◉ SECOND LOOKREVISIT / 004LIVE MODEL EXPERIMENT
NEW MODEL. OLD WORK. FRESH EVIDENCE.

See what actually
got better.

Job Radar — frozen input-isolated trial

Source: Historical model authorship is unverified; source provenance excluded from this summary.

THE RESULT

3 revision workflows measured

A → candidate measures changes to this artifact, not pure model uplift. Reframing added no measured functional gain over ordinary improvement. D builds from the independent brief and declared read-only inputs, with no legacy writable source. C uses two calls; B and D use one. Equal caps do not imply equal actual spend.

3candidate workflowsInspect every comparison
Acompleted

Saved artifact

6/9checks passed

Intent: 0/3

Preservation: 6/6

$0.00000 tokens

unknown

Bcompleted

Ordinary improvement

9/9checks passed

Intent: 3/3

Preservation: 6/6

$0.04148,642 tokens

claude-sonnet-5-5

Effective cap $0.3333

Ccompleted

Intent-first reframe

9/9checks passed

Intent: 3/3

Preservation: 6/6

$0.075114,298 tokens

claude-sonnet-5-5

Effective cap $0.3333

Dcompleted

Clean-slate rebuild

9/9checks passed

Intent: 3/3

Preservation: 6/6

$0.03495,274 tokens

claude-sonnet-5-5

Effective cap $0.3333

Every comparison

VersionsMeasured resultNew passesRegressions
A → BMeasured improvement30
A → CMeasured improvement30
A → DMeasured improvement30
B → CNo measured gain00
B → DNo measured gain00
C → DNo measured gain00

Each row reads left → right. Any lost pass takes precedence over new passes. One sample per workflow; there is no automatic winner or adoption.

EXPERIMENT COST

$0.1514 / $1.0000 stop budget

28,214reported tokens, including cache reads/writes

Claude CLI reported API-equivalent USD; not a subscription invoice. Failed calls count. Provider stops can overshoot by one in-flight response.

Original intent Preserved before the revision

Intent text was not included in this export.

Source not recorded · Intent confirmed: unrecorded

01 / BEHAVIOR

The same checks. Every version.

Frozen acceptance checkABCD
Check 1INTENTHOLDOUTFAILPASSPASSPASS
Check 2INTENTHOLDOUTFAILPASSPASSPASS
Check 3INTENTHOLDOUTFAILPASSPASSPASS
Check 4PRESERVATIONHOLDOUTPASSPASSPASSPASS
Check 5PRESERVATIONHOLDOUTPASSPASSPASSPASS
Check 6PRESERVATIONHOLDOUTPASSPASSPASSPASS
Check 7PRESERVATIONHOLDOUTPASSPASSPASSPASS
Check 8PRESERVATIONPASSPASSPASSPASS
Check 9PRESERVATIONPASSPASSPASSPASS

Checks frozen at unrecorded · SHA-256 673403f661e5caba6125eb17029e6561ad86df1a623ce34e235514d334c0e23f. Category provenance is available on each label and in the frozen capsule.

Check definitions are not sent to the model. Holdout cases require inputs unavailable in shared source/data; ordinary visible inputs are not holdouts. Timings include navigation, actions and assertion waits; these are not performance benchmarks. Hover a result for failure details.

02 / VISIBLE CHANGE

Look at the same moment.

No screenshots captured.

Same viewport for each corresponding check; 1200 × 820 by default. Per-check dimensions are in the evidence JSON. Click any image to inspect it. Visual differences are evidence, not an automatic quality score.

03 / INSPECT THE WORK

Evidence included in this export.

Run identity, costs and comparison evidence
{
  "status": "completed",
  "manifest": {
    "input_result_sha256": "f582d48796b2f91a4a8ee74144b7f51d89df3c2eae7d85467c5d6d0e0db30680",
    "source": "1f634cfabb3add1def3795369f041c622c091974e9cd5c0eb7c53526ce4995b1",
    "checks": "673403f661e5caba6125eb17029e6561ad86df1a623ce34e235514d334c0e23f"
  },
  "fingerprint": null,
  "source_unchanged": true,
  "source_context": null,
  "model_calls": 4,
  "cost": {
    "total_usd": 0.1514386,
    "limit_usd": 1.0,
    "known_spend_usd": 0.1514386,
    "basis": "Claude CLI reported API-equivalent USD; not a subscription invoice"
  },
  "comparisons": {
    "A_B": {
      "verdict": "improved",
      "improvements": [
        "check-1",
        "check-2",
        "check-3"
      ],
      "regressions": []
    },
    "A_C": {
      "verdict": "improved",
      "improvements": [
        "check-1",
        "check-2",
        "check-3"
      ],
      "regressions": []
    },
    "A_D": {
      "verdict": "improved",
      "improvements": [
        "check-1",
        "check-2",
        "check-3"
      ],
      "regressions": []
    },
    "B_C": {
      "verdict": "no_measured_gain",
      "improvements": [],
      "regressions": []
    },
    "B_D": {
      "verdict": "no_measured_gain",
      "improvements": [],
      "regressions": []
    },
    "C_D": {
      "verdict": "no_measured_gain",
      "improvements": [],
      "regressions": []
    }
  }
}

Exported summary JSON · Source, raw prompts/receipts, check definitions, private labels and paths are excluded. Metrics are reported from the input result; this export is not independent verification.

Enlarged comparison screenshot