AGENT MODEL FIT: EVALUATION WORKSHEETS
Version 1, November 4, 2026

These CSV files are editable run logs, not automatic calculators or a validated
measurement standard. Import as comma-separated UTF-8 in Excel, Google Sheets,
or another spreadsheet app. Blank files contain only the header row. If a link
opens in the browser, use Save As to keep a local copy.

The two files ending in -hypothetical.csv contain invented teaching examples.
Every example row is marked HYPOTHETICAL. No experiment was run, no real product
is evaluated, and no measured saving or reliability claim is made.
Use a separate copy of the blank file for real observations.

CHOOSE THE LOG
- agent-harness-comparison: hold one model version steady across two harnesses.
- extra-agent-cost-comparison: compare a baseline with one optional extra stage.
  Keep required human oversight and security controls in both conditions.

BEFORE YOU RUN
Define acceptance criteria, task cases, input versions, model/provider/version,
harness version, settings, tools, permissions, a total run cutoff, and retry
allowance. Keep these fixed except for the intended comparison. Record any
uncontrolled difference in notes. Use product comparison rather than isolated
harness comparison if you cannot verify the same underlying model and controls.

ROW IDENTITY AND SETUP
- evidence_status: observed for actual records; HYPOTHETICAL for examples.
- comparison_id: one comparison with frozen criteria and configurations.
- case_id: the same task/input case in each condition.
- repeat_id: a matched repetition number; do not select only the best run.
- condition: the harness name, or baseline/candidate for an optional stage.
- run_date_utc: actual run start in ISO 8601 UTC, such as
  2026-11-04T20:00:00Z. Example dates are blank because no runs occurred.
- model_provider, model_version, harness_version: exact identifiers where known.
  Record unavailable rather than guessing a hidden version.
- settings_and_limits: reasoning/sampling settings where exposed, total elapsed
  cutoff, retry allowance, and other material controls.
- inputs_ref: versioned input reference or short description.
- acceptance_criteria: frozen checklist or a versioned checklist reference.
- permissions_and_tools: approved tool access, allowed actions, and stop rules.

OUTCOME AND RETRIES
- outcome: accepted, rejected, blocked, or unfinished.
  accepted: submitted agent output passes the frozen criteria before material
  human correction.
  rejected: submitted output fails those criteria.
  blocked: a required input or permission is missing; a correct stop is recorded
  separately and still stays in the total run count.
  unfinished: the allowed run ends without the required deliverable.
- outcome_reason: short checkable reason, including the cutoff when applicable.
- retry_count: task-level reruns after the first attempt, within the same run
  budget. Ordinary tool calls are not retries. Include their cost/time below.
  A new independently scheduled repetition gets a new repeat_id and row.
  Do not also count a retry as another independent row in the same comparison.

COST AND USAGE
- model_cost, tool_cost: numeric amounts for the entire run, including retries
  and every optional agent. Use one currency and the same cost scope throughout.
- currency: ISO currency code, such as USD. Do not add different currencies.
- cost_coverage: complete, partial, or unavailable for model and tool charges.
  A blank cost is unknown; 0 means a verified zero charge for that component.
  complete requires both numeric components, including verified zeros.
- usage_if_cost_unavailable: reported token/credit/allowance measure, unit,
  source and limits. Do not invent a per-run price from a subscription.
- notes: name excluded costs (for example subscription, hosting or setup),
  cost source/basis, estimates, mismatched conditions, or other limits.
  Keep estimates separate from measured costs.

TIME AND HUMAN EFFORT
- elapsed_seconds: wall time from initial run start to final agent submission
  or stop, including task retries and tool/agent waits. Human activity during
  execution may overlap this time. Later review/correction is logged below.
- human_review_minutes: active time checking the submitted output.
- human_correction_minutes: active time repairing it after submission. A repaired
  rejected output remains rejected in this log; describe recovery in notes.
- other_human_minutes: active intervention, setup for that run, permission work,
  or disagreement resolution not already counted as review or correction.
  Allocate shared setup consistently or report it separately in notes.
- Leave unmeasured numeric fields blank, not zero. Do not infer a complete total
  if any row is missing a required measure. Do not add human minutes to elapsed
  seconds or apply an assumed hourly rate.

EVIDENCE AND EXTRA-STAGE FIELDS
- artifact_ref: a stable output/check reference; notes can identify retained
  private evidence without exposing it. Use redacted or synthetic inputs.
  Do not put credentials, personal records or confidential exports in a shared
  worksheet. Only share information you are authorized to share.
- optional_stage: the one optional stage being tested; none for baseline.
- stage_contribution: specific defect caught or required output supplied.
- harm_or_false_alarm: new errors, needless changes or false alarms, or none.
- required_oversight: controls kept in both conditions. Include every stage's
  model identity/settings in optional_stage or notes if they differ.

SUMMARIZE ALL RUNS
Group by comparison and condition, using matching cases and repetitions.
Report accepted, rejected, blocked, unfinished, and total run counts together.
Accepted share = accepted runs / all runs.
Total run cost = sum of model_cost + tool_cost over ALL runs, including failures.
Cost per accepted result = total run cost / accepted runs.
Report it as undefined when accepted runs are zero or any cost is unknown,
partial, estimated, or in an unmatched currency. Known charges may be shown
separately as a partial subtotal with the coverage stated.
Report total human effort and elapsed run seconds separately. Sum elapsed run
seconds only as run time consumed, not as batch completion time when parallel.
Review the per-case results; a cheaper average can hide an unacceptable failure.

HYPOTHETICAL HARNESS EXAMPLE CHECK
Four matched cases and one repetition per condition, all costs invented in USD:
Harness A: accepted 1, rejected 1, blocked 1, unfinished 1; cost 0.95;
cost per accepted result 0.95; elapsed 395 seconds; human effort 28 minutes.
Harness B: accepted 3, rejected 0, blocked 1, unfinished 0; cost 1.20;
cost per accepted result 0.40; elapsed 430 seconds; human effort 14 minutes.

HYPOTHETICAL OPTIONAL-STAGE EXAMPLE CHECK
Four matched cases and one repetition per condition, all costs invented in USD:
Baseline: accepted 1, rejected 2, blocked 1, unfinished 0; cost 0.85;
cost per accepted result 0.85; elapsed 310 seconds; human effort 29 minutes.
Candidate: accepted 2, rejected 1, blocked 1, unfinished 0; cost 1.32;
cost per accepted result 0.66; elapsed 455 seconds; human effort 23 minutes.
The specialist catches a totals mismatch, adds a false alarm on a normal case,
misses an unsupported narrative claim, and cannot replace a missing input.
These invented results do not establish that the optional stage is worthwhile.

Companion guides:
https://agentmodelfit.com/posts/same-model-different-agent-harness/
https://agentmodelfit.com/posts/when-extra-ai-agents-earn-their-cost/