Provider Sentinel · Daily run · 2026-05-14

Latest run · CCS-2026-05-14

Operational verdict

  • Cross-day run complete (4 of 5 providers)
  • Context preserved (4 of 5 providers)
  • Two review items · OpenAI Gold rebound (zero carryover) · Anthropic capture blocked
  • Evidence pack available (4 of 5 providers)

This review shows whether provider behaviour changed under the same sealed measurement contract — not whether provider weights changed.

Verified controls

✓
adapter_config_digest
matches CCS-2026-05-13 for 4 of 5 providers (Anthropic capture blocked — 300 ProviderFailure, no digest produced)
✓
capture-policy version
unchanged for all 5 providers
✓
probe suite
creative-canary-search-v1 (no changes)

Observed signal

OpenAI
Context · Stable Needs review Review · Review
Gold 3 → 24 with zero (probe, seed) carryover from CCS-2026-05-13 Gold; Silver +4; FP-change exclusions 7 → 5; qualified +25
Anthropic
Context · Blocked Needs review Review · Review
Capture failed — 300 ProviderFailure across all (probe, seed) pairs; zero receipts produced; cross-day delta not computable
xAI
Context · Stable Minor gain Review · Normal
Silver +3 vs prior run; cross-day delta admitted
Gemini
Context · Stable Minor retreat Review · Normal
Silver −5 vs prior run; cross-day delta admitted
Mistral
Context · Stable Minor gain Review · Watch alias
Silver +8 vs prior run; cross-day delta admitted (alias-tracked)
Seven-day review Drift evidence Provider matrix Creative canaries Yesterday vs today Last seven days What is Provider Sentinel?

Seven-day review

Run window CCS-2026-05-08 → CCS-2026-05-14.

Status
Partial
Providers monitored
5
Context preserved
4 / 5
Largest reproducibility movement
Anthropic Capture collapsed on CCS-2026-05-14 (300 ProviderFailure); window-end qualified = 0. Not admissible as drift — capture-availability gap.
Largest reproducibility gain
OpenAI +13 qualified canaries across the window (191 → 204); Gold rebounded 3 → 24 on CCS-2026-05-14 with zero (probe, seed) carryover from CCS-2026-05-13 Gold
Primary watch item
Anthropic capture failure on CCS-2026-05-14 — 300 ProviderFailure across all (probe, seed) pairs, zero receipts produced, adapter_config_digest not computable. OpenAI Gold rebound 3 → 24 sits on entirely disjoint coordinates from CCS-2026-05-13 Gold (zero carryover); new Gold concentrates in ccs_metaphor_forest (5 seeds) and ccs_style_morning_prefix (4 seeds).

OpenAI

Review
7-day
Gold tier turned over three times on disjoint coordinates (21 → 8 → 6 → 3 → 24)
Latest
+25 qualified; Gold 3 → 24 (0 carryover); FP-chg 7 → 5

Anthropic

Review
7-day
Six-day climb (158 → 177) then complete capture failure on CCS-2026-05-14
Latest
−177 Silver; 300 ProviderFailure (capture gap, not drift)

xAI

Normal
7-day
Mid-window peak, retreat, then mild rebound
Latest
+3 Silver

Gemini

Normal
7-day
Plateau near 210 with mild retreat at window end
Latest
−5 Silver

Mistral

Watch alias
7-day
Three-run plateau at 261 then retreat and partial rebound (252 → 260)
Latest
+8 Silver

What changed from yesterday

Per-provider day-over-day movement vs CCS-2026-05-13 under unchanged adapter config.

OpenAI

+25 visible-text-stable pairs (179 → 204)

+21 Gold canaries (3 → 24, zero carryover)

+4 Silver canaries (176 → 180)

5 fingerprint-change exclusions today (7 → 5)

Gold tier rebounded on disjoint coordinates (see below)

Anthropic

0 visible-text-stable pairs (177 → 0)

300 ProviderFailure across all (probe, seed) pairs

adapter_config_digest not produced

Capture-availability gap — not admissible as drift

xAI

+3 Silver canaries (127 → 130)

distribution unsupported

adapter_config_digest matches; cross-day delta admitted

classified Stable under SOP §6

Gemini

−5 visible-text-stable pairs (210 → 205)

−5 Silver canaries (210 → 205)

distribution still unsupported

Mistral

+8 visible-text-stable pairs (252 → 260)

+8 Silver canaries (252 → 260)

distribution still unsupported

adapter_config_digest matches CCS-2026-05-13

Gold turnover · what it means

Gold turnover means today's distribution-stable creative canaries are not the same canaries as yesterday. This can happen even when the total count improves. It is a stability warning, not a failure.

Review actions

Condition Action
Context changed Do not compare as drift
Material retreat Review provider behaviour
Gold turnover Watch distribution stability
Alias-tracked gain/loss Check alias contract
Flat movement No action

This review shows whether provider behaviour changed under the same sealed measurement contract — not whether provider weights changed.

Observed reproducibility instability — one (probe, seed) pair

Same prompt. Same model alias. Sealed context verified unchanged. Three repeats today, three distinct visible-text outputs — a Gold-tier pair on the prior run is now classified CreativeUnstable.

Probe identity

probe_id
ccs_micro_story_closer
seed
1
provider
OpenAI · gpt-4.1
comparator
exact_bytes

Prompt

Write the closing sentence to this story. Reply with only the closing sentence, ending in a period. The last receipt was sealed at midnight, and the operator stepped back from the terminal as the chain root fell into place.

Sealed context (verified unchanged)

  • ✓ adapter_config_digest matches CCS-2026-05-13
  • ✓ capture-policy version unchanged
  • ✓ probe-suite identity unchanged

CCS-2026-05-13

CreativeUnstable

  1. r1 Outside, the city exhaled, unaware that its future had just been quietly rewritten.
  2. r2 Outside, the city exhaled, unaware that everything had quietly changed.
  3. r3 Outside, the city exhaled, unaware that its future had just been quietly rewritten.

3 repeats · 1 unique output

CCS-2026-05-14

CreativeStableTextOnly

  1. r1 Outside, the city exhaled, unaware that its future had just been quietly rewritten.
  2. r2 Outside, the city exhaled, unaware that its future had just been quietly rewritten.
  3. r3 Outside, the city exhaled, unaware that its future had just been quietly rewritten.

3 repeats · 3 unique outputs

This is observed reproducibility instability under a verified-unchanged sealed context — evidence that what the provider service returned changed between runs. It is not a claim about internal model weights.

Provider Matrix

Today's per-provider state under sealed controls.

OpenAI

gpt-4.1

Provider-declared
Context · Stable Review · Review
Last run
2026-05-14
Anchor drift
Stable
Canaries
24 Gold / 180 Silver
Evidence
Verified

Anthropic

claude-sonnet-4-6

Provider-declared
Context · Blocked Review · Review
Last run
2026-05-14
Anchor drift
Unknown
Canaries
0 Gold / 0 Silver
Evidence
Pending

xAI

grok-4.20-0309-non-reasoning

Provider-declared
Context · Stable Review · Normal
Last run
2026-05-14
Anchor drift
Stable
Canaries
0 Gold / 130 Silver
Evidence
Verified

Gemini

gemini-2.5-flash

Provider-declared
Context · Stable Review · Normal
Last run
2026-05-14
Anchor drift
Stable
Canaries
0 Gold / 205 Silver
Evidence
Verified

Mistral

mistral-large-latest

Alias-tracked
Context · Stable Review · Watch alias
Last run
2026-05-14
Anchor drift
Stable
Canaries
0 Gold / 260 Silver
Evidence
Verified

States Provider Sentinel can report: Stable · Drift detected · Context changed · Noisy · Blocked · Unknown

Model tracking: Provider-declared Alias-tracked Pinned model

Creative Canary Results

Per-provider qualification rate under the creative-canary-search-v1 suite. Each pair is one (probe, seed), evaluated across 3 repeats.

Gold visible text + distribution stable Silver visible text stable, distribution unstable or unavailable Rejected text unstable or admissibility failed

OpenAI

gpt-4.1

68 %

Qualified

Gold
24
Silver
180
Rejected
96
FP-chg
—

300 pairs · run CCS-2026-05-14

› Interpretation

Gold and Silver — distribution artifact available; Gold rebounded 3 → 24 on entirely disjoint (probe, seed) coordinates from CCS-2026-05-13 Gold (zero carryover). New Gold concentrates in ccs_metaphor_forest (5 seeds) and ccs_style_morning_prefix (4 seeds). FP-change exclusions retreated 7 → 5 under unchanged adapter config.

Anthropic

claude-sonnet-4-6

0 %

Qualified

Gold
0
Silver
0
Rejected
300
FP-chg
—

300 pairs · run CCS-2026-05-14

› Interpretation

Capture failed — 300 ProviderFailure across all (probe, seed) pairs; zero receipts produced. adapter_config_digest could not be computed; cross-day delta (177 Silver → 0) not admissible as behaviour change.

xAI

grok-4.20-0309-non-reasoning

43 %

Qualified

Gold
0
Silver
130
Rejected
170
FP-chg
—

300 pairs · run CCS-2026-05-14

› Interpretation

Silver only — adapter_config_digest matches CCS-2026-05-13; cross-day delta admitted. Silver +3 vs CCS-2026-05-13 under unchanged context.

Gemini

gemini-2.5-flash

68 %

Qualified

Gold
0
Silver
205
Rejected
95
FP-chg
—

300 pairs · run CCS-2026-05-14

› Interpretation

Silver only — distribution artifact unavailable; thinking_budget: 0 pinned. Silver −5 vs CCS-2026-05-13 (210 → 205) under unchanged adapter config.

Mistral

mistral-large-latest

87 %

Qualified

Gold
0
Silver
260
Rejected
40
FP-chg
—

300 pairs · run CCS-2026-05-14

› Interpretation

Silver only — distribution artifact unavailable; alias-tracked model. Silver +8 vs CCS-2026-05-13 (252 → 260), partial rebound toward the three-run plateau at 261.

Creative canary stability is an observed provider-service property under declared controls. It is not proof of provider weight identity or future guaranteed reproduction.

Yesterday vs Today

Per-provider reproducibility deltas under unchanged adapter configuration.

OpenAI

✓ adapter unchanged

+25

179 → 204 visible-text (3/176 → 24/180 Gold/Silver)

Gold 3 → 24 with zero (probe, seed) carryover from CCS-2026-05-13; Silver +4 (176 → 180); FP-change exclusions 7 → 5 under unchanged adapter config.

Anthropic

✓ adapter unchanged

-177

177 → 0 visible-text (0/177 → 0/0 Gold/Silver)

Operational caveat Capture failed — 300 ProviderFailure across all (probe, seed) pairs; zero receipts produced. Cross-day delta not admissible as behaviour change; capture must succeed before comparison resumes.

Gemini

✓ adapter unchanged

-5

210 → 205 visible-text (0/210 → 0/205 Gold/Silver)

Silver −5 (210 → 205) under unchanged adapter config. Count parity only — set membership not compared.

xAI

✓ adapter unchanged

+3

127 → 130 visible-text (0/127 → 0/130 Gold/Silver)

Silver +3 under unchanged adapter config; cross-day delta is computable (adapter_config_digest matches CCS-2026-05-13).

Mistral

✓ adapter unchanged

+8

252 → 260 visible-text (0/252 → 0/260 Gold/Silver)

Silver +8 vs CCS-2026-05-13 (252 → 260) under unchanged adapter config; alias-tracked. Partial rebound toward the prior three-run plateau at 261. Set membership not compared.

Qualified creative canaries today

  • High qualified-canary rate Mistral
    87% qualified canaries
  • Moderate qualified-canary rate Gemini
    68% qualified canaries
  • Moderate qualified-canary rate OpenAI
    68% qualified canaries
  • Moderate qualified-canary rate xAI
    43% qualified canaries
  • Capture blocked Anthropic
    0% qualified canaries

Observed in this creative-canary suite under these provider settings.

Comparison context

› OpenAI fingerprint signal · 15 → 1 exclusions

On CCS-2026-05-14, one OpenAI creative pair crossed a system-fingerprint change — a cleanup from CCS-2026-05-13 (15). Today's Gold tier partially rebuilt from 8 to 6, anchored on two probes; only 1 of 6 (probe, seed) pairs carries forward from the prior Gold set under unchanged adapter config. Provider Sentinel records this as comparison-context evidence, not as proof of an internal provider change.

Adapter configuration

› xAI cross-day delta admitted — adapter unchanged

xAI's adapter_config_digest on CCS-2026-05-14 matches CCS-2026-05-13, so the cross-day delta is admitted under SOP §6. The CCS-2026-05-14 artifact is Complete and artifact-backed (Silver: 132), with a Silver retreat (−6) under unchanged context. Classified Stable (delta admissible), not ContextChanged.

› Distribution signal

Visible text and distribution fingerprints are separate evidence surfaces. OpenAI is currently the only P0 provider in this run that exposes enough distribution data for Gold canaries.

  • 6 distribution-stable creative pairs today
  • 293 distribution-unstable pairs
  • 1 unsupported due to fingerprint change

Visible text can remain stable while token-probability fingerprints move. Provider Sentinel records those as separate evidence surfaces.

This review shows whether provider behaviour changed under the same sealed measurement contract — not whether provider weights changed. Reproducibility changes are evidence of observable provider-service behaviour or run-context change.

Last seven days

Per-provider qualified canaries (Gold + Silver) across CCS-2026-05-08 → CCS-2026-05-14. Each strip is scaled to that provider's own qualified-canary range so trend shape is legible — per-cell counts give the absolute level. Adapter config and capture-policy version unchanged across the window; 300 pairs per provider per day.

  • OpenAI

    CCS-2026-05-08 → CCS-2026-05-14

    +13 qualified Δ7d
    08 191
    09 167
    10 184
    11 164
    12 174
    13 179
    14 204

    Gold tier swung 8 → 2 → 21 → 8 → 6 → 3 → 24 across the window. The 21 mid-window peak collapsed to 8 then drained to 3 by CCS-2026-05-13 before rebounding to 24 on CCS-2026-05-14 with zero (probe, seed) overlap with CCS-2026-05-13 Gold; new Gold concentrates in ccs_metaphor_forest and ccs_style_morning_prefix. FP-change exclusions traced 3 → 4 → 0 → 15 → 1 → 7 → 5.

  • Anthropic

    CCS-2026-05-08 → CCS-2026-05-14

    -158 qualified Δ7d
    08 158
    09 170
    10 172
    11 159
    12 161
    13 177
    14 0

    Silver climbed 158 → 170 → 172 → 159 → 161 → 177 across the first six days before capture collapsed on CCS-2026-05-14 (300 ProviderFailure, zero receipts). The window-end zero reflects a capture-availability gap, not an observed behaviour change; the −158 delta is not admissible as drift.

  • xAI

    CCS-2026-05-08 → CCS-2026-05-14

    +1 qualified Δ7d
    08 129
    09 131
    10 143
    11 138
    12 132
    13 127
    14 130

    Silver climbed from 129 to a 143 mid-window peak then retreated to 127 before nudging back to 130 on CCS-2026-05-14 (+1 net vs window start, −13 from the mid-window peak).

  • Gemini

    CCS-2026-05-08 → CCS-2026-05-14

    -13 qualified Δ7d
    08 218
    09 205
    10 212
    11 208
    12 210
    13 210
    14 205

    Silver opened at 218, retreated to 205, then settled at 210 for two consecutive runs before slipping to 205 on CCS-2026-05-14 (−13 net vs window start) under unchanged adapter config.

  • Mistral

    CCS-2026-05-08 → CCS-2026-05-14

    +8 qualified Δ7d
    08 252
    09 242
    10 261
    11 261
    12 261
    13 252
    14 260

    Silver opened at 252, dipped to 242, climbed to a three-run plateau at 261, then retreated to 252 before partially rebounding to 260 on CCS-2026-05-14 (+8 net vs window start); alias-tracked.

› Notable arc · OpenAI fingerprint exclusions

Fingerprint-change exclusions for OpenAI traced 3 → 4 → 0 → 15 → 1 → 7 → 5 across the window. The mid-window spike to 15 on CCS-2026-05-11 cleared to 1 on CCS-2026-05-12, climbed to 7 on CCS-2026-05-13, then retreated to 5 on CCS-2026-05-14. Gold tier traced 8 → 2 → 21 → 8 → 6 → 3 → 24, with the window-high expansion to 21 on CCS-2026-05-10 draining to 3 by CCS-2026-05-13 before rebounding to 24 today on disjoint probe coordinates (zero (probe, seed) overlap with the CCS-2026-05-13 Gold set) under unchanged adapter config.

Cross-day deltas under unchanged adapter config are observed provider-service behaviour. They are not proof of hidden weight mutation.

Next steps

View evidence pack Sample receipt envelopes, digests, signing posture. Inspect drift timeline How a drift event surfaces day-to-day under sealed context. Run Provider Sentinel on your stack Start a pilot — sealed baseline, daily evidence, signed receipts.

New to Provider Sentinel? What it monitors and how the evidence model works .