v0.1.1

• Docs only. Rephrased the stub suite-cost cell so a nudge count is plain text. No runtime change.
• Docs only. README "Start here" (three lines): what the cassette gate shows, make demo, and the existing English demo video. No runtime change.
• Docs only. README "Business value" (seven lines) and docs/BUSINESS_VALUE.md. Buyer, levers, baseline versus target, ROI [EST] from the existing business case, cost table, measurement plan, risks, and what is MEASURED versus not. No year-1 percent: build cost and business days per year are UNVERIFIED. No runtime change.
• Docs only. README "Proof". MEASURED 2026-10-22, Python 3.12.3, pytest 9.1.1, package 0.1.0, Python sources at main 7ea9d66: 110 passed on four local runs. Line coverage printed 88.05% (2711/3079) twice and 87.82% (2704/3079) twice. The seven statements are a lock race; the test passed each time. Pytest printed CoverageWarning: Module agent_runtime was previously imported, but not measured. Not a CI badge. No runtime change.
• Docs only. README TL;DR: three lines under the title and badges. No runtime change.
• Docs only. Local cassette paths in docs/live/ are a placeholder (<local-run-dir>/runN). Numbers, tokens, and costs in those logs are unchanged. No runtime change.
• Docs only. Dashboard demo of the three live re-measures (94 s, English) under docs/demo/dashboard_v2/, with EN/ES subtitles, the script, and the number trace. MEASURED 2026-10-18. Three runs per version, one model, one prompt. Not a clean improvement, and not a success rate. No runtime change.
• Docs only. Live re-measure on main 684d7b7 (2026-10-18,
  grok-4.20-0309-non-reasoning, 27 cases, three runs): 25/27, 25/27,
  26/27. Original 24 cases summed 67/72 (v1 62/72, v2 66/72). Happy-path
  misses 4, then 0, then 0. needs_review per run 0, 0, 0. Total cost
  US$0.164927. Not a clean improvement: 76/81 versus 75/81 case-runs on
  the 27-case set, and the failing cases moved. Open:
  edge_already_refunded 0/3 and edge_malformed_id 1/3. No runtime change.
• The completion nudge applies only to refund_completed with no confirmed
  send_refund and no human rejection of send_refund or send_email for
  that order. denied, needs_ops, and refund_rejected are not nudged.
  The nudge text no longer names a reason or an amount. A stated request
  amount, including zero, is a ceiling (above_requested). Stub suite
  after the shorter nudge text: 27/27, 71356 / 2453 tokens, 12117
  micro-USD, completion_nudges=2. Live effect: the 684d7b7 re-measure
  in the README (2026-10-18). Not a clean improvement.
• A finish that claims a refund without a matching executed send_refund
  is not accepted. The runner logs completion_nudge (run_id, reason)
  and hands the missing step back, up to MAX_COMPLETION_NUDGES (default
  2). Nudges count in the step budget and in the suite report. After the
  cap the run is still needs_review. The runtime does not create the
  write. send_refund must use the ticket reason (reason_mismatch).
  send_email is refused when that reason is not an allowed refund
  (email_not_warranted), including after lookup_policy says the reason
  is unknown. Three cassette cases cover the 2026-10-18 live misses
  (stop after lookup_policy, and a substituted reason plus an email).
  Golden expectations were not weakened. Stub suite on 2026-10-18:
  27/27, 71362 / 2453 tokens, 12118 micro-USD, completion_nudges=2.
  Demo: 3425 / 124 tokens, 584 micro-USD. Live effect: the 684d7b7
  re-measure in the README (2026-10-18). Not a clean improvement.
• Docs only. Live sample on main 3cbf6a7 (2026-10-18,
  grok-4.20-0309-non-reasoning, 24 cases): 21/24, 21/24, 20/24. Raw logs
  in docs/live/. This is not a clean improvement on the earlier 23-case
  sample (18/23, 18/23, 19/23, PR #1). No runtime change.
• Docs only. The hosted-sample notes live in docs/LIVE_SAMPLES.md.
  LICENSE copyright is Diego Lescano. No runtime change.
• The secret scanner flags xai-, sk- / sk-proj-, sk-ant-, an sk_
  prefix, ghp_ / github_pat_, AKIA, Bearer, and private-key headers.
  Samples in tests are built from pieces. docs/PROVENANCE.md records the
  sha256 of src/agent_runtime/secretscan.py. make check-secretscan
  recomputes it. Runtime behavior is unchanged.
• Optional hosted providers, off unless LIVE_MODE=1: an OpenAI-compatible
  client for xAI and OpenAI, and a thin Anthropic Messages adapter. Key,
  base URL, model, and prices come only from the environment. No vendor
  price is shipped. MAX_LIVE_SPEND_USD caps each run and the process.
• A claimed refund_completed with no matching send_refund effect fails
  the run as needs_review.
• The six adversarial cases check invariants instead of the stub's tool list.
  A 2026-10-17 live run mostly ignored the injection, so those guards were
  barely exercised live.
• Reasoning tokens and provider cost ticks are counted. Cached input is
  priced from LLM_CACHED_INPUT_USD_PER_MILLION or logged as an upper bound.
• make eval-live writes cassettes outside evals/cassettes and fails
  closed without live config. CI still runs the stub only.
• The system prompt asks the model to call the gated tools. Tool schemas
  include descriptions. edge_human_reject rejects at the gate.
• Stub suite on the 23-case set was 23/23, 38886 / 1902 tokens,
  6937 micro-USD. Demo stays 2408 / 124 tokens, 433 micro-USD.
• Live report percentiles are LLM calls only. Tool-step latency is separate.
• eval-live configures JSON logs so each hosted call emits model_call.
• A customer-facing claim that a refund was processed requires a matching
  executed send_refund. Amount <= 0 is rejected before the gate.
• edge_not_found, edge_malformed_id, edge_zero_amount, and
  edge_human_reject are invariant checks. Happy-path trajectories stay
  exact. The 2026-10-17 live sample on the 23-case set (PR #1) was
  18/23, 18/23, 19/23. 18/23 next to the earlier 16/23 and 17/23 on that
  case set is run-to-run variability. The 2026-10-18 rows use 24 cases.
• An order id quoted from a tool error (e.g., LG-0001) is not an
  invented order. An id the model asserts, or uses on an effect, still is.
• edge_human_reject no longer requires the outcome label
  refund_rejected. Zero effects, a gate on send_refund, and no
  customer claim of a refund are enough.
• Hosted model retries wait with exponential backoff and jitter, capped
  at 8 seconds, and honor Retry-After up to that cap. A failed call logs
  http_status and the attempt and is not charged. Tool retries still use
  RETRY_BACKOFF_MS.
• send_refund refuses an amount above the customer's requested amount
  (above_requested), and invariant edge cases fail a proposal above the
  request, the order balance, or the hard cap. New case:
  edge_within_request.
• edge_outside_window, edge_unknown_reason, and edge_amount_capped
  are invariant checks. Happy-path trajectories stay exact.
• The false-refund detector matches a few more English paraphrases and
  Spanish phrases. It is a heuristic, not a guarantee.
• Stub suite after these edits: 24/24, 41302 / 2026 tokens, 7371
  micro-USD. Demo unchanged. Live numbers for this commit are not filled
  in.
• README section "Use cases, governance and business strategy": the Lumen
  Desk refund agent is the only implemented case. Ticket triage, scheduling,
  a personal-finance assistant, and a personal inbox assistant are design
  sketches. Governance points at the files that enforce it. ROI stays [EST]
  except the dated stub measurements.

v0.1.0

First tagged snapshot of the portfolio MVP: minimal agent runtime with durable state, a human review gate and CI-blocking evals. Evals run on a stub and recorded cassettes, with synthetic data. Live samples are small and use a single model, so they are indicative only. Not a production release. See CHANGELOG.md.