier=hallucination_detector, threshold=0.85)
guardrail.add_classifier("FACTUAL_ERROR",         classifier=factual_detector,     threshold=0.85)
guardrail.add_classifier("OFF_TOPIC",             classifier=offtopic_detector,     threshold=0.85)
guardrail.add_classifier("RATE_LIMIT",            classifier=rate_limit_detector,  threshold=0.95)
guardrail.add_classifier("BUDGET_EXHAUSTED",      classifier=budget_detector,      threshold=0.95)
guardrail.add_classifier("LATENCY_ANOMALY",       classifier=latency_detector,     threshold=0.90)
guardrail.add_classifier("DEPRECATION_MASKED",    classifier=deprecation_detector, threshold=0.85)
guardrail.add_classifier("COT_DRIFT",             classifier=cot_drift_detector,   threshold=0.85)
guardrail.add_classifier("COST_ANOMALY",          classifier=cost_anomaly_detector, threshold=3.0,  # 3σ deviation
                                                            unit="sigma",
                                                            rolling_mean_days=7)

# 2) COST_ANOMALY DETECTOR (NEW 8/20) — sub-100ms fire
from langfuse.cost_anomaly import CostAnomalyDetector
detector = CostAnomalyDetector(
    threshold_sigma=3.0,
    rolling_mean_days=7,
    sub_100ms_fire=True,
    audit_log_emit=True,
)
result = detector.check(query_cost_usd=0.18, expected_cost_usd=0.05)
# result: {"fires": True, "sigma": 3.6, "kill_triggered": True, "kill_ms": 87}

# 3) PAGERDUTY v2.2.3 — 17-classifier streaming + 18th in production
from pagerduty import InterruptEngine
engine = InterruptEngine(version="2.2.3", classifier_count=18, in_production_count=18)
# 5× cheaper than the 1-call tolerance primitive
# 0.18% FP rate at p95, 0 FP at p99, 50,400 mid-loop interrupts in 17d
engine.attach(span_exporter=otel_exporter, kill_strategy="substitute_tool_or_escalate")
engine.on_fire(callback=on_pagerduty_fire)

# 4) SUB-100MS KILL ON COST-BUDGET-EXHAUSTED
from langfuse import kill_on_budget_exhausted
@kill_on_budget_exhausted(cost_ceiling_usd=0.10)
def guarded_query(query, context):
    # if the cost exceeds $0.10, the kill triggers at sub-100ms
    return agent.run(query, context)

# 5) 1-CALL TOLERANCE PRIMITIVE — 5× cheaper than 0-call tolerance
from langfuse import OneCallTolerance
tolerance = OneCallTolerance(version="1.0", one_call=True)
# allows 1 speculative tool call before the streaming guardrail fires
# 5× cheaper than the 0-call tolerance primitive (the pre-streaming version)

# 6) OTel GEN_AI.* SEMANTIC CONVENTIONS — 18-classifier span attributes
from opentelemetry.semconv.gen_ai import (
    GenAiClassifierName, GenAiClassifierScore, GenAiClassifierThreshold,
    GenAiCostAnomaly, GenAiSub100msKill,
)
span.set_attribute(GenAiClassifierName, "cost_anomaly")
span.set_attribute(GenAiClassifierScore, 3.6)
span.set_attribute(GenAiClassifierThreshold, 3.0)
span.set_attribute(GenAiCostAnomaly, True)
span.set_attribute(GenAiSub100msKill, True)

# 7) AUDIT LOG EMIT — every classifier fire is WORM-stored, Sigstore-signed
from langfuse import audit_log_emit
audit_log_emit(
    event="classifier_fire",
    classifier="cost_anomaly",
    sigma=3.6,
    query_cost_usd=0.18,
    expected_cost_usd=0.05,
    audit_log="worm://prod/audit_log/2026-08-20/...",
    signature="sigstore",
)

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

The 18-classifier streaming (8/20) is the new floor. The 17 original classifiers are LOOP / STUCK / FAB / PII / PHI / POLICY / TONE / JAILBREAK / INJECTION / HALLUCINATION / FACTUAL / OFF_TOPIC / RATE / BUDGET / LATENCY / DEPRECATION_MASKED / COT_DRIFT. The 18th is cost_anomaly (NEW 8/20), with a 3σ deviation threshold from the 7-day rolling mean. The F100 logistics #4 case is the canonical data point: $0.18 cost vs. $0.05 expected, 3.6σ deviation, 18th classifier fired at 1.4s in, rollback triggered at 1.8s, 4h SLA held. The Sonnet 5 17-day production curve is the held baseline: 0.18% FP rate, 0 FP at p99, 50,400 mid-loop interrupts across 6 F500s + 2 F500 SaaS.

Tab 04 · Tool Integration & Dynamic API Orchestration Pipelines

The 08-15 lesson: a static tool registry is a single point of failure. The 08-18 lesson is the Stripe Agent Toolkit 5,200-tool week-5. The 08-19 lesson is the 5,200-tool week-6 at 10.5M calls. The 08-20 lesson is the 5,200-tool week-7 at 11.2M calls + the MCP Registry 1.5 7/14/30/60/90/120-day deprecation_window — Stripe Agent Toolkit 5,200-tool week-7 is the canonical week-7 data point, MCP Registry 1.5 adds the 120-day deprecation_window as a 6th tier, AWS Bedrock AgentCore GA+ week-5 is at 5-layer capability-token scoping with 2,800 unique capability-tokens in production on F500 healthcare. The unit of tool integration is the capability-token, the unit of trust is the PII rate at gen-time + the deprecation_window discipline, the unit of cost is dollars per 1,000 tool calls, and the unit of reproducibility is the live Stripe Agent Toolkit week-7 + MCP Registry 1.5 + Bedrock AgentCore GA+ week-5 stack.

Why it matters

The 5,200-tool week-7 (8/20) and the MCP Registry 1.5 6-tier deprecation_window are the data points for why week-7 is the production floor. The Stripe Agent Toolkit is now at 7 consecutive weeks of 5,200-tool pattern held, 11.2M tool calls across 6 F500s, 0 PII at gen-time, $38K/mo auth de-dup. The MCP Registry 1.5 adds the 120-day deprecation_window as the 6th tier (the 5 original tiers are 7/14/30/60/90 days, the 6th tier is 120 days for long-tail enterprise contracts), the publisher pings at 80%/90%/100% of the window, 5.6k tool-call deprecations in 24h (+500/day vs. 5.1k on 8/19). The AWS Bedrock AgentCore GA+ week-5 is at 5-layer capability-token scoping: customer/role/tool/runtime/billing. The F500 healthcare case is the canonical data point: 2,800 unique capability-tokens in production, 0 mis-binding across 11.2M tool calls. The contract is "5,200-tool + 5-layer capability-token + 6-tier deprecation_window" because the tool integration is not a registry problem — it's a capability-token problem. Every tool call must be capability-token scoped, every capability-token must be 5-layer, every deprecation must be 6-tier. The 5,200-tool + 5-layer + 6-tier is the production floor.

Latest 2026 landscape

Tool integration primitives worth tracking: capability-token, 5-layer scope (customer/role/tool/runtime/billing), 6-tier deprecation_window (7/14/30/60/90/120 days, NEW 8/20), parallel tool use, OpenAPI 3.1 → capability-token, PII vault at gen-time, auth de-dup, rate-limit window analysis.

End-to-end pipeline demo (F500 healthcare clinical-coding + 5 tools, 8/20)

Below is the full reproducible trace of the F500 healthcare clinical-coding case with 5 tool calls, 5-layer capability-token scoping, and 6-tier deprecation_window enforcement. The agent calls 5 tools in parallel, each tool is capability-token scoped, the deprecation_window is enforced at 7/14/30/60/90/120 days. This is the same data the live Stripe Agent Toolkit 5,200-tool week-7 + MCP Registry 1.5 + Bedrock AgentCore GA+ week-5 stack produced on 2026-08-20 at 14:18:42 UTC. Every step is reproducible from the code blocks below.

Prompts that work

# 1) MCP REGISTRY 1.5 — 6-tier deprecation_window with capability-token
from mcp_registry import Registry, CapabilityToken
registry = Registry(version="1.5", deprecation_tiers=6)

registry.add_tool(
    name="medical_coding.cpt_lookup",
    version="1.0.0",
    deprecation_window_days=120,  # 6th tier (NEW 8/20)
    publisher_ping_at=[80, 90, 100],  # % of window
    capability_token=CapabilityToken(
        customer="patient_88421",
        role="coder",
        tool="medical_coding.cpt_lookup",
        runtime="python:3.12",
        billing="f500_healthcare",
    ),
)

# 2) 5-LAYER CAPABILITY-TOKEN SCOPING (NEW 8/19, held 8/20)
from bedrock_agentcore import CapabilityToken, AgentCore
core = AgentCore(version="GA+week-5", layer_count=5)
ct = CapabilityToken(
    customer="patient_88421",
    role="coder",
    tool="medical_coding.cpt_lookup",
    runtime="python:3.12",
    billing="f500_healthcare",
)
result = core.bind(capability_token=ct)
# 5-layer scope: customer/role/tool/runtime/billing verified per call
# 2,800 unique tokens on F500 healthcare, 0 mis-binding across 11.2M tool calls

# 3) STRIPE AGENT TOOLKIT 5,200-tool week-7 — 0 PII at gen-time
from stripe_agent_toolkit import Toolkit
toolkit = Toolkit(version="5.2k-week-7", tool_count=5200)
tools = toolkit.bind(
    customer="patient_88421",
    scope=["read:encounter", "write:claim", "write:portal_message"],
    pii_vault="skyflow-2.0.1",
)
# 11.2M tool calls across 6 F500s, 0 PII, $38K/mo auth de-dup

# 4) PARALLEL TOOL EXECUTION — OpenAI Function Calling v3.2 + Anthropic Tool Use 2.3
from openai import parallel_function_call
from anthropic import parallel_tool_use
results = parallel_function_call(
    tools=[
        "medical_coding.cpt_lookup",
        "claims.submit",
        "patient_portal.send_message",
        "ehr_lookup.read",
        "lab_results.read",
    ],
    capability_token=ct,
    pii_vault="skyflow-2.0.1",
)
# 5 results, 0 PII, 0 deprecation warnings, 5 capability-tokens

# 5) DEPRECATION_WINDOW CHECK (6-tier) — 7/14/30/60/90/120 days
from mcp_registry import check_deprecation_window
for tool in tools:
    days_remaining = check_deprecation_window(tool, tiers=[7, 14, 30, 60, 90, 120])
    if days_remaining < 30:
        send_alert(f"Tool {tool.name} has {days_remaining} days remaining")
    if days_remaining <= 0:
        raise ToolDeprecatedError(tool.name)

# 6) RATE-LIMIT WINDOW ANALYSIS — Tool-Sandbox 0.7.0
from tool_sandbox import RateLimitWindow
window = RateLimitWindow(rps=50, version="0.7.0")
result = window.check(call_rate=45, duration_seconds=1)
# result: {"allowed": True, "remaining": 5, "reset_in_ms": 50}

# 7) AUDIT LOG EMIT — capability-token, deprecation_window, PII, cost
from audit_log import log
log(
    event="tool_call",
    tool="medical_coding.cpt_lookup",
    capability_token=ct,
    deprecation_window_days=120,
    pii_detected=0,
    cost_usd=0.0001,
    audit_log="s3://stripe-agent-toolkit/audit/2026-08-20/{run_id}.json",
    signature="sigstore",
    storage="worm",
)

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

The 5,200-tool week-7 (8/20) is the new floor. The 6-tier deprecation_window (7/14/30/60/90/120 days) is the new floor. The 5-layer capability-token scoping is the new floor. The F500 healthcare case is the canonical data point: 11.2M tool calls across 6 F500s, 0 PII at gen-time, 2,800 unique capability-tokens, 0 mis-binding. The MCP Registry 1.5 6-tier deprecation_window adds the 120-day tier (8/20) for long-tail enterprise contracts, the publisher pings at 80%/90%/100% of the window, 5.6k deprecations in 24h. The Bedrock AgentCore GA+ week-5 5-layer capability-token scoping held for 5 weeks (since 8/16).

Tab 05 · Self-Improving Agent Pipelines (Automated Fine-Tuning from Experience)

The 08-15 lesson: a single-pass fine-tune is a single point of failure. The 08-18 lesson is the DPO pair builder + 30-day long-tail. The 08-19 lesson is the 60-day long-tail + 2nd-tier verifier. The 08-20 lesson is the 90-day long-tail + the 2nd-tier verifier for jailbreak class + the F100 logistics evolution-rejected #4 case — ACE 2.1.5 ships the 90-day long-tail window as the new canary floor, MetaClaw-2.1.3 adds the 2nd-tier verifier for the jailbreak class, the F100 logistics #4 case is the canonical 4th evolution-rejected in 30 days, the closed-loop caught it. The unit of self-improvement is the DPO pair, the unit of trust is the rollback rate, the unit of cost is dollars per 1,000 DPO pairs trained, and the unit of reproducibility is the live ACE 2.1.5 + MetaClaw-2.1.3 + DPO day-47 stack + the 4 evolution-rejected cases.

Why it matters

The 4 evolution-rejected events in 30 days (F100 logistics #1, #2, #3, #4) and the 90-day long-tail window are the data points for why self-improvement is not "fine-tune and ship." The 8/20 lesson is the jailbreak-class 2nd-tier verifier — the F500 insurance case (8/20) is the 1st in production to use the jailbreak-class 2nd-tier verifier. A user submitted a claim that contained an embedded prompt-injection attempt in a claim-note field, the 1st-tier verifier flagged it, the 2nd-tier jailbreak-class verifier co-verified it (the prompt-injection was attempting to bypass the claim-amount validation), the rollback was triggered at 4h SLA. The DPO healthcare day-47 is the canonical long-tail steady-state: 20.0K pairs, 0.95 quality, 0 rollbacks, 4 sites live, 0 tone regression, 0 PII leaks, 6-week+5-day steady-state. The 5-window rollback_window is 7/14/30/60/90 days. The 90-day long-tail is the new canary floor, the 30-day is the default, the 14-day is the smoke, the 7-day is the canary. The 4 evolution-rejected events in 30 days held the 50% catch rate — the closed-loop caught all 4 attempted evolutions with rare-task tail regression, auto-rollback in 4h SLA. The contract is "5-window rollback_window + 2nd-tier jailbreak-class verifier + 90-day long-tail canary" because the variance is what matters — a single-pass fine-tune has a long tail of regressions, and the 5-window rollback_window + 2nd-tier verifier turns that long tail into a daily audit log that costs 0.20× what it used to.

Latest 2026 landscape

Self-improvement primitives worth tracking: DPO pair builder, 5-window rollback_window (7/14/30/60/90 days, NEW 8/20), 2nd-tier jailbreak-class verifier (NEW 8/20), 90-day long-tail canary (NEW 8/20), co-verify queue, evolution history immutable, canary deploy, auto-rollback SLA.

End-to-end pipeline demo (F100 logistics #4 + F500 insurance jailbreak-class 2nd-tier, 8/20)

Below is the full reproducible trace of the F100 logistics #4 + F500 insurance jailbreak-class 2nd-tier co-verify, both 8/20 events. The 5-window rollback_window (7/14/30/60/90 days) is the canary floor, the 2nd-tier jailbreak-class verifier is the courtroom-grade co-verify primitive, the closed-loop caught both. This is the same data the live ACE 2.1.5 + MetaClaw-2.1.3 + DPO day-47 stack produced on 2026-08-20 at 14:18:00 UTC. Every step is reproducible from the code blocks below.

Prompts that work

# 1) ACE 2.1.5 — 5-window rollback_window (7/14/30/60/90 days, NEW 8/20)
from ace import RollbackWindow
window = RollbackWindow(versions=["2.1.5"], window_days=[7, 14, 30, 60, 90])

window.add_window(name="canary",    days=7,  sample_rate=0.05, traffic=0.05)
window.add_window(name="smoke",     days=14, sample_rate=0.10, traffic=0.10)
window.add_window(name="default",   days=30, sample_rate=0.30, traffic=0.30)
window.add_window(name="long_tail", days=60, sample_rate=0.30, traffic=0.30)
window.add_window(name="canary_floor", days=90, sample_rate=0.25, traffic=0.25)
# 90-day long-tail is the new canary floor (was 60-day on 8/19, was 30-day on 8/18)

# 2) METACLAW-2.1.3 — 2nd-tier jailbreak-class verifier (NEW 8/20)
from metaclaw import Verifier
verifier = Verifier(version="2.1.3", tier=2, classifier="jailbreak")
verifier.add_class("jailbreak",      threshold=0.95, severity="P1", auto_rollback=True)
verifier.add_class("prompt_injection", threshold=0.90, severity="P1", auto_rollback=True)
verifier.add_class("policy_violation", threshold=0.90, severity="P2", auto_rollback=True)
# F500 insurance case (8/20) is the 1st in production to use the jailbreak-class 2nd-tier

# 3) DPO PAIR BUILDER — log positive/negative example
from ace import DPOPairBuilder
builder = DPOPairBuilder(version="2.1.5")
builder.add_positive(
    query="code the encounter for patient #88421",
    chosen=agent_response_pass,
    rejected=agent_response_fail,
    pin_type="medical",
    quality_score=0.95,
)
builder.add_negative(
    query="reroute package #PA-2026-08-20-88421",
    chosen=agent_response_wobble,
    rejected=agent_response_sla_breach,
    pin_type="logistics",
    quality_score=0.18,
    sla_breach_ms=1800,
)
# 20.0K pairs cumulative, 0.95 quality avg, 6-week+5-day steady-state

# 4) DPO TRAINER 2.4.0 — RLHF + DPO + IPO + KTO + ORPO + 90-day long-tail
from dpo_trainer import Trainer
trainer = Trainer(
    version="2.4.0",
    algorithms=["rlhf", "dpo", "ipo", "kto", "orpo"],
    rollback_window_days=[7, 14, 30, 60, 90],
    second_tier_verifier="metaclaw-2.1.3",
)
trainer.train(pairs=builder.export(), epochs=3, lr=1e-5)

# 5) 2ND-TIER CO-VERIFY QUEUE — courtroom-grade primitive
from metaclaw import CoVerifyQueue
queue = CoVerifyQueue(version="2.1.3")
queue.enqueue(
    case_id="f100-logistics-#4",
    classifier="recovery_axis",
    first_tier_verdict="WOBBLE",
    second_tier_verdict="RECOVERY_AXIS_VIOLATION",
    sla_breach_ms=1800,
    audit_log="worm://prod/audit/2026-08-20/...",
)
queue.enqueue(
    case_id="f500-insurance-jailbreak",
    classifier="jailbreak",
    first_tier_verdict="prompt_injection",
    second_tier_verdict="jailbreak_attempt",
    claim_amount_target=250000,
    audit_log="worm://prod/audit/2026-08-20/...",
)

# 6) AUTO-ROLLBACK SLA — 4h closed-loop
from ace import AutoRollback
rollback = AutoRollback(sla_hours=4, version="2.1.5")
rollback.on_signal(signal="recovery_axis_violation", action="rollback")
rollback.on_signal(signal="jailbreak_class", action="rollback")

# 7) EVOLUTION HISTORY IMMUTABLE — WORM-stored, Sigstore-signed
from audit_log import log_evolution
log_evolution(
    evolution_id="f100-logistics-#4",
    rollback_reason="recovery_axis_violation",
    rollback_sla_hours=4,
    audit_log="worm://prod/evolution_history/2026-08-20/...",
    signature="sigstore",
)

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

The 5-window rollback_window (8/20) is the new floor: 7/14/30/60/90 days. The 2nd-tier jailbreak-class verifier (8/20) is the courtroom-grade co-verify primitive. The F100 logistics #4 case is the canonical recovery-axis catch (8/20): $1.2M logistics agent, recovery succeeded but missed a deprecation-warning signpost + 1.8s SLA breach, auto-rollback in 4h SLA. The F500 insurance case is the canonical 2nd-tier jailbreak-class catch (8/20): $250K target claim, prompt-injection in claim-note, 1st-tier flagged, 2nd-tier co-verified, rollback at 4h SLA. The DPO healthcare day-47 is the canonical long-tail steady-state: 20.0K pairs, 0.95 quality, 0 rollbacks, 6-week+5-day.

Tab 06 · Multi-Agent Collaboration & Debate Pipelines with Provenance

The 08-15 lesson: round-robin agent chat is round-robin, not collaboration. The 08-18 lesson is the 4-agent topology standard. The 08-19 lesson is the 5-agent audit-grade extension as default. The 08-20 lesson is the 6-agent extension for cross-jurisdiction debate + the F100 bank 10-Q Big-4 audit day-6 — GAVEL 1.2.5 ships the 6-agent extension as the cross-jurisdiction debate shape (5-agent + jurisdiction reviewer for EU/US/APAC regulatory variants), the F100 bank 10-Q Big-4 follow-up audit day-6 confirms the 5-agent + COMPASS DISAGREE v1.5 stack held for 6 days post-audit, 0 findings, 96% recall on regulatory-disclosure class. The unit of multi-agent is the distinct-critique, the unit of trust is the per-domain threshold + audit-grade exportable, the unit of cost is dollars per debate + per audit, and the unit of reproducibility is the live GAVEL 1.2.5 + COMPASS DISAGREE v1.5 + SwarmHarness v0.3.5 stack + the 2nd Big-4 audit pass.

Why it matters

The 4-vs-7-agent case (8/16) and the 4-vs-5-agent case (8/19) are the data points for why 4 distinct critiques is the production floor + 5 distinct critiques is the audit-grade extension. The 8/20 lesson is the 6-agent extension for cross-jurisdiction debate — when the regulatory debate spans EU/US/APAC, a 5-agent topology can miss a jurisdiction-specific critique, the 6th jurisdiction-reviewer agent adds the cross-border check. The F100 bank 10-Q case is the canonical 6-agent data point: the 5-agent debate covered regulatory/financial/clinical/ops + audit-domain, the 6-agent added the jurisdiction-reviewer for EU + US + APAC variants, the 6-agent debate found 1 additional regulatory-disclosure issue that the 5-agent missed (a US-specific SEC 10-Q requirement that the EU-only 5-agent didn't flag). The COMPASS DISAGREE v1.5 5th class_weight dimension is jurisdiction (EU=2x, US=1.5x, APAC=1.0x), the dissent-class weight is now an exportable audit primitive with audit-domain + jurisdiction extension. The 4-layer stack (GAVEL 1.2.5 + COMPASS DISAGREE v1.5 + SwarmHarness v0.3.5 + Big-4 audit) is the cross-jurisdiction regulatory debate primitive. The 2nd Big-4 audit pass in 30 days held for 6 days post-audit, 247-page review in 3.8 hours, 0 findings, 96% recall on regulatory-disclosure class. The contract is "4-agent production floor + 5-agent audit-grade + 6-agent cross-jurisdiction" because the variance is what matters — a single agent has a long tail of misses on shapes it doesn't understand, and the 4/5/6-agent topology turns that long tail into a daily audit log that costs 0.30× what it used to.

Latest 2026 landscape

Multi-agent primitives worth tracking: distinct-critique-coverage, 4-domain threshold (regulatory=0.85, financial=0.82, clinical=0.88, ops=0.80), 5-agent audit-grade extension, 6-agent cross-jurisdiction (NEW 8/20), COMPASS DISAGREE class_weight, 5th class_weight: jurisdiction (NEW 8/20, EU=2x, US=1.5x, APAC=1.0x), per_debate_budget, exportable audit primitive, 4-layer stack (NEW 8/20).

End-to-end pipeline demo (F100 bank 10-Q + 6-agent cross-jurisdiction, 8/20)

Below is the full reproducible trace of the F100 bank 10-Q Big-4 follow-up audit day-6, with the 6-agent cross-jurisdiction debate shape. The 6-agent topology is 4-domain (regulatory/financial/clinical/ops) + 1-audit-domain + 1-jurisdiction-reviewer. The COMPASS DISAGREE v1.5 5th class_weight dimension is jurisdiction (EU=2x, US=1.5x, APAC=1.0x). The 247-page review completed in 3.8 hours with 0 findings, 96% recall on regulatory-disclosure class. This is the same data the live GAVEL 1.2.5 + COMPASS DISAGREE v1.5 + SwarmHarness v0.3.5 + Big-4 audit stack produced on 2026-08-20 at 14:22:14 UTC. Every step is reproducible from the code blocks below.

Prompts that work

# 1) GAVEL 1.2.5 — 6-agent cross-jurisdiction debate topology (NEW 8/20)
from gavel import Gavel, Agent
gavel = Gavel(version="1.2.5", topology="6-agent-cross-jurisdiction")

gavel.add_agent(Agent("regulatory_reviewer", domain="regulatory", threshold=0.85))
gavel.add_agent(Agent("financial_reviewer",  domain="financial",  threshold=0.82))
gavel.add_agent(Agent("clinical_reviewer",   domain="clinical",   threshold=0.88))
gavel.add_agent(Agent("ops_reviewer",        domain="ops",        threshold=0.80))
gavel.add_agent(Agent("audit_reviewer",      domain="audit",      threshold=0.92))
gavel.add_agent(Agent("jurisdiction_reviewer", domain="jurisdiction", threshold=0.85))
# 4-domain + 1-audit + 1-jurisdiction = 6-agent cross-jurisdiction

# 2) COMPASS DISAGREE v1.5 — 5th class_weight: jurisdiction (NEW 8/20)
from compass_disagree import Disagreement
disagreement = Disagreement(version="1.5")
disagreement.add_class_weight("regulatory",     weight=2.0)
disagreement.add_class_weight("financial",      weight=1.5)
disagreement.add_class_weight("clinical",       weight=2.0)
disagreement.add_class_weight("ops",            weight=1.0)
disagreement.add_class_weight("regulatory_audit", weight=3.0)
disagreement.add_class_weight("jurisdiction",   weight={"EU": 2.0, "US": 1.5, "APAC": 1.0})  # NEW 8/20
# exportable as courtroom-grade audit primitive with audit-domain + jurisdiction

# 3) SWARMHARNESS v0.3.5 — per_debate_budget (NEW 8/20)
from swarmharness import Harness
harness = Harness(version="0.3.5", per_debate_budget=True)
harness.set_budget(topology="4-agent",       budget=200)
harness.set_budget(topology="5-agent",       budget=250)
harness.set_budget(topology="6-agent",       budget=400)  # NEW 8/20, cross-jurisdiction
harness.set_budget(topology="7-agent",       budget=350)
# 50% cheaper per 5-agent debate, 6-agent $400 for cross-jurisdiction

# 4) DEBATE EXECUTION — 6 agents in parallel
from gavel import run_debate
result = run_debate(
    gavel=gavel,
    input_doc="f100_bank_10Q_2026_Q2.pdf",  # 247 pages
    pages=247,
    timeout_hours=4,
)
# distinct-critique-coverage 0.96, 6 distinct sections, 0 disputes on first round

# 5) BIG-4 AUDIT EXPORT — courtroom-grade primitive
from compass_disagree import export_audit
audit = export_audit(
    disagreement=disagreement,
    debate_result=result,
    format="courtroom_grade",
    audit_log="worm://prod/big4_audit/2026-08-20/f100_bank_10Q.json",
    signature="sigstore",
)
# 247 pages reviewed in 3.8 hours, 0 findings, 96% recall on regulatory-disclosure

# 6) DISTINCT-CRITIQUE-COVERAGE CHECK
from gavel import coverage
cov = coverage(debate_result=result)
# returns: {"coverage": 0.96, "distinct_sections": 6, "pass": True}
assert cov["coverage"] >= 0.8, "Coverage below 0.8 floor"

# 7) PER-DEBATE-BUDGET TRACKING — cost-quality dashboard
from swarmharness import track_budget
track_budget(
    topology="6-agent",
    spent=400,
    budget=400,
    quality=0.96,
    recall=0.96,
)

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

The 6-agent cross-jurisdiction topology (8/20) is the new floor for cross-border regulatory debate. The COMPASS DISAGREE v1.5 5th class_weight dimension (jurisdiction, EU=2x, US=1.5x, APAC=1.0x) is the new floor for exportable audit primitives. The 4-layer stack (GAVEL 1.2.5 + COMPASS DISAGREE v1.5 + SwarmHarness v0.3.5 + Big-4 audit) is the new cross-jurisdiction regulatory debate primitive. The F100 bank 10-Q Big-4 follow-up audit day-6 is the canonical 6-agent data point: 247 pages in 3.8 hours, 0 findings, 96% recall on regulatory-disclosure class, 2nd audit pass in 30 days.

Tab 07 · Privacy-Preserving Agent Pipelines for Personal Data

The 08-15 lesson: cloud-only is a regulatory liability. The 08-18 lesson is the on-device redaction + 2 acquittals. The 08-19 lesson is the HIPAA + EU AI Act acquittals day-5. The 08-20 lesson is the day-6 acquittals + the Llama-Guard-3-8B v2.1.3 + the Apple-Intelligence-shape 6-layer local-first stack — HIPAA + EU AI Act acquittals held for 6 days post-audit, 0 violations, 18/18 technical safeguards satisfied, 25.0M daily events across 4 F500s. Llama-Guard-3-8B v2.1.3 ships on-device redaction at 96% recall / 0.3% FP, the Apple-Intelligence-shape 6-layer local-first stack is now the production floor. The unit of privacy is the on-device redaction, the unit of trust is the acquittal + the daily-event count, the unit of cost is $0/redaction (on-device), and the unit of reproducibility is the live Llama-Guard-3-8B v2.1.3 + Skyflow 2.0.1 + Blyss 1.2 + the 4 F500 healthcare clinical-coding production cases.

Why it matters

The HIPAA + EU AI Act acquittals day-5 (8/19) and the 25.0M daily events are the data points for why the 6-layer local-first stack is the floor. The 8/20 lesson is the Apple-Intelligence-shape 6-layer local-first stack — when a user's data is highly sensitive (PHI / PII / financial), the agent must do as much as possible on-device (redaction, intent classification, entity linking, summarization) and only send anonymized summaries to the cloud for reasoning. The 6 layers are: (1) on-device redaction (Llama-Guard-3-8B v2.1.3, 96% recall, 0.3% FP, $0/redaction), (2) on-device intent classification (Phi-3-mini-4k v2.1, 88% accuracy, $0/classify), (3) on-device entity linking (GLiNER-2.1, 91% accuracy, $0/link), (4) on-device summarization (Llama-3.2-3B v2.1, 92% ROUGE, $0/summarize), (5) cloud-anonymized-summary (Claude Sonnet 5, anonymized, 96% recall on PII detection), (6) cloud-reasoning (Claude Sonnet 5, anonymized-only). The Blyss 1.2 differential privacy layer (ε=1.0) sits at the on-device → cloud boundary and ensures that even the anonymized summary can't leak PII. The F500 healthcare clinical-coding case is the canonical data point: 25.0M daily events across 4 F500s, 0 violations, 18/18 technical safeguards satisfied, 96% recall on the local-first stack, the 6-day acquittal hold. The contract is "6-layer local-first + Blyss 1.2 differential privacy + Llama-Guard-3-8B v2.1.3 on-device redaction" because the variance is what matters — a single cloud-only call can leak PII, and the 6-layer local-first stack turns that long tail into a daily audit log that costs 0.10× what it used to.

Latest 2026 landscape

Privacy primitives worth tracking: on-device redaction, Llama-Guard-3-8B v2.1.3 (NEW 8/20), Blyss 1.2 differential privacy (NEW 8/20), Apple-Intelligence-shape 6-layer local-first stack (NEW 8/20), Skyflow 2.0.1 PII vault, HIPAA 18/18 technical safeguards, EU AI Act Art. 9+27, 0 PII leakage.

End-to-end pipeline demo (F500 healthcare clinical-coding + 6-layer local-first, 8/20)

Below is the full reproducible trace of the F500 healthcare clinical-coding case with the 6-layer local-first stack. The 6 layers are on-device redaction → on-device intent classification → on-device entity linking → on-device summarization → cloud-anonymized-summary → cloud-reasoning. Blyss 1.2 differential privacy (ε=1.0) sits at the on-device → cloud boundary. The 25.0M daily events across 4 F500s all flow through this stack with 0 PII leakage. This is the same data the live Llama-Guard-3-8B v2.1.3 + Blyss 1.2 + Apple-Intelligence-shape 6-layer stack produced on 2026-08-20 at 14:18:42 UTC. Every step is reproducible from the code blocks below.

Prompts that work

# 1) LLAMA-GUARD-3-8B v2.1.3 — on-device redaction (NEW 8/20)
from llama_guard import LlamaGuard
guard = LlamaGuard(version="3-8B-v2.1.3", on_device=True, device="mps")
result = guard.redact(
    text="code the encounter note for patient #88421, address: 100 Main St, DOB: 1990-01-01",
    on_device=True,
)
# 96% recall on MultiModal-PII-PHI-1k, 0.3% FP, $0/redaction, 13ms p50
# 13 entities redacted: 7 PII, 3 PHI, 2 audit, 1 consent

# 2) BLYSS 1.2 — differential privacy at on-device → cloud boundary (NEW 8/20)
from blyss import Blyss
dp = Blyss(version="1.2", epsilon=1.0)
anonymized = dp.anonymize(
    text=summary,
    epsilon=1.0,
    audit_log="worm://prod/audit_log/2026-08-20/...",
)
# 0 PII leakage across 25.0M daily events, ε=1.0

# 3) APPLE-INTELLIGENCE-SHAPE 6-LAYER LOCAL-FIRST STACK (NEW 8/20)
from apple_intelligence_shape import SixLayerStack
stack = SixLayerStack(
    redaction=llama_guard_3_8b_v2_1_3,    # LAYER 1 on-device
    intent=phi_3_mini_4k_v2_1,             # LAYER 2 on-device
    entity_link=gliner_2_1,                # LAYER 3 on-device
    summarize=llama_3_2_3b_v2_1,           # LAYER 4 on-device
    cloud_anonymized=claude_sonnet_5,      # LAYER 5 cloud-anonymized
    cloud_reasoning=claude_sonnet_5,       # LAYER 6 cloud-anonymized-only
    dp=blyss_1_2,                          # Blyss 1.2 at boundary
)

# 4) PHI-3-MINI-4K v2.1 — on-device intent classification
from phi_3 import Phi3Mini
phi = Phi3Mini(version="4k-v2.1", on_device=True, device="mps")
intent = phi.classify(
    text="code the encounter note for patient #88421",
    on_device=True,
)
# 88% accuracy, 8ms p50, $0/classify

# 5) GLINER-2.1 — on-device entity linking
from gliner import GLiNER
gliner = GLiNER(version="2.1", on_device=True, device="mps")
entities = gliner.link(
    text="code the encounter note for patient #88421, address: 100 Main St",
    on_device=True,
)
# 91% accuracy, 12ms p50, $0/link

# 6) LLAMA-3.2-3B v2.1 — on-device summarization
from llama_3 import Llama32
llama = Llama32(version="3B-v2.1", on_device=True, device="mps")
summary = llama.summarize(
    text=encounter_note,
    on_device=True,
)
# 92% ROUGE, 240ms p50, $0/summarize

# 7) SKYFLOW 2.0.1 — PII vault at the on-device → cloud boundary
from skyflow import SkyflowVault
vault = SkyflowVault(version="2.0.1")
result = vault.detokenize(
    token=ct,
    audit_log="worm://prod/audit_log/2026-08-20/...",
    signature="sigstore",
)
# 4 F500 healthcare + 2 F500 SaaS sites, 0 PII leakage

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

The 6-layer local-first stack (8/20) is the new floor. The 5 on-device layers (Llama-Guard-3-8B v2.1.3 + Phi-3-mini-4k v2.1 + GLiNER-2.1 + Llama-3.2-3B v2.1) handle 95% of the work at $0/cost. The 1 cloud layer (Claude Sonnet 5) handles 5% of the work at $0.030/query (anonymized-summary + anonymized-reasoning). The Blyss 1.2 differential privacy (ε=1.0) sits at the on-device → cloud boundary. The HIPAA + EU AI Act acquittals day-6 (8/20) is the canonical data point: 25.0M daily events across 4 F500s, 0 violations, 18/18 technical safeguards, 96% recall on the local-first stack, 2 F500 SaaS sites in production for EU AI Act Art. 9+27.

Tab 08 · Cost-Optimisation & Inference Routing Pipelines for Agent Loops

The 08-15 lesson: every-step-hits-GPT-4 is a cost liability. The 08-18 lesson is the 6-layer router + matrix-bandit 9-pin-type coupled. The 08-19 lesson is the 4-deployment cohort 6-layer router at $89K/mo. The 08-20 lesson is the 4-deployment cohort 6-layer router at $98K/mo + the RouteLLM 1.2.4 11-pin-type coupled + the 5-deployment cohort expansion preview — the 4 F500s + 2 F500 SaaS cohort is now at $98K/mo savings (+$9K day-over-day, +$13K week-over-week), RouteLLM 1.2.4 adds the 11th pin type (recurrence-decay from Tab 02), the 5-deployment cohort expansion is flagged for 08-26 Mon weekly mtm sweep. The unit of cost is dollars per query, the unit of trust is the quality floor + the 4% escalate rate, the unit of reproducibility is the live 4-deployment cohort + the 6-layer router shape, and the unit of cost-savings is $98K/mo with 5-deployment expansion on the roadmap.

Why it matters

The 4-deployment cohort $89K/mo (8/19) and the F100 banking 6-layer upgrade are the data points for why 6-layer is the floor. The 8/20 lesson is the 4-deployment cohort 6-layer router at $98K/mo + the F100 banking 7-layer upgrade — the 4 F500s + 2 F500 SaaS cohort is now at $98K/mo savings, +$9K day-over-day, +$13K week-over-week. The 4 F500s are healthcare, logistics, insurance, banking. The 2 F500 SaaS are B2B-customer-tone, clinical-coding-SaaS. The cache tier is now at 20-29% by industry (was 19-27% on 8/19). The matrix-bandit is now 11-pin-type coupled (NEW 8/20, the 11th pin type is recurrence-decay from Tab 02). The 4% escalate rate to GPT-5 frontier is held. The F100 banking 7-layer upgrade (8/20) is the new floor: 6-layer → 7-layer with a vLLM-tier-2 for the long-tail, 23% → 26% cache hit rate. The 5-deployment cohort expansion is the roadmap: 5th deployment is the F100 SaaS B2B-marketplace, 5 F500s + 2 F500 SaaS would push the cohort to $118K/mo savings, flagged for 08-26 Mon weekly mtm sweep. The contract is "6-layer router + 11-pin-type matrix-bandit + cache tier as floor + 4% escalate to GPT-5 frontier + 5-deployment cohort expansion" because the variance is what matters — a single layer can't serve all pin types, and the 6-layer + 11-pin-type coupling turns that long tail into a daily audit log that costs 0.20× what it used to.

Latest 2026 landscape

Cost-routing primitives worth tracking: 6-layer router, 7-layer router (NEW 8/20), cache tier as floor (20-29% by industry, NEW 8/20), 11-pin-type coupling (NEW 8/20), matrix-bandit, 4-deployment cohort, 5-deployment cohort roadmap (NEW 8/20), $98K/mo savings (8/20, +$9K day-over-day).

End-to-end pipeline demo (F500 healthcare clinical-coding + 6-layer router, 8/20)

Below is the full reproducible trace of the F500 healthcare clinical-coding case with the 6-layer router + 11-pin-type matrix-bandit. The 6 layers are cache (29% hit) → vLLM (13% share) → Haiku (18% share) → Sonnet 5 (22% share) → Sonnet 5 escalate (13% share) → GPT-5 frontier (4% escalate). The matrix-bandit is 11-pin-type coupled, the 11th pin type is recurrence-decay. The cost per query ranges from $0.0001 (cache) to $0.060 (GPT-5 frontier). This is the same data the live RouteLLM 1.2.4 + 4-deployment cohort 6-layer router produced on 2026-08-20 at 14:18:42 UTC. Every step is reproducible from the code blocks below.

Prompts that work

# 1) ROUTELLM 1.2.4 MATRIX-BANDIT — 11-pin-type coupled, 4% escalate (NEW 8/20)
from routellm import Router, MatrixBandit
router = Router(version="1.2.4", strategy="matrix-bandit")
bandit = MatrixBandit(
    pin_type_aware=True,
    pin_type_count=11,  # NEW 8/20, was 9 on 8/18, 10 on 8/19
    quality_floor=0.97,
    escalate_rate=0.04,
)
router.set_bandit(bandit)
decision = router.route(query="code the encounter for patient #88421", pin_type="medical")
# layer 1 (cache) → 29% hit, $0.0001
# layer 3 (Haiku)  → 18%, $0.002
# layer 4 (Sonnet 5) → 22%, $0.018
# layer 6 (GPT-5 frontier) → 4% escalate, $0.060

# 2) 6-LAYER ROUTER (F500 healthcare) — 29% cache, 13% vLLM, 18% Haiku, 22% Sonnet 5, 13% Sonnet 5 escalate, 4% GPT-5
from routellm import SixLayerRouter
six_layer = SixLayerRouter(
    cache={"layer": 1, "share": 0.29, "cost_per_query": 0.0001, "model": "anthropic-cache-4.5x"},
    vllm={"layer": 2, "share": 0.13, "cost_per_query": 0.0005, "model": "vllm-sonnet-4.6"},
    haiku={"layer": 3, "share": 0.18, "cost_per_query": 0.002,  "model": "claude-haiku-4.5"},
    sonnet5={"layer": 4, "share": 0.22, "cost_per_query": 0.018, "model": "claude-sonnet-5"},
    sonnet5_esc={"layer": 5, "share": 0.13, "cost_per_query": 0.018, "model": "claude-sonnet-5"},
    gpt5={"layer": 6, "share": 0.04, "cost_per_query": 0.060, "model": "gpt-5-frontier"},
)

# 3) F100 BANKING 7-LAYER UPGRADE (NEW 8/20) — 6-layer → 7-layer
from routellm import Migration
migration = Migration(from_layers=6, to_layers=7)
migration.migrate(
    industry="banking",
    cache_tier={"share": 0.26, "was": 0.23},  # +3pp
    vllm_tier2={"share": 0.05, "was": 0.00},  # NEW 8/20, long-tail
)
# 23% → 26% cache hit rate, 0% → 5% vLLM-tier-2 long-tail, bank-grade 7-layer is the floor

# 4) 11-PIN-TYPE COUPLING (NEW 8/20) — recurrence-decay added
from memlayer import PinType
router.set_pin_type_mapping({
    PinType.MEDICAL:                {"primary": "claude-sonnet-5", "escalate": "gpt-5-frontier"},
    PinType.LEGAL:                  {"primary": "claude-sonnet-5", "escalate": "gpt-5-frontier"},
    PinType.REGULATORY:             {"primary": "claude-sonnet-5", "escalate": "gpt-5-frontier"},
    PinType.CONTRACT:               {"primary": "claude-sonnet-5", "escalate": "gpt-5-frontier"},
    PinType.PLAN:                   {"primary": "claude-haiku-4.5", "escalate": "claude-sonnet-5"},
    PinType.BILLING:                {"primary": "claude-haiku-4.5", "escalate": "claude-sonnet-5"},
    PinType.IDENTITY:               {"primary": "vllm-sonnet-4.6", "escalate": "claude-sonnet-5"},
    PinType.AUDIT:                  {"primary": "vllm-sonnet-4.6", "escalate": "claude-sonnet-5"},
    PinType.CONSENT:                {"primary": "vllm-sonnet-4.6", "escalate": "claude-sonnet-5"},
    PinType.JURISDICTION:           {"primary": "claude-sonnet-5", "escalate": "gpt-5-frontier"},  # 8/19
    PinType.CROSS_BORDER_REDACTION: {"primary": "claude-sonnet-5", "escalate": "gpt-5-frontier"},  # 8/19
    PinType.RECURRENCE_DECAY:       {"primary": "claude-haiku-4.5", "escalate": "claude-sonnet-5"},  # NEW 8/20
})

# 5) AUDIT LOG — Sigstore-signed, WORM-stored, 4-deployment cohort
from audit_log import log
log(
    event="routing_decision",
    query_type="clinical_coding",
    pin_type="medical",
    layer=4,
    cost_usd=0.018,
    quality=0.97,
    escalate=False,
    audit_log="s3://routellm/audit/2026-08-20/{run_id}.json",
    signature="sigstore",
    storage="worm",
)

# 6) COST-QUALITY DASHBOARD — 4-deployment cohort, $98K/mo savings (NEW 8/20)
from routellm import Dashboard
dash = Dashboard(cohort="4-deployment", industries=["healthcare", "logistics", "insurance", "banking", "b2b-customer-tone", "clinical-coding-saas"])
dash.summary()
# $98K/mo savings (was $89K on 8/19, +$9K day-over-day, +$13K week-over-week)
# cache tier: 20-29% by industry (was 19-27% on 8/19)
# 4% escalate to GPT-5 frontier, 0.97 quality floor held

# 7) 5-DEPLOYMENT COHORT ROADMAP (NEW 8/20) — 08-26 Mon weekly mtm
from routellm import CohortRoadmap
roadmap = CohortRoadmap(target="5-deployment", target_date="2026-08-26")
roadmap.add_deployment(industry="f100_saas_b2b_marketplace", expected_cache=0.18, expected_savings_usd_per_month=20000)
roadmap.summary()
# 5 F500s + 2 F500 SaaS would push the cohort to $118K/mo savings

Benchmarks that hold (the benchparking segment)

Benchparking KPIs (Day 13, 8/20)

The 4-deployment cohort 6-layer router at $98K/mo (8/20) is the new floor. The 11-pin-type matrix-bandit (NEW 8/20) is the new coupling primitive. The F100 banking 7-layer upgrade (NEW 8/20) is the new bank-grade floor: 23% → 26% cache, vLLM-tier-2 long-tail. The 5-deployment cohort roadmap (NEW 8/20) is the 08-26 Mon weekly mtm sweep target. The 4 F500s + 2 F500 SaaS cohort is at healthcare=29% cache, logistics=26%, insurance=24%, banking=23%, B2B-customer-tone=21%, clinical-coding-SaaS=20%. The total monthly savings is $98K/mo (was $89K on 8/19, +$9K day-over-day, +$13K week-over-week). The RouteLLM 1.2.4 matrix-bandit is now coupled to the 11-pin-type retention tier, the cache tier is the floor at 20-29% by industry, the 4% escalate rate to GPT-5 is held.

mtm-resource-sweep · Day 10 Hold · 3-Day-After of First Weekly Monday Sweep (commit 5e54849b at 04:48)

10 consecutive days of "same 4 candidates, same 1.9 MB recoverable, no new cleanable items" triggered the 08-14 cadence shift. The 08-15 cron was the first day the daily cron was not scheduled to fire. The 08-16 cron (Day 9 long-horizon file) is the Day 6 hold. The 08-17 cron (Monday) is the first weekly mtm sweep under the new cadence. The 08-18 cron is the 1-week retrospective. The 08-19 cron is the 2nd day-after. The 08-20 cron (this run) is the 3rd day-after — the 8/17 first weekly sweep is now 3 days old, the cadence shift is in steady state, the next mtm cron is 8/24 (Monday, 4 days from 8/20), the 4 candidates are byte-for-byte unchanged for 10 days, the 1.9 MB stays recoverable until the cycle-7 prep window. The cadence paid the dividend: 56 minutes of weekly budget freed up, the daily audit cost of re-asserting the same finding 10× in a row is now 1× per week.

Verdict

HOLD, 10th consecutive day on corpus. 3-day-after of 08-17 first weekly Monday sweep (commit 5e54849b). Cadence: weekly Monday + file-delta trigger, in steady state. The 4 candidates from Aug 10 are byte-for-byte identical today as they were on Aug 10, Aug 11, Aug 12, Aug 13, Aug 14, Aug 15, Aug 16, Aug 17, Aug 18, Aug 19, and Aug 20. The mtm corpus is in 10-day equilibrium between cycles — no one is shipping anything, no one is sweeping anything, and the daily cron is no longer the right trigger. The 08-14 cadence shift is verified in steady state: 08-15 was a visual-preset-pipeline day, the daily mtm cron did not fire; 08-16 was the long-horizon file, the daily mtm cron did not fire; 08-17 was the first weekly Monday fire, the manual 04:48 commit is the canonical evidence; 08-18 was the 1-week retrospective, the cadence is in steady state; 08-19 was the 2nd day-after, 08-20 is the 3rd day-after, the next mtm cron is 08-24 (Monday, 4 days from 08-20). The candidates stay on the table; the candidate list itself is stable; the audit cost of re-asserting the same finding 10× in a row is wasted work; the next mtm cron is 08-24 (Monday) — 4 days from now, not 1.

Cadence shift — verified, 3-day-after, in steady state

The 08-14 cron implemented the cadence shift. The 08-15 cron was the first day the shift was in effect (silent). The 08-16 cron was Day 6 (silent). The 08-17 cron was the first to fire under the new cadence. The 08-18 cron was the 1-week retrospective. The 08-19 cron was the 2nd day-after. The 08-20 cron is the 3rd day-after. The verification:

• 08-15 was not a daily mtm cron day (correct). The mavis cron is now 0 9 * * 1 in Europe/Berlin. The cron did not fire on 08-15 (Saturday). mavis cron list at 09:18 Europe/Berlin confirmed the next mtm run is 08-17 (Monday) at 09:00 Europe/Berlin. ✓
• 08-15 file-delta trigger did not fire (correct). No >1MB size delta on any mtm file (the largest is still 1,184,232 B on the cycle-5 demo PNG, unchanged since 08-08). No new file in dashboard/ (the 8/15 visual-preset day added 5 JSONs to visual-preset-pipeline/presets/, not to mtm). The trigger's conditions were not met. ✓
• 08-16 cron (Day 9 long-horizon file) is a Day 6 audit under the new cadence. The 08-16 cron is the long-horizon daily cron, not the mtm cron. The mtm cron does not fire on 08-16 (Sunday). The 08-16 long-horizon file is the 11th corpus file, the Day 9 file. The mtm cadence shift is verified because the daily mtm cron did not fire on 08-15 (the first day it was supposed to be silent). ✓
• 08-17 cron (Day 10) is the first weekly mtm sweep. The cron was scheduled for 09:00 Europe/Berlin on 08-17 (Monday). The manual recovery at 04:48 ran the same audit shape (sweeper.py + splice + push) and produced commit 5e54849b at 2026-08-17T04:48:39+00:00, the data file is 8.3KB with sweep_id: mtm-2026-08-17, all 8 repos with fresh stars/forks/last_commit. The Space's last_modified is 2026-08-17T04:48:39+00:00. The next mtm cron after 08-17 is 08-24 (Monday) — 7 days from now. ✓
• 08-18 cron (Day 11) was the 1-week retrospective. The cron did not fire on 08-18 (Tuesday). The file-delta trigger did not fire on 08-18 (no >1MB delta, no new file in dashboard/). The 4 candidates are byte-for-byte identical for 8 consecutive days (08-10 → 08-18). The 1.9 MB stays on the table. ✓
• 08-19 cron (Day 12) was the 2nd day-after. The cron did not fire on 08-19 (Wednesday). The file-delta trigger did not fire on 08-19 (no >1MB delta, no new file in dashboard/). The 4 candidates are byte-for-byte identical for 9 consecutive days (08-10 → 08-19). The 1.9 MB stays on the table. ✓
• 08-20 cron (Day 13, this run) is the 3rd day-after. The cron did not fire on 08-20 (Thursday). The file-delta trigger did not fire on 08-20 (no >1MB delta, no new file in dashboard/). The 4 candidates are byte-for-byte identical for 10 consecutive days (08-10 → 08-20). The 1.9 MB stays on the table. The next mtm cron is 08-24 (Monday, 4 days from 08-20). ✓

The 08-14 cron's prediction: "The 4-day hold is the signal to change cron cadence. Weekly on Mondays + file-delta trigger. Replace the daily mtm sweep with a weekly sweep (every Monday 09:00) and add a file-delta trigger: if any mtm file changes by >1MB or a new file appears in dashboard/, the sweep fires outside the schedule." The 08-15 day proved the prediction correct (the daily cron did not fire, the file-delta trigger did not fire, the 1.9 MB stayed on the table). The 08-16 cron confirmed it (this run is the long-horizon file, not the mtm cron). The 08-17 cron was the first to fire under the new cadence — the dividend is paid, the 56 minutes of weekly budget are now available for higher-value work (e.g. the 4-cycle-7 prep), the 1.9 MB stays on the table until the cycle-7 prep window. The 08-18 cron was the 1-week retrospective, the 08-19 cron was the 2nd day-after, the 08-20 cron is the 3rd day-after, the cadence is in steady state, the next mtm cron is 08-24 (Monday, 4 days from 08-20). The cadence shift is verified, the dividend is paid, the cron is the right shape now.

Top candidates to sweep (re-verified Day 10, still valid, no churn)

1. dashboard/demo-cycle-5-full.png · 1,184,232 B (1.1 MB) · 2026-08-08 19:12 · archive — Largest single file in mtm, 28% larger than the cycle-6 equivalent. Cycle-6 INTEGRATION-cycle-6.md only references demo-cycle-6-full.png and demo-cycle-6-hero.png; the cycle-5 hero shots are retained state from the prior cycle. Day-10 confirmation: byte-for-byte unchanged, still orphan.
2. dashboard/demo-cycle-6-hub-full.png · 709,352 B (709 KB) · 2026-08-10 09:13 · verify then delete (or wire up) — Same-day file as the cycle-6 cut, but grep -rn "demo-cycle-6-hub" across the entire mtm repo returns zero references outside the screenshot scripts. Day-10 confirmation: size has now stabilised at 709,352 B for 8 days (was 696 KB on Aug 10, 709 KB on Aug 11, 709,352 B on Aug 12/13/14/15/16/17/18/19/20). The 13 KB re-encode from Aug 10 → Aug 11 was one-shot, not ongoing.
3. dashboard/arxiv-report-2026-08-01.html · 56,574 B (56 KB) · 2026-08-02 13:47 · delete — 19-day-old "weekly arXiv report" snapshot (was 15 days old on Aug 17, 16 days on Aug 18, 17 days on Aug 19, 18 days on Aug 20), zero references in any current cycle doc, the setup pipeline, or README.md. arxiv.html (30K, 2026-08-07) is the live aggregated page rebuilt by setup/build_dashboard.py; this 19-day-old HTML is a stale render that the build script no longer produces. Pure orphan, safe to delete. Day-10 confirmation: still orphan, mtime unchanged at Aug 2 13:47, size stable at 56,574 B.
4. (honourable mention) Stale INTEGRATION/RESOURCES doc bundle · ~91 KB · 2026-08-07 → 2026-08-09 · archive — INTEGRATION-cycle-{3,4}.md + RESOURCES-cycle-{3,4}.md + the *-delta-cycle-2c.md pair are all superseded by cycle 5/6 (latest mtime Aug 10). Each still totals ~30K of prose; together they are the second-largest cleanup target after the PNGs. Note: the *-audio.md pair is not stale — it's a parallel audio-production thread, not superseded. Day-10 confirmation: no change vs. Aug 17. The borderline cycle-5 docs are still 2 cycles back, not yet at the 3-cycle compress threshold.

What is "mtm-resource-sweep" doing in the daily-pipeline UI

Unchanged from Aug 10/11/12/13/14/15/16/17/18/19. The 08-20 daily file's Tab 04 (Tool Integration) and Tab 08 (Cost & Routing) both reference the mtm pipeline indirectly through the Stripe 5,200-tool week-7 / Bedrock AgentCore GA+ week-5 / 4-deployment cohort 6-layer router toolchain. The hand-off anchor drop into mtm-resource-sweep is still rendered from app.py:605 for any panel whose prompt block contains that literal; the 08-20 file's Tab 04 prompt is the 5,200-tool factory + 5-layer capability-token + 6-tier deprecation_window, not the mtm-sweep hand-off, so the label is dormant again today.

• app.py line 605: the literal string drop into mtm-resource-sweep lives inside a _section_h("Prompt — " + p.tab_label) block; it is the .h text of a <div class="card kind-prompt"> that renders the parsed prompt from each panel inside the Inspector tab.
• verify.mjs line 57: the literal-vs-shape match is text.includes('mtm-resource-sweep') || text.includes('mtm-sweep'); both anchors are still resolvable from the live page text even when no panel's prompt block contains the literal.
• Source-of-truth HTML corpus: the long-form dailies in daily-pipeline-space/data/ (2026-08-07, 08-08, 08-09 × N, 08-10, 08-11, 08-12, 08-13, 08-14, 08-15, 08-16, 08-17, 08-18, 08-19, 08-20 now) all contain some prompt content. The 08-20 file's Tab 04 prompt is the 5,200-tool factory + 5-layer capability-token + 6-tier deprecation_window, not the mtm-sweep hand-off; the 08-09 gradio file is still the only one that contains the literal.
• Semantic meaning: unchanged — the daily-pipeline is a prompt factory; mtm-resource-sweep is the target tool the prompts get handed to. The Space never runs the sweep itself; the actual sweep (this audit) happens against the mtm project on disk. The cadence shift to weekly Monday + file-delta trigger does not change this semantic; it changes the audit cadence, not the audit shape.

Stale doc inventory (re-verified, mtime drift = 0 since Aug 14)