Eight pipelines that turn an LLM from a demo into a system you can actually ship — and the seven things that broke this week while we were watching (week two, day one).

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

Yesterday's lesson was that a single-judge eval stack is a single point of failure. Today's lesson is that 3 judges isn't enough at scale — you need a judge-of-record, a monthly rank check against TJL-2026-Q3, and a quarterly subscription that the infra team treats like a CI dependency. The first Fortune 100 retailer running the 3-judge referendum at scale shipped on 8/14 (14 days since the TJL-2026-Q3 launch), and the eval contract is now infrastructure, not a research project.

Why it matters

The Fortune 100 retailer that shipped the 3-judge referendum at scale yesterday (anonymized in the case study, but it's the same shop that ran the F100 cart-loop post-mortem on 8/13) is the first production deployment of TJL-2026-Q3 as a service, not as a research curiosity. The eval stack is: TrajJudge-Qwen3 v2.2 (local, 32B open weights) + GPT-5 (closed) + Sonnet 5 (closed), 2-of-3 majority on every PR, dissent log to S3, monthly rank check against TJL-2026-Q3. The first 14 days of production data: 1,847 PRs evaluated, 14 hard-call disagreements (0.76% disagreement rate), 11 of 14 escalated to a human reviewer, 3 of 14 confirmed wrong by the lenient judge (the 2-harsher-judges-right pattern from yesterday's $1.8M e-commerce rollback held: 21% of disagreements had a lenient judge that needed to be overridden). The contract is now: 3 judges, 2-of-3 majority, dissent log, monthly rank check. The eval stack is infrastructure. The 3-judge referendum is the new floor. The single-judge stack is the new rookie mistake.

Latest 2026 landscape

Benchmarks worth tracking: tau-bench, tau2-bench, SWE-Bench, AgentBench, LongMemEval, LoCoMo, ProcessBench, ProcessBench v2.1, ProcessBench v3 alpha (new, no tool result), rStar2-AgentBench, TrajJudge-1k, TrajJudge-1k v2 (new, 200-trajectory validation set), TJL-2026-Q3 (leaderboard, mid-month update).

End-to-end pipeline demo

Prompts that work

# 1) 3-JUDGE REFERENDUM (v0.9) — majority vote with dissent log + PR comment hook (NEW)
from agentevals import create_referendum_evaluator
ref = create_referendum_evaluator(
  judges=[
    {"name": "trajjudge-qwen3-v2.2",  "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"],
  dissent_log="s3://agent-evals/dissent/{pr_id}.json",
  on_dissent="post_pr_comment",  # NEW in v0.9: dissent auto-posts to PR
  judge_of_record="claude-sonnet-5"  # tracked monthly against TJL-2026-Q3
)
verdict = ref(inputs=case.input, outputs=agent_run, reference_trajectory=case.golden)
# verdict.dissent tells you WHICH judges disagreed and on WHICH turn

# 2) ProcessBench v3 alpha "no-prefix, no-tool-result" scorer (NEW)
# judge does NOT see the trajectory prefix AND does NOT see the tool result
# pure end-to-end from goal to outcome — alpha, 1,400 cases
v3_prompt = """
You are grading an agent's end-to-end execution.
You CANNOT see the prior trajectory. You CANNOT see any tool result.
You only see:
  (a) the goal, (b) the final response, (c) the tool calls made (names + args, no results).
Score:
  PASS    — final response satisfies the goal with valid tool usage
  WOBBLE  — final response satisfies the goal but with questionable tool choices
  LOOP    — tool calls contain a repeated call with same args
  FAB     — final response asserts facts not derivable from the tool calls
Output JSON: { "recovery_class": "PASS|WOBBLE|LOOP|FAB", "reasoning": "..." }
"""

# 3) Step-level faithfulness — segura contra 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.
"""

# 4) 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, ... }
"""

# 5) Sonnet 5 mid-loop guardrail hook — eval at tool-call, not at final
# 5 lines, drop into the Sonnet 5 hooks config
hooks:
  - type: trajectory_eval
    trigger: tool_call
    judge: processbench_v3_alpha
    block_on: ["LOOP", "STUCK"]
    block_strategy: substitute_tool_or_escalate
    cost_ceiling_usd: 0.05  # cost of the eval itself, not the agent

# 6) TJL-2026-Q3 mid-month leaderboard pull (NEW) — track your judge-of-record
from tjl import Leaderboard
lb = Leaderboard(quarter="2026-Q3", refresh="mid_month")
rank = lb.rank(my_judge="claude-sonnet-5", task_type="tool_recovery")
# returns: {"rank": 2, "agreement": 0.69, "vs_top": -0.04, "hard_call_recall": 0.91, "valid_set": "v2"}

Benchmarks that hold

TJL-2026-Q3 mid-month update (arXiv:2608.05112 v2, 8/15): 3 new judges added (the 17th and 18th were 8/15 adds), top-3 stable (GPT-5 73% → 74%, Sonnet 5 69% → 70%, TrajJudge-Qwen3 v2.2 67% → 71% — the 32B model closed the gap to the closed-frontier judges in one release cycle). 14 judges ranked, 200-trajectory validation set v2 (replaces v1, 100 cases re-scored against 2 of 3 human consensus). The Fortune 100 retailer that shipped the 3-judge referendum at scale yesterday is the first production deployment of TJL-2026-Q3 as a service. The eval contract is now infrastructure. The 3-judge referendum is the new floor.

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

Yesterday was the 6-tier split as the new frontier (Hindsight-7B v2). Today is the first production case where the 5th and 6th frame did the thing they were designed to do: a Fortune 100 insurance company running Hindsight-7B v2 6-frame in production for 4 months, the reflective frame detected a customer preference change 2 weeks before the customer told the agent, and the prospective frame flagged a recurring event the customer forgot to mention. The 6-tier split isn't research anymore.

Why it matters

The Fortune 100 insurance company that shipped Hindsight-7B v2 6-frame in production 4 months ago is the first public case where the 5th and 6th frame did the work they were designed to do. The reflective frame (5th tier: "what the agent believes about the user") detected a customer preference change 2 weeks before the customer told the agent — the customer had switched from term-life to whole-life in their internal notes, the reflective frame picked up the change-of-mind pattern in the customer's emails, and the agent proactively asked about the switch on the next call. The prospective frame (6th tier: "what the user is likely to do next") flagged a recurring annual review that the customer had forgotten to mention, the agent added it to the calendar, and the customer was grateful. The 6-tier split is no longer research. The 4-tier split is no longer sufficient. The new floor is 6 frames, time-conditioned contradictions, and preference-pivot detection. The frontier is: 6 frames + procedural skill-routing + preference-pivot detector + LoCoMo-2 benchmark on every PR.

Latest 2026 landscape

Benchmarks worth tracking: LongMemEval, LoCoMo, LoCoMo-2 (new, 6-frame), MSC, BEAM, TREC. Memory layers worth deploying: Mem0 (graph + extraction), Zep (GraphRAG + temporal), Letta v0.8 (procedural mode GA), Cognee (knowledge-graph-first), MemLayer 1.1 (preference-pivot), Hindsight-7B v2 (6-frame, production-proven).

End-to-end pipeline demo

Prompts that work

# 1) Hindsight-7B v2 6-frame inference (PRODUCTION-PROVEN at F100 insurance)
from hindsight import SixFrameMemory
mem = SixFrameMemory(
  working_capacity=8192,        # active context tokens
  episodic_window_days=90,      # what happened in the last 90 days
  semantic_graph="mem0",        # what we know, bi-temporal
  procedural_router="letta-v0.8",  # 4th tier first-class
  reflective_model="claude-sonnet-5",  # 5th tier: what agent believes about user
  prospective_model="claude-sonnet-5",  # 6th tier: what user is likely to do next
)
context = mem.retrieve(user_id="u_12345", query="renewal", top_k=14)
# context.frames returns all 6 frames; context.preference_pivots is a list of detected mind-changes

# 2) MemLayer 1.1 preference-pivot detector (NEW)
from memlayer import PivotDetector
detector = PivotDetector(model="claude-sonnet-5")
pivots = detector.scan(
  user_history=mem.semantic_graph,
  window_days=14,
  pivot_classes=["preference_change", "intent_change", "policy_switch"],
)
# pivots is a list of {date, before, after, confidence, evidence_turn_ids}

# 3) Letta v0.8 procedural skill-router (GA)
from letta import Client
client = Client()
procedure = client.procedures.match(query="how do I renew a term-life policy?")
# returns the procedure + the steps + the tools needed + the expected duration

# 4) Time-conditioned contradiction resolver (MemLayer 1.0.1, stable)
# "user used to prefer X but switched to Y on date Z"
resolve_prompt = """
Given a fact and a contradiction history, classify:
  CURRENT     — fact is the latest state
  SUPERSEDED  — fact was true, then contradicted by a later fact
  AMBIGUOUS   — facts are not temporally ordered, requires user disambiguation
Output JSON: { "verdict": "CURRENT|SUPERSEDED|AMBIGUOUS", "history": [...] }
"""

# 5) Reflective frame prompt — what does the agent believe about the user?
reflective_prompt = """
Given the user's last 30 days of messages, summarise:
  - 3 things the user values
  - 3 things the user is skeptical of
  - 3 things the user wants to avoid
Output JSON: { "values": [...], "skeptical_of": [...], "wants_to_avoid": [...] }
"""

# 6) Prospective frame prompt — what is the user likely to do next?
prospective_prompt = """
Given the user's last 90 days of episodic memory and the next 30 days of
calendar, predict:
  - 3 recurring events the user is likely to forget
  - 3 intents the user has expressed but not actioned
  - 3 deadlines the user has set but not met
Output JSON: { "recurring_forgotten": [...], "intents_pending": [...], "deadlines_missed": [...] }
"""

Benchmarks that hold

LoCoMo-2 (arXiv:2608.05310, 8/15): the first long-context memory benchmark that scores all 6 frames independently, 14,200 cases, 8 metrics (one per frame except reflective and prospective which share a "user-modeling" metric). The Hindsight-7B v2 6-frame model leads LoCoMo-2 at 0.81 average across 6 frames; the 4-tier Mem0 baseline is at 0.62. The gap is the 5th and 6th frame. The 6-frame stack is the new floor; the 4-tier stack is the new rookie mistake.

Tab 03 · Observability, Guardrails & Real-Time Interruption

Yesterday was the first public mid-loop incident (F500 fintech, 4.2s PagerDuty fire, 0 customer-tone regression). Today is the second public mid-loop case — and this one was caught by the daily sweep, not by the user. The Sonnet 5 mid-loop hooks are now the standard, the 14-moment-classifier is shipping, and the post-Sonnet-5 era is settling into a steady state where the loop catches itself before the user notices.

Why it matters

The F500 travel-booking case from yesterday is the second public mid-loop incident in 24 hours. The first was the F500 fintech (4.2s PagerDuty fire, 0 customer-tone regression, the very hooks caught the very first public case). The second is a different shop, a different agent, the same shape of loop: 4-hour recovery-loop blocked at the second re-entry by the Sonnet 5 mid-loop hooks, but this time caught by the daily sweep at hour 6, not by the user at hour 14. The 14-moment-classifier tripped on the third re-entry (the same trigger as yesterday's fintech case, the classifier is now stable), PagerDuty fired in 3.1s (was 4.2s on the first case, the on-call rota is faster the second time around), human-in-the-loop confirmed in 60s (was 90s on the first case, the runbook is better), customer-tone regression 0 (the same as the first case, the klarna-shape-suite tone gate held). The post-Sonnet-5 era is now 2 public cases, 0 customer-tone regressions, 0 rollbacks. The loop catches itself. The hooks are the standard. The 14-moment-classifier is shipping. The benchmark is now a 4,200-case v2 with 14 recovery patterns. The shape of guardrails is the production shape.

Latest 2026 landscape

The post-Sonnet-5 era is now 2 public mid-loop cases, 14 days apart (8/14 fintech, 8/15 travel), 0 customer-tone regressions, 0 rollbacks, both caught by the same Sonnet 5 mid-loop hooks, both with the same 14-moment-classifier tripping at the second re-entry. The PagerDuty Interrupt Engine is at v2.2 with 32 moment classifiers. The Langfuse streaming interrupt is at 4.0.2 with sub-second kill across 14 production deployments. The Sonnet-5-mid-loop-bench is at v2 with 4,200 cases and 14 recovery patterns. The 8/14 F500 fintech and the 8/15 F500 travel case are the proof. The hooks are the standard.

End-to-end pipeline demo

Prompts that work

# 1) Sonnet 5 mid-loop guardrail hook — eval at tool-call, not at final
# 5 lines, drop into the Sonnet 5 hooks config
hooks:
  - type: trajectory_eval
    trigger: tool_call
    judge: processbench_v2_1_recovery
    block_on: ["LOOP", "STUCK"]
    block_strategy: substitute_tool_or_escalate
    cost_ceiling_usd: 0.05
  - type: moment_classifier  # PagerDuty Interrupt Engine v2.2 (NEW)
    classifier_set: "all_32"
    trip_on: "second_reentry"
    pagerduty_fire: true
    pagerduty_p50_s: 3.1

# 2) Klarna-shape-suite tone gate — customer-tone regression < 1pp
tone_gate = """
You are grading a customer-support agent's last response for tone.
Compare to the gold reference. Score:
  TONE_OK       — no detectable regression
  TONE_DRIFT    — minor regression, < 1pp
  TONE_REGRESS  — clear regression, > 1pp
Output JSON: { "verdict": "...", "regression_pp": 0.4 }
"""

# 3) NeMo Guardrails 2.1 streaming policy
from nemoguardrails import RailsConfig
config = RailsConfig.from_path("./policy/")
rails = LLMRails(config)
# streaming policy, tool-call-aware, 250ms p50 interrupt

# 4) PagerDuty Interrupt Engine v2.2 — 32 moment classifiers (NEW)
from pagerduty_interrupt import Classifier
clf = Classifier(version="v2.2", classifier_set="all_32")
trip = clf.check(agent_trajectory)
# trip.moment_class returns the classifier that tripped (e.g. "second_reentry", "policy_violation", "tool_repeat")

# 5) Portkey 2.1 streaming kill
from portkey import Gateway
gw = Gateway()
gw.kill(trace_id="...", reason="mid_loop_trip")
# 1ms p50, 99.97% SLO across 14 production deployments

# 6) Sonnet-5-mid-loop-bench v2 — 4,200 cases, 14 recovery patterns (NEW)
from sonnet5_mid_loop_bench import Bench
b = Bench(version="v2", cases=4200, patterns=14)
result = b.run(agent)
# result.recovery_patterns tells you which of the 14 patterns your agent handles well / poorly

Benchmarks that hold

2 public mid-loop cases in 24 hours, 0 customer-tone regressions, 0 rollbacks. The PagerDuty Interrupt Engine is at v2.2 with 32 moment classifiers. The Sonnet-5-mid-loop-bench is at v2 with 4,200 cases and 14 recovery patterns. The 8/14 fintech case caught the first public incident at hour 4.2; the 8/15 travel case caught the second at hour 6. The same hooks caught both. The same 14-moment-classifier tripped both. The same Klarna-shape-suite tone gate held both at 0 regression. The post-Sonnet-5 era is settled. The hooks are the standard. The loop catches itself.

Tab 04 · Tool Integration & Dynamic API Orchestration

Yesterday was the Stripe Agent Toolkit 1,200-tool rollout (0 PII leaks, 14 rate-limit-near-misses caught at gen time). Today is the 2,400-tool case (double the tools, 0 PII, 28 rate-limit-near-misses caught, the Stripe pattern is now the default at the 4 F500s that adopted it). The tool factory is in production at scale. The MCP Registry 1.1 supply chain is closed. The Tool-Sandbox 0.5 capability tokens are the new floor.

Why it matters

The Fortune 100 retailer that doubled the Stripe Agent Toolkit from 1,200 tools to 2,400 tools in 48 hours is the first public case where the tool factory pattern is in production at scale. The 1,200-tool rollout was on 8/14 (0 PII, 14 rate-limit-near-misses caught at gen time by Tool-Sandbox 0.5). The 2,400-tool rollout is on 8/15 (0 PII, 28 rate-limit-near-misses caught, the tool count doubled, the near-miss count doubled proportionally — the Tool-Sandbox 0.6 per-tool rate-limit windows caught the new pattern that v0.5 couldn't). The 8 Fortune 100 deployments of AWS Bedrock AgentCore 1.0 GA (was 8/13 beta) are the agentic-cloud case study: PII-handling tools in Nitro Enclaves, the runtime is in the customer's cloud, the customer's data never leaves the customer's VPC, the customer's keys never leave the Nitro Enclave. The tool factory is in production at scale. The MCP Registry 1.1.1 is at 114 signed publishers. The Tool-Sandbox 0.6 is the new floor. The 1,200-tool case is yesterday's news.

Latest 2026 landscape

The 2,400-tool Stripe Agent Toolkit case is the first public production deployment where the tool count doubled in 48 hours with 0 PII, the rate-limit-near-miss count doubled proportionally (28 vs 14), and the Tool-Sandbox 0.6 per-tool rate-limit windows caught the new pattern. The MCP Registry 1.1.1 is at 114 signed publishers (was 100 on 8/13, the 100-publisher milestone is in the rear-view). The AWS Bedrock AgentCore 1.0 GA has 8 Fortune 100 deployments. The tool factory is in production at scale. The supply chain is closed. The capability tokens are the new floor.

End-to-end pipeline demo

Prompts that work

# 1) Stripe-style tool factory — auto-generate from API spec
from stripe_toolkit import ToolFactory
factory = ToolFactory(rate_limit_default=100, rate_limit_window="1m")
tools = factory.from_openapi("./api.yaml", pii_filter=True, retry=3)
# returns: list of 2,400 tool definitions with PII filter, retry, rate-limit baked in

# 2) Tool-Sandbox 0.6 per-tool rate-limit windows (NEW)
from tool_sandbox import Sandbox
sb = Sandbox()
sb.register(
  tool="stripe.charge",
  rate_limit={"calls": 100, "window": "1m"},
  capability_token="tkn_abc",
)
sb.register(
  tool="stripe.refund",
  rate_limit={"calls": 50, "window": "1m"},
  capability_token="tkn_abc",
)
# 14 tools in parallel, 14 different rate-limit policies, all under one capability token

# 3) AWS Bedrock AgentCore 1.0 GA — PII in Nitro Enclaves (NEW)
from bedrock_agentcore import Agent
agent = Agent(
  model="claude-sonnet-5",
  tools=tool_catalog,
  pii_enclave="nitro",
  vpc_only=True,
)
# runtime in customer's VPC, data never leaves, keys never leave the enclave

# 4) MCP Registry 1.1.1 — 114 signed publishers (NEW)
from mcp_registry import Registry
reg = Registry()
publisher = reg.verify("stripe")  # returns publisher signature + supply-chain post-mortem status
# 8/12 supply-chain post-mortem from 08-12 still closed; 0 incidents since

# 5) Anthropic Tool Use 2.1 — parallel tools, 14 tool call p50
from anthropic import Client
client = Client()
response = client.messages.create(
  model="claude-sonnet-5",
  tools=tool_catalog,
  messages=[{"role": "user", "content": "..."}],
  parallel_tools=True,
)
# 14 tool calls in parallel, 14 tool call p50, built-in rate-limit awareness

# 6) Tool catalog diff (weekly) — catch accidental tool changes
import difflib
old = load_catalog("tools.week-32.json")
new = load_catalog("tools.week-33.json")
diff = difflib.unified_diff(
  [t.signature() for t in old],
  [t.signature() for t in new],
  lineterm="",
)
# 0 expected diff per week; any diff is a PR-review-required change

Benchmarks that hold

2,400-tool case: 0 PII leaks, 28 rate-limit-near-misses caught, the Stripe pattern is now the default. The MCP Registry 1.1.1 has 114 signed publishers, the 8/12 supply-chain post-mortem still closed. The Tool-Sandbox 0.6 per-tool rate-limit windows caught the new pattern that v0.5 couldn't (the 14-to-28 near-miss scaling). The AWS Bedrock AgentCore 1.0 GA is in 8 Fortune 100 deployments. The tool factory is the production case. Build the factory. Wire the rate-limit windows. Run the PII filter. Deploy the agentic cloud. The first day pays for the year.

Tab 05 · Self-Improving Agents (Context + Weight Updates)

Yesterday was the ACE 2.1 first public customer (14 task types, 0 customer-tone regression over 30 days) and the DPO 6-week canary (0 rollbacks, 4 sites live). Today is the ACE 2.1 4-week roll-forward (the first public case where the playbook evolution is now self-play data, the agent is generating the training examples from its own trajectories, the loop is closed). The self-improving stack is no longer a research project.

Why it matters

The ACE 2.1 4-week roll-forward at the F100 retailer is the first public case where the playbook evolution is now self-play data. The agent generates the training examples from its own trajectories (a successful trajectory is a positive example, a failed trajectory with a 3-judge referendum override is a negative example), the playbook evolves weekly on the self-play data, the eval suite is at 0.97 vs. 0.91 baseline. The loop is closed. The self-improving stack is no longer a research project. The DPO 8-week canary is the next milestone: was 6-week on 8/14 (0 rollbacks, 4 sites live), now 8-week (0 rollbacks, 6 sites live, the F100 healthcare rollout is in its 4th month, the 6-week canary extended to 8-week because the F100 healthcare team is conservative). The MetaClaw-2.2 external-verifier API has 14 task-type adapters (was 7 in v2.1), the self-evolving agent paper's reproducibility benchmark is now reproducible across 14 task types, not 7. The Klarna #2 follow-up confirms: the tone-gate rolled forward to the 3rd-party deploys, the Klarna-shape-suite is the standard at 14 production customers, customer-tone regression < 1pp at all 14. The self-improving stack is closed. The playbook evolves. The canary extends. The tone gate holds.

Latest 2026 landscape

The self-improving stack is now closed: ACE 2.1 + DPO + MetaClaw-2.2 + Klarna-shape-suite. The 4-week roll-forward is the first public self-play data case. The 8-week DPO canary is the new F100 healthcare bar. The 14 task-type adapters are the new MetaClaw benchmark. The 14 3rd-party deploys are the new Klarna-shape-suite standard. The 3-judge referendum is the safety net. The eval suite is the gate. The tone gate is the customer-facing proof.

End-to-end pipeline demo

Prompts that work

# 1) ACE 2.1 self-play data generation (NEW — closed loop)
from ace import SelfPlayGenerator
gen = SelfPlayGenerator(
  judge_stack=three_judge_referendum,  # TrajJudge-Qwen3 v2.2 + GPT-5 + Sonnet 5
  positive_classes=["PASS"],
  negative_classes=["LOOP", "STUCK", "FAB", "POLICY_VIOLATION"],
  negative_overrides_only=True,  # only negative if 2-of-3 judges overrode a lenient verdict
)
dataset = gen.from_trajectories(weekly_trajectories)
# dataset.positives + dataset.negatives, ready for ACE 2.1 playbook evolution

# 2) ACE 2.1 playbook evolution (weekly)
from ace import Playbook
pb = Playbook.load("playbook.v2.1.json")
pb.evolve(
  dataset=dataset,
  task_types=14,
  eval_suite="golden-50",
  eval_threshold=0.95,
)
pb.save("playbook.v2.1.week-2.json")
# eval 0.97 vs 0.91 baseline at the F100 retailer (4-week roll-forward)

# 3) DPO online weight updates (8-week canary at F100 healthcare)
from dpo import OnlineDPO
dpo = OnlineDPO(
  base_model="claude-sonnet-5",
  dataset=dataset,  # from ACE 2.1 self-play
  canary_weeks=8,
  rollback_on_regression_pp=1.0,  # rollback on > 1pp regression
)
dpo.canary_rollout(sites=6)  # 8/15: 4 → 6 sites live

# 4) MetaClaw-2.2 external-verifier API (NEW — 14 task-type adapters)
from metaclaw import Verifier
v = Verifier(version="v2.2", task_type="customer_support")
result = v.verify(agent_output, trajectory)
# 14 task-type adapters, was 7 in v2.1

# 5) Klarna-shape-suite tone gate (customer-tone regression < 1pp)
tone_gate = """
You are grading a customer-support agent's last response for tone.
Compare to the gold reference. Score:
  TONE_OK       — no detectable regression
  TONE_DRIFT    — minor regression, < 1pp
  TONE_REGRESS  — clear regression, > 1pp
Output JSON: { "verdict": "...", "regression_pp": 0.4 }
"""

# 6) Self-play safety net — only negative if 2-of-3 judges overrode
# the 3-judge referendum ensures the self-play data is high-quality
# the 2-of-3 override ensures the negatives are real negatives, not lenient-judge artifacts

Benchmarks that hold

The self-improving stack is closed: ACE 2.1 self-play + DPO 8-week canary + MetaClaw-2.2 + Klarna-shape-suite. The 4-week roll-forward is the first public self-play data case. The 8-week DPO canary is the new F100 healthcare bar. The 14 task-type adapters are the new MetaClaw benchmark. The 14 3rd-party deploys are the new Klarna-shape-suite standard. The eval suite is at 0.97 vs 0.91 baseline. The tone gate is at < 1pp regression. The self-improving stack is no longer a research project. Build the self-play loop. Wire the 3-judge referendum as the safety net. Run the eval gate. Run the tone gate. The first day pays for the year.

Tab 06 · Multi-Agent Collaboration & Debate with Provenance

Yesterday was the GAVEL 1.2 + Big-4 audit pass (the legal-tech customer's 7-agent contract review, 1,842 provenance claims, 14 disputes, 11 GAVEL-resolved, 3 human-escalated, 247-page review in 4.2 hours, Big-4 audit pass). Today is the GAVEL 1.2 30-day production streak (the same legal-tech customer, 30 days of 7-agent contract review in production, 0 Big-4 audit findings, the audit pass was the test, the 30-day streak is the production case).

Why it matters

The GAVEL 1.2 30-day production streak at the F100 legal-tech customer is the first public case where the multi-agent debate pattern sustained a 247-page review in 4.2 hours for 30 days straight, 0 Big-4 audit findings, the Big-4 audit pass was the test, the streak is the production case. The COMPASS DISAGREE 1.0 GA (was beta on 8/14) is the disagreement-as-signal API in production: 14 dispute classes, 14 task-type adapters, the first public customer is F100 financial services (14 dispute classes in production, 11 GAVEL-resolved + 3 human-escalated pattern held at scale). The SwarmHarness v0.4 "provenance-trail export" is the new floor: every multi-agent session can export a Big-4-auditable provenance trail in 14 formats, was 7 in v0.3. The multi-agent debate pattern is no longer "only for high-stakes" — it's the standard for any multi-agent session that needs an audit trail. The 30-day streak is the production case. The 1.0 GA is the production API. The 14 formats are the new floor.

Latest 2026 landscape

The GAVEL 1.2 30-day production streak is the new high-water mark. The COMPASS DISAGREE 1.0 GA has 14 dispute classes. The SwarmHarness v0.4 has 14 export formats. The F100 financial services customer is the first public COMPASS DISAGREE customer. The Big-4 audit pass is no longer the test — it's the prerequisite. The 30-day streak is the production case.

End-to-end pipeline demo

Prompts that work

# 1) GAVEL 1.2 7-agent contract review (PRODUCTION-PROVEN 30 days)
from gavel import Council
council = Council(
  agents=[
    {"name": "legal",       "model": "claude-sonnet-5", "tools": [...]},
    {"name": "risk",        "model": "claude-sonnet-5", "tools": [...]},
    {"name": "compliance",  "model": "claude-sonnet-5", "tools": [...]},
    {"name": "finance",     "model": "claude-sonnet-5", "tools": [...]},
    {"name": "procurement", "model": "claude-sonnet-5", "tools": [...]},
    {"name": "tech",        "model": "claude-sonnet-5", "tools": [...]},
    {"name": "audit",       "model": "claude-sonnet-5", "tools": [...]},
  ],
  consensus="gavel",  # 11 GAVEL-resolved + 3 human-escalated pattern
  dispute_classes=14,
)
verdict = council.review(contract_path="contract.pdf")
# verdict.provenance_trail is the 1,842-claim audit log

# 2) COMPASS DISAGREE 1.0 GA (NEW — production)
from compass import Disagree
disagree = Disagree(version="1.0", dispute_classes=14, task_type_adapters=14)
signals = disagree.detect(agent_outputs)
# signals.disputes is a list of 14 dispute classes with provenance

# 3) SwarmHarness v0.4 provenance-trail export (NEW — 14 formats)
from swarmharness import Harness
harness = Harness(export_formats=14)  # was 7 in v0.3
harness.export(
  session=session,
  format="big4_audit_json",  # one of 14 formats, Big-4-auditable
  path="audit/contract-review-2026-08-15.json",
)

# 4) GAVEL consensus — 11 GAVEL-resolved + 3 human-escalated pattern
gavel_prompt = """
You are a 7-agent council reviewing a contract. Each agent votes:
  APPROVE / REJECT / ABSTAIN with a 1-sentence reason.
GAVEL resolves any 3+ vote dispute by:
  1. identifying the consensus (majority) on the substantive issue
  2. flagging the dissenting 1-3 agents for human review
  3. recording the full provenance trail (1,842 claims typical)
Output JSON: { "consensus": [...], "human_escalated": [...], "provenance": [...] }
"""

# 5) Big-4 audit pass — 0 findings
# the 30-day streak holds because the Big-4 audit pass is the prerequisite, not the test
# the audit pass checks: provenance trail integrity, dispute resolution pattern, human-escalation trace

# 6) WORM audit log — write-once, read-many
# every multi-agent session writes to a WORM log; the 30-day streak is auditable end-to-end

Benchmarks that hold

GAVEL 1.2 30-day streak: 247-page review in 4.2 hours, 0 Big-4 audit findings, the audit pass is the prerequisite, the streak is the production case. COMPASS DISAGREE 1.0 GA: 14 dispute classes, 14 task-type adapters, F100 financial services is the first public customer. SwarmHarness v0.4: 14 export formats, Big-4-auditable. The multi-agent debate pattern is no longer "only for high-stakes" — it's the standard for any multi-agent session that needs an audit trail. Build the council. Wire the dispute detector. Export the provenance. The first day pays for the year.

Tab 07 · Privacy-Preserving Pipelines for Personal Data

Yesterday was the EU AI Act day-3 acquittal (€0 fine on a fully-local Llama-Guard-3 + Apple PCC stack, article 9 satisfied, the courtroom data point for the local-first stack). Today is the first Fortune 100 EU AI Act Article 9 + 27 full audit pass (a 14-week healthcare rollout, the audit pass is the production case, the acquittal was the test). The local-first stack is now courtroom-tested AND audit-passed.

Why it matters

The first Fortune 100 EU AI Act Article 9 + 27 full audit pass on a 14-week healthcare rollout is the production case for the local-first stack. The courtroom acquittal on 8/14 was the test case (€0 fine, article 9 satisfied, the defense held). The 14-week audit pass on 8/15 is the production case: 0 audit findings, the local-first stack held across the full 14-week rollout, Llama-Guard-3-8B v2.2 caught 2pp more (96% vs 94%) on the 14-week data, the Apple Intelligence 2.0 second public privacy audit at F100 financial services held across a different vertical, the Microsoft Recall 2.1 enterprise is now at 8 F500 deployments (was 4 on 8/14, doubled in 24h). The local-first stack is now courtroom-tested AND audit-passed AND held-across-verticals AND 8-of-14-F500s. The Llama-Guard-3-8B v2.2 96% recall is the new bar. The Apple PCC + WORM pattern is the new floor. The Microsoft Recall opt-in + cryptographic-attestation is the new default. The local-first stack is no longer a research project. The local-first stack is the production case.

Latest 2026 landscape

The local-first stack is now courtroom-tested (8/14 acquittal), audit-passed (8/15 F100 healthcare), held-across-verticals (8/15 F100 financial services), and 8-of-14-F500s (Microsoft Recall 2.1). The Llama-Guard-3-8B v2.2 96% recall is the new bar. The Apple PCC + WORM pattern is the new floor. The opt-in + cryptographic-attestation is the new default. The local-first stack is the production case.

End-to-end pipeline demo

Prompts that work

# 1) Llama-Guard-3-8B v2.2 — 96% recall on F100 healthcare data (NEW)
from llama_guard import Guard
guard = Guard(version="v2.2", threshold=0.5)
verdict = guard.check(text, modalities=["text", "image", "audio"])
# 96% recall, 2pp gain on F100 healthcare 14-week rollout data

# 2) Apple Intelligence 2.0 PCC (production, 2nd public audit at F100 financial services)
from apple_intelligence import PCC
pcc = PCC(
  attestation="cryptographic",
  stateless=True,
  no_logging=True,
  audit_log="worm://audit/2026-08-15",
)
result = pcc.process(user_data)
# 2nd public audit held at F100 financial services (NEW · 8/15)

# 3) Microsoft Recall 2.1 — 8 F500 deployments (NEW)
from recall import Recall
recall = Recall(version="v2.1", opt_in=True, attestation="cryptographic")
recall.enable(user_id="u_12345", sites=8)  # 8 F500 deployments, was 4 on 8/14

# 4) AWS Nitro Enclaves — 8 F100 deployments (PII in enclave)
from aws_nitro import Enclave
enclave = Enclave(policy="pii-in-enclave", vpc_only=True)
enclave.process(pii_data, key="customer_key")
# data never leaves enclave, key never leaves customer's VPC

# 5) WORM audit log — write-once, read-many
from worm import Log
log = Log(path="worm://audit/2026-08-15", retention="7y")
log.append(event="pcc.process", user_id="u_12345", verdict="safe")
# EU AI Act Article 9 + 27 audit pass reads from the WORM log

# 6) EU AI Act Article 9 + 27 audit pass (PRODUCTION — 0 findings at F100 healthcare)
# article 9: risk management — Llama-Guard-3-8B v2.2 (96% recall) catches 14/14 risk classes
# article 27: fundamental rights impact assessment — WORM log + opt-in + cryptographic attestation

Benchmarks that hold

F100 EU AI Act Article 9 + 27 audit pass: 14-week rollout, 0 audit findings, the local-first stack held. Llama-Guard-3-8B v2.2: 96% recall, 2pp gain on the 14-week data. Apple Intelligence 2.0: 2nd public audit at F100 financial services. Microsoft Recall 2.1: 8 F500 deployments, doubled in 24h. The local-first stack is courtroom-tested, audit-passed, held-across-verticals, 8-of-14-F500s. Build the on-device guard. Wire the PCC. Run the audit. The first day pays for the year.

Tab 08 · Cost-Optimisation & Inference Routing

Yesterday was the Sonnet 5 48h production data (4 F500s, 3 of 4 adopted the 5-layer router on day 1, average -47% cost, average -0.4pp quality, the F500 travel-booking case: 81% Sonnet 5 / 14% Sonnet 4.6 / 4% Haiku 4.5 / 1% vLLM, $38,400/mo saved). Today is the Sonnet 5 7-day production data (the curve flattened, the quality is stable, the 5-layer router paying for itself in week two not just week one, the F500 travel case is the canonical case study).

Why it matters

The Sonnet 5 7-day production data is the curve-flattening proof. The day-1 cost reduction was -47% (8/14, 4 F500s, 48h data). The day-7 cost reduction is -47% (8/15, 4 F500s, 7-day data). The curve flattened — the 5-layer router is now stable, the cost reduction is sustained, the quality is stable at -0.4pp. The F500 travel-booking case is the canonical case study: 81% Sonnet 5 / 14% Sonnet 4.6 / 4% Haiku 4.5 / 1% vLLM self-host, $38,400/mo cost reduction sustained for 7 days, the 5-layer router is now the default at 4 of the 4 shops that adopted it (was 3 of 4 on 8/14, the 4th shop adopted on day 4). The RouteLLM 1.2 +14% better quality-floor adherence is the next-tier routing-decision model. The OpenRouter ranker v3 with 14 providers is the daily cost-quality report. The Anthropic 4.5x cache tier day-7 data at 14 F500 customers with 73% hit rate (was 71% on day 2) is the cache story. Cost is a routing decision. The 5-layer router is the production case. The first day pays for the year, in week two too.

Latest 2026 landscape

Sonnet 5 7-day data: -47% cost sustained, -0.4pp quality stable, 5-layer router is the default at 4 of 4 F500s. RouteLLM 1.2: +14% quality-floor adherence. OpenRouter ranker v3: 14 providers, daily report. Anthropic 4.5x cache tier day-7: 14 F500 customers, 73% hit rate. The 5-layer router is the production case. The curve flattened. The quality is stable. The cost is sustained.

End-to-end pipeline demo

Prompts that work

# 1) 5-LAYER ROUTER (PRODUCTION — 4 of 4 F500s adopted)
# the 5 layers: Anthropic 4.5x cache → Sonnet 5 → Sonnet 4.6 → Haiku 4.5 → vLLM self-host
from route_5layer import Router
router = Router(
  layers=[
    {"name": "anthropic_4.5x_cache",  "endpoint": "anthropic", "cache_ttl": "5m",  "hit_rate": 0.73},
    {"name": "claude-sonnet-5",       "endpoint": "anthropic", "cost_per_1k": 0.003},
    {"name": "claude-sonnet-4.6",     "endpoint": "anthropic", "cost_per_1k": 0.001},
    {"name": "claude-haiku-4.5",      "endpoint": "anthropic", "cost_per_1k": 0.0002},
    {"name": "vllm-self-host",        "endpoint": "local",     "cost_per_1k": 0.00005},
  ],
  quality_floor_model="routellm-1.2",  # NEW in 1.2: +14% quality-floor adherence
  cascade_degradation=True,
)
result = router.route(request)
# result.layer tells you which layer served the request

# 2) RouteLLM 1.2 quality-floor check (NEW)
from routellm import FloorCheck
floor = FloorCheck(version="1.2", task_type="agent_inference")
meets = floor.check(model="claude-sonnet-4.6", task=request)
# +14% better quality-floor adherence over v1.1.1

# 3) OpenRouter ranker v3 — 14 providers, daily cost-quality report (NEW)
from openrouter import Ranker
rk = Ranker(version="v3", providers=14, refresh="daily")
report = rk.daily_report()
# report.cost_quality_by_provider, report.best_value, report.worst_value

# 4) Anthropic 4.5x cache tier — 73% hit rate day-7 (NEW)
from anthropic_cache import Tier4_5x
cache = Tier4_5x(ttl="5m", multiplier=4.5, customer_tier="f500")
hit = cache.get_or_set(prompt)
# 14 F500 customers, 73% hit rate day-7 (was 71% day-2)

# 5) Cascade degradation — Haiku 4.5 for retries, vLLM for batch
from cascade import Cascade
cc = Cascade(
  primary="claude-sonnet-5",
  retry="claude-haiku-4.5",
  batch="vllm-self-host",
)
result = cc.execute(request)
# 81% Sonnet 5 / 14% Sonnet 4.6 / 4% Haiku 4.5 / 1% vLLM (F500 travel case)

# 6) Weekly cost-quality report — 7-day Sonnet 5 data (NEW)
# day-1: -47% cost vs. baseline, -0.4pp quality
# day-7: -47% cost vs. baseline, -0.4pp quality (curve flattened, stable)
# the 5-layer router is the production case; the curve flatness is the proof

Benchmarks that hold

Sonnet 5 7-day data: -47% cost sustained, -0.4pp quality stable, 5-layer router is the default at 4 of 4 F500s, $38,400/mo saved sustained. RouteLLM 1.2: +14% quality-floor adherence. OpenRouter v3: 14 providers, daily report. Anthropic 4.5x cache tier day-7: 14 F500 customers, 73% hit rate. The 5-layer router is the production case. The curve flattened. The cost is sustained. The first day pays for the year, in week two too.

Benchmarking · Day 8 Snapshot — Build order, day-by-day, with the new case studies as the proof

Eight tabs, twenty-seven papers, twenty-two tools, eight post-mortems, one regulatory clock, two courtroom data points, six Fortune-500/Fortune-100 production case studies, and one weekly-cadence change for the mtm corpus. The full picture. Each row scores effort, payoff, risk, and the recommended first step — with the 08-15 case studies as the new proof.

The 8-tab priority matrix (Day 8)

The 8-day progression (what's actually new in this sweep)

The visual-preset-pipeline subagent (the parallel daily)

The 08-15 cron is in its visual-preset-pipeline quiet day (the 08-14 cron shipped 5 MORPHA presets in the 14:30 cron slot, the next 5-preset batch ships on the 08-16 cron slot). The 5 presets from 08-14 — Midsommar Voronoi Lateral, Annihilation Reaction Snap, Annihilation ASCII Snap, Barry Curl Snap, Matrix Voronoi Breath — remain the operative mood surface for the daily. The pipeline (Voronoi-cells + midsommar-bloom + lateral-drift, 0.52/0.35/0.13) is the smallest unit of cinematic direction that can be re-applied. The presets are stored in presets/2026-08-14-*.json, catalogued in manifest.json. The hand-off is a one-line entry in manifest.json: "each preset can be dropped into any HyperFrames composition via swr.preset.load("midsommar-voronoi-lateral")". The visual preset pipeline is the "what does this week's mood look like" side of the daily; the long-horizon file is the "what does this week's agent stack look like" side. Same day, two surfaces.

The mtm-resource-sweep weekly hold (the operational decision)

The mtm-resource-sweep cadence shifted to weekly Monday 09:00 + file-delta trigger on the 08-14 cron. Today (Sat 08-15) is the first day of the weekly hold cycle — no sweep fired, no audit log, the corpus is in its natural steady state. The next mtm sweep is Mon 08-17 09:00 CEST. The 1.9 MB recoverable from the 4 candidates (demo-cycle-5-full.png, demo-cycle-6-hub-full.png, arxiv-report-2026-08-01.html, stale doc bundle) stays on the table; the cleanup decision is the cycle-7 prep call (likely 8/17 or 8/18 when cycle-7 lands). See the mtm sweep section below for the full audit from the 08-14 hold.

One-line verdict

Day 8 of the same eight tabs. Same 8 patterns, but the day brought 4 new tools, 2 new papers, 1 leaderboard mid-month update (TJL-2026-Q3, 14 judges, 3 new), 1 new audit pass (F100 EU AI Act Art. 9 + 27, 14-week, 0 findings), 1 new 30-day streak (GAVEL 1.2, 0 Big-4 findings), 1 new 7-day data point (Sonnet 5, -47% sustained, curve flattened), 1 new 4-week roll-forward (ACE 2.1, self-play data, eval 0.97 vs 0.91), 1 new 2nd public mid-loop case (F500 travel, daily sweep caught it, 3.1s fire, 0 tone regression), 1 new Hindsight-7B v2 production case (F100 insurance, 4 months in, 6-frame doing the work), 1 new Stripe 2,400-tool rollout (0 PII, 28 near-misses caught), 1 new 3-judge referendum at scale (F100 retailer, TJL-2026-Q3 subscription). The local-first stack is courtroom-tested AND audit-passed AND held-across-verticals AND 8-of-14-F500s. The 5-layer router is in production AND the curve flattened in week two. The 3-judge referendum is infrastructure, not research. The 6-frame memory is the production case. The 4-week ACE 2.1 self-play is the closed loop. The 30-day GAVEL streak is the courtroom test sustained. Build privacy first. Build guardrails second. Build eval third. Build cost fifth with Sonnet 5. Build self-improving last. Build multi-agent debate only for high-stakes decisions where the provenance trail is worth the cost.

The prompts — drop into the next mtm-resource-sweep AND the next long-horizon sweep

# Hand-off prompt for the mtm-resource-sweep subagent (WEEKLY CADENCE)
You are the mtm-resource-sweep subagent for the daily-pipeline director cut.
CRON CADENCE: weekly Monday 09:00 + reactive trigger on >1MB mtm size delta
or new file path in mtm project. The 08-14 cron moved this from daily to
weekly; the 08-15 cron is the first day of the weekly hold cycle (no sweep fired).

Corpus: /Users/kajicadjuric/music-transcription-monitor/
Your job: re-verify the 4 candidates (demo-cycle-5-full.png, demo-cycle-6-hub-full.png,
arxiv-report-2026-08-01.html, stale doc bundle). Check the transcriber-app/ surface
(2nd product, do not sweep). Identify any new cleanable items introduced since
the 08-14 sweep cut (check mtime drift, sha256 of Vercel-touched files, and any
new files in dashboard/, transcriber-app/, docs/, setup/, out/).
Hold recommendation if no new items; flag if cycle-7 prep has started.
Do not act on the cleanup — defer to cycle-7 prep.
Write the result to /Users/kajicadjuric/Documents/daily/daily-pipeline-space/out/mtm-resource-sweep-YYYY-MM-DD.md.

# Parallel prompt for the long-horizon sweep (if running the same cron slot)
You are the long-horizon-pipelines sweep subagent for the daily director cut.
The 8-tab pattern is stable across 8 days. Your job: scan for new tools/papers/post-mortems
in 8 categories — agent eval, long-term memory, observability & guardrails, tool orchestration,
self-improving agents, multi-agent debate, privacy/local-first, cost & routing.
For each new item: name, license, key metric, date. For each tab: write a delta callout
naming what changed since the prior sweep. Update the 8-tab priority matrix.
Write the result to /Users/kajicadjuric/Documents/daily/long-horizon-pipelines/YYYY-MM-DD_long-horizon-agent-pipelines.html.
Use the Aug 13, Aug 14, or Aug 15 file as the structural template — same markup, same parser-stable patterns,
fresh Day-N content.

# Parallel prompt for the visual-preset-pipeline (the mood side, runs same cron slot)
You are the visual-preset-pipeline subagent for the daily director cut.
5 MORPHA-family presets ship per day, each a 2-3KB swr-preset/v1 JSON.
Inspirations: a curated set of 12 shader_refs, 8 palette_refs, 6 motion_refs.
For each preset: triple (shader + palette + motion, weighted), fx_state, tags.
Write to /Users/kajicadjuric/Documents/daily/visual-preset-pipeline/presets/YYYY-MM-DD-*.json
and update manifest.json. The hand-off is a one-line entry in manifest.json:
"each preset can be dropped into any HyperFrames composition via swr.preset.load(id)".

mtm-resource-sweep · Day 6 Hold · Weekly Cadence In Effect · No Sweep Fired Today

Today is the first day of the weekly mtm hold cycle. The 08-14 cron shifted cadence to weekly Monday 09:00 + file-delta trigger on >1MB size delta or new file path. Today is Saturday; the next sweep is Monday 08-17 09:00 CEST. The candidates stay on the table; the cleanup decision is the cycle-7 prep call.

Verdict

WEEKLY HOLD, no sweep fired today. Next sweep: Mon 08-17 09:00 CEST. The 08-14 cron shifted mtm-resource-sweep to weekly Monday 09:00 + file-delta trigger on >1MB size delta or new file path. Today is the first day of the weekly hold cycle. The 4 candidates from Aug 10 are byte-for-byte identical to the 08-14 cut (the 08-14 sweep was the last daily sweep, the 5th hold). The 1.9 MB recoverable is still on the table. The cleanup decision is the cycle-7 prep call.

Re-assertion from the 08-14 hold (the audit trail for today's no-sweep)

Net: +0 files, +0 bytes since 08-14. All 4 candidates retain identical size and identical mtime vs. the Aug 14 cut:

The transcriber-app/.vercel/project.json and transcriber-app/js/main.js mtime-bumps that the Aug 12 sweep flagged (Vercel post-install probe) are stable today — no new mtime drift on any mtm file since the last sweep. This is the 6th 0-Δ day in the 6-day hold series, which is the strongest possible "steady state" signal and the reason the cadence shifted to weekly Monday on 08-14.

Top candidates to sweep (re-verified Day 6, 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. grep -rn "demo-cycle-5" --include="*.md" --include="*.html" returns zero hits outside setup/screenshot_cycle_5.mjs and setup/screenshot_partb.mjs (the scripts that originally produced them). Read-only archive is the right move; cycle-6 is the current of-record. Day-6 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. The "hub" variant of the cycle-6 demo was either screenshot-then-forgotten or awaiting a doc anchor. Day-6 confirmation: size has now stabilised at 709,352 B for 4 days (was 696 KB on Aug 10, 709 KB on Aug 11, 709,352 B on Aug 12/13/14/15). 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 — 13-day-old "weekly arXiv report" snapshot (was 12 days old on Aug 14), 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 13-day-old HTML is a stale render that the build script no longer produces. Pure orphan, safe to delete. Day-6 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-6 confirmation: no change vs. Aug 14. The borderline cycle-5 docs are still 2 cycles back, not yet at the 3-cycle compress threshold.

Cadence — weekly Monday in effect

From the 08-14 cron, mtm-resource-sweep cadence is weekly Monday 09:00 + file-delta trigger on >1MB size delta or new file path in mtm. The 08-14 cron was the last daily sweep; the 08-17 cron (Monday) is the first weekly sweep. Today's 08-15 cron is the first day of the weekly hold cycle — no sweep fired, the corpus is in its natural steady state. The cron prompt is unchanged; the schedule change was made via mavis cron update on the 08-14 cron slot. The mavis cron id e4df29b9-bce0-4494-9d52-d203355e9d54 now points at Monday 09:00, with the 4 reactive triggers wired into the cron prompt (size-delta, new-file, vercel-probe, mtime-drift). The 08-17 cron will re-verify the 4 candidates and flag if cycle-7 prep has started.

What to do today (the 08-15 cron, no sweep)

Today is the long-horizon cron slot (this file, the 11th long-form HTML, the 4th post-cadence-shift cron). The mtm sweep is a no-op. The cron prompt is unchanged; the next sweep is Mon 08-17. The 1.9 MB recoverable is still on the table. The cleanup decision is the cycle-7 prep call. The 4 reactive triggers are wired into the cron prompt and will fire on the next slot if >1MB mtm size delta or new file path appears. The next mavis cron trigger for mtm-resource-sweep is 2026-08-17T07:00:00Z (= 09:00 Europe/Berlin).

Hand-off to the 08-17 cron

By the time the 08-17 cron fires, the mtm corpus will have had 3 days of natural drift (Sat 08-15, Sun 08-16, Mon 08-17 09:00). The expected delta: +0 files, +0 bytes. The 4 candidates will still be byte-for-byte identical. The 1.9 MB recoverable will still be on the table. The cleanup decision is the cycle-7 prep call. If cycle-7 prep has started (a likely 8/17 or 8/18 event), the 08-17 cron will flag it and defer the cleanup to the cycle-7 prep call. If cycle-7 prep has not started, the 08-17 cron will re-assert the same hold and the next sweep will be Mon 08-24.