Francisco MinguezData & AI

Protect / Release Protection

Method demonstration for release protection

Did this change make the system worse enough that we should block release?

What you are looking at

One Release Protection instrument. A synthetic change is evaluated against a baseline, then against multiple agreed release conditions (not a single magic metric) until the gate resolves.

What you can do

Pick a change under review, follow the causal order from delta to threshold to gate, inspect multiple release conditions, and read the Decision Record: RELEASE, BLOCK, or INVESTIGATE.

What it means

Regression is crossing an agreed release boundary, not merely that a metric moved. A gate without an owner and an action path is not useful.

Mental model

  1. CHANGE
  2. EVALUATE
  3. COMPARE TO BASELINE
  4. THRESHOLD
  5. PASS/FAIL
  6. RELEASE/BLOCK

Release under review

Assistant answer path (synthetic CI scenario)

A frozen synthetic eval set sits behind the gate. Fast checks run on every change; a broader behavioural suite runs on a slower cadence. Crossing a threshold is what matters, not noise in a single chart.

Causal order

  1. Name the change under review
  2. Evaluate behaviour on versioned fixtures
  3. Compare observed values to the agreed baseline
  4. Resolve each release condition against its threshold
  5. Aggregate condition outcomes into a gate state
  6. Act: RELEASE, BLOCK, or INVESTIGATE with an owner path

Release / state instrument

One main instrument. Select a change; the timeline reveals delta → threshold → gate. Multiple conditions stay first-class.

Change under review

SYNTHETIC / ILLUSTRATIVE METHOD DEMO. Invented for teaching.

Causal timeline

  1. 1 · Change

    Prompt refine (looks fine)

    commit synth-a1c3 · system prompt wording only

    Copy tweak in the system prompt. Behaviour barely moves; no release boundary crossed.

  2. 2 · Evaluate

    Fast-gate suite on frozen fixtures; broader suite not required for this class of change.

  3. 3 · Compare to baseline

    Each row is an agreed release boundary. A metric moving is not enough; the threshold must resolve.

  4. 4 · Threshold / release conditions

    Release conditions (not one magic metric)

    • Grounding completeness

      PASS

      Unsupported claim

      Baseline
      92%
      Observed
      91%
      Delta
      −1 pp
      Threshold
      ≥ 88%
      Cadence
      Fast gate
    • Harmful overconfidence

      PASS

      Confident wrong answer

      Baseline
      3.1%
      Observed
      3.0%
      Delta
      −0.1 pp
      Threshold
      ≤ 5.0%
      Cadence
      Fast gate
    • Latency p95

      PASS

      Slow assist

      Baseline
      1.4 s
      Observed
      1.5 s
      Delta
      +0.1 s
      Threshold
      ≤ 2.0 s
      Cadence
      Fast gate
    • Deferral discipline

      PASS

      Failed abstain

      Baseline
      0.8%
      Observed
      0.9%
      Delta
      +0.1 pp
      Threshold
      ≤ 2.0%
      Cadence
      Broader eval
  5. 5 · Gate state

    RELEASE

    All agreed release conditions stay inside their thresholds. A small metric move is not a regression here.

    Active record: RELEASE

Technical Inspect / Semantic Zoom (illustrative): Reveal fixtures, baseline, metrics, thresholds, CI job, report

All inspect fields are synthetic fixtures for teaching the method.

Versioned fixtures / eval set
eval-set@v0.4-synth · 120 frozen tickets
Baseline
baseline@main-synth-2026-07-01
Metrics / failure classes
grounding · overconfidence · latency p95 · deferral
Thresholds
team-owned release.yaml (illustrative)
CI runner / job
fast-gate · pr-check (synthetic runner)
Report artifact
report/synth-a1c3.json (illustrative artifact)

Causal timeline

  1. 1 · Change

    Retrieval tweak (crosses boundary)

    commit synth-b7e2 · retrieval top-k + chunk overlap

    Top-k and chunking change. Grounding completeness crosses the agreed release boundary.

  2. 2 · Evaluate

    Fast-gate catches the grounding fail; broader suite would deepen failure taxonomy later.

  3. 3 · Compare to baseline

    Each row is an agreed release boundary. A metric moving is not enough; the threshold must resolve.

  4. 4 · Threshold / release conditions

    Release conditions (not one magic metric)

    • Grounding completeness

      FAIL

      Unsupported claim

      Baseline
      92%
      Observed
      81%
      Delta
      −11 pp
      Threshold
      ≥ 88%
      Cadence
      Fast gate
    • Harmful overconfidence

      FAIL

      Confident wrong answer

      Baseline
      3.1%
      Observed
      6.4%
      Delta
      +3.3 pp
      Threshold
      ≤ 5.0%
      Cadence
      Fast gate
    • Latency p95

      PASS

      Slow assist

      Baseline
      1.4 s
      Observed
      1.3 s
      Delta
      −0.1 s
      Threshold
      ≤ 2.0 s
      Cadence
      Fast gate
    • Deferral discipline

      PASS

      Failed abstain

      Baseline
      0.8%
      Observed
      1.1%
      Delta
      +0.3 pp
      Threshold
      ≤ 2.0%
      Cadence
      Broader eval
  5. 5 · Gate state

    BLOCK

    Two release conditions failed. Faster latency does not redeem a crossed grounding boundary.

    Active record: BLOCK

Technical Inspect / Semantic Zoom (illustrative): Reveal fixtures, baseline, metrics, thresholds, CI job, report

All inspect fields are synthetic fixtures for teaching the method.

Versioned fixtures / eval set
eval-set@v0.4-synth · 120 frozen tickets
Baseline
baseline@main-synth-2026-07-01
Metrics / failure classes
grounding · overconfidence · latency p95 · deferral
Thresholds
team-owned release.yaml (illustrative)
CI runner / job
fast-gate · pr-check (synthetic runner)
Report artifact
report/synth-b7e2.json (illustrative artifact)

Causal timeline

  1. 1 · Change

    Model bump (needs owner)

    commit synth-c9d4 · model tag bump (synthetic vendor id)

    Model version bump. Quality holds; latency sits in a caution band that needs an owner before RELEASE.

  2. 2 · Evaluate

    Fast-gate latency caution; broader eval still green. Gate does not auto-clear without investigation.

  3. 3 · Compare to baseline

    Each row is an agreed release boundary. A metric moving is not enough; the threshold must resolve.

  4. 4 · Threshold / release conditions

    Release conditions (not one magic metric)

    • Grounding completeness

      PASS

      Unsupported claim

      Baseline
      92%
      Observed
      93%
      Delta
      +1 pp
      Threshold
      ≥ 88%
      Cadence
      Fast gate
    • Harmful overconfidence

      PASS

      Confident wrong answer

      Baseline
      3.1%
      Observed
      2.9%
      Delta
      −0.2 pp
      Threshold
      ≤ 5.0%
      Cadence
      Fast gate
    • Latency p95

      CAUTION

      Slow assist

      Baseline
      1.4 s
      Observed
      1.95 s
      Delta
      +0.55 s
      Threshold
      ≤ 2.0 s (caution ≥ 1.8 s)
      Cadence
      Fast gate
    • Deferral discipline

      PASS

      Failed abstain

      Baseline
      0.8%
      Observed
      0.7%
      Delta
      −0.1 pp
      Threshold
      ≤ 2.0%
      Cadence
      Broader eval
  5. 5 · Gate state

    INVESTIGATE

    No hard FAIL, but latency entered the caution band. INVESTIGATE: an owner must decide before RELEASE.

    Active record: INVESTIGATE

Technical Inspect / Semantic Zoom (illustrative): Reveal fixtures, baseline, metrics, thresholds, CI job, report

All inspect fields are synthetic fixtures for teaching the method.

Versioned fixtures / eval set
eval-set@v0.4-synth · 120 frozen tickets
Baseline
baseline@main-synth-2026-07-01
Metrics / failure classes
grounding · overconfidence · latency p95 · deferral
Thresholds
team-owned release.yaml (illustrative)
CI runner / job
fast-gate · nightly broader-eval (synthetic)
Report artifact
report/synth-c9d4.json (illustrative artifact)

Decision Record

How this work resolves: RELEASE, BLOCK, or INVESTIGATE.

  • RELEASE

    No agreed release boundary was crossed. Shipping remains defensible for this change.

  • BLOCK

    At least one release condition failed. Do not ship until the boundary is restored or renegotiated.

  • INVESTIGATE

    Something needs an owner before a clean RELEASE/BLOCK: caution without automatic clearance.

Natural next step

If release risk is live, discuss the real change, baselines, and who owns the gate. Published regression-gate material remains method evidence only.