# Review Methods 1.0 — experimental design edition

Advisory only. Criteria and calculation fixed before the walkthrough hand pass;
not a frozen API or a validated service. Five dimensions, each weight 1/5, share
the foundation below. Domain anchors change what is judged, not the denominator.
These are judgment anchors; arithmetic over them does not make them objective.

| Dimension | Prediction | Strategy | Project |
|---|---|---|---|
| Purpose | Decision-relevant, precise event and answer | Repeatable objective with identifiable markets and invalidation | Audience, intended action and a coherent reason to act |
| Supporting evidence | Timely, checked grounding; contrary findings considered | Out-of-sample discipline, realistic data and costs, contradictory regimes | Actual audience/problem support; reliable claims and attribution |
| Feasibility | Resolvable source/rule, executable horizon | Instrument, liquidity, execution and operational constraints | Delivery, resources, distribution and dependencies |
| Potential value | Informative coherent distribution and justified departures from reference | Plausible net benefit under costs and reference alternatives | Distinctive useful proposition compared with a relevant alternative |
| Downside | Sensitivities, ambiguity, disagreement and tail possibilities handled | Loss exposure, exits, failure modes and scope of risk limits | Misunderstanding, consent, reversibility and failure costs |

For each dimension use anchors `0` (known contradiction or unusable), `25`
(substantial demonstrated weakness), `50` (mixed support), `75` (strong reasoned
support with a material reservation), `100` (exceptionally strong support under
the declared conditions). Intermediate multiples of `5` express the evaluator's
reasoned interpolation, which must be written down. They do not imply precision
to five points. Unknown is null with a reason, not zero or a changed criterion.

Examples: a binary distribution whose probabilities sum correctly can still have
weak grounding. A proposal with draft tickets can be feasible to review without
being executable. A short campaign can have a clear, distinctive action without
audience evidence. An impossible production budget is known bad feasibility;
an unstated budget is unknown if the mechanism cannot reasonably be assessed.
An unsupported supplied claim is evidence of what was claimed, not verification
that the claim is true. An assumption or interpretation is never Evidence.

## Calculation and band

Let K be the known dimensions out of the fixed five. `S = sum(s_i)/K` for known
dimensions. Show S only if `K >= 3` and purpose and feasibility are known. Keep
all five dimensions in the Review. Missing width `M=100*(5−K)/5`.

For an independent second evaluation on the same anchored dimensions,
let J be the count of dimensions both can judge and use
`D=sum(abs(s_i−t_i))/(2*J)` on that intersection; separately disclose
that intersection's size and each unknown. If it does not cover the scored set,
use the greater of that measured width and the `5`-point unmeasured reserve.
With zero overlap, use the reserve and say unmeasured; do not divide by zero.
Without a second anchored evaluation use the `5`-point reserve and say unmeasured.
A generic headline disagreement is a diagnostic, not a substitute for dimension
disagreement. Future sensitivity `F` is the average explicitly reasoned dimension
sensitivity, between `0` and `20`, under stated plausible scenarios. The hand
walkthroughs use the following dimension sensitivity judgments. Each known
dimension receives a `10`-point sensitivity, not an observed error estimate:

| Domain | Purpose scenario | Support scenario | Feasibility scenario | Value scenario | Downside scenario |
|---|---|---|---|---|---|
| Prediction | wording becomes less decision-relevant | late lineup changes weaken grounding | source reliability changes near horizon | the reference case changes | tail circumstances become material |
| Strategy | market conditions change the use | data regime changes undermine support | quotes or liquidity change | costs remove the benefit | exits behave differently under stress |
| Project | visitors read the action differently | a broader audience challenges initial support | delivery availability changes | a similar offer reduces distinctiveness | confusion or operational constraints become material |

These are evaluator judgments that plausible changes could move each known
assessment by about ten points; at an endpoint movement is one-sided. For K known
dimensions `F=(10*K)/K=10`; unknown dimensions do not enter this average. With
no known dimensions there is no headline or centered band. New reviews should
record different sensitivities if justified; this fixed hand pass uses these
declared scenarios consistently. Range:
`[max(0,S−M−D−F), min(100,S+M+D+F)]`.

This deliberately broad sensitivity envelope is not a probability, posterior,
confidence interval or percentile. Correlated unknowns make precision especially
unwarranted. A new adverse fact can lower a Score; absence alone has no negative
numeric value. Mean-known normalization is not a claim about unknowable facts.
It may change when facts arrive; deleting known facts cannot reset it.
No headline means no band center, not an invented `0`. Publish dimension
judgments and uncertainty reasons instead. Calculate with Decimal, round export
to two decimal places HALF_UP for Review points, and integer display HALF_UP.
Result calculations use their own stated precision and never the Review scale.

## Whole-idea judgment and choices

Every Review includes the strongest objection, the main uncertainty and a useful
next step. Material findings outside the expected criteria are recorded as
supporting items with the effect on the recommendation and the dimension they
inform, if any. Do not silently add a weight, penalty or sixth dimension.
The infeasible film stress case scored `46` in 1.0 despite a decisive no-go:
that is a method weakness, retained in the challenge set. A numeric total is
never permission to proceed. Version 1.0 remains experimental for this use.

Comparison uses blind pairwise reasons under one explicit purpose, audience,
resource envelope, stage, method and reference. Tie, inconclusive and conditional
recommendation are valid outcomes. If premises differ, align them visibly or
refuse the comparison; never rank incomparable headline numbers. Explain what
would reverse the choice. A single Idea can be compared with a named reference.
Do not invent an outcome for an unimplemented alternative.

## Corrections, omission, independence and versions

Preserve original words, visible inferences and supporting material. Customers
can correct kind, purpose and consequential assumptions; create a new version.
Same mechanism/audience/objective with new prose is a revision. A materially new
decision can become a related Idea, with applicable adverse history retained.
A rebuttal needs support and leaves its predecessor visible. An undetected
duplicate or never-supplied fact remains an acknowledged limit.

Exclude owner identity, prior performance and willingness to pay from the judge's
inputs. Show affiliation and the prior Record beside the Review. Outcome-blind
review input extraction must also exclude outcome-revealing attachments; when a
historical author already saw outcomes, label the pass historical, as here.
Review reruns preserve the original, reason for rerun, evaluator/method versions
and numerical variation; independent agreement is not proof. A changed criterion
or calculation mints a new Method version and recomputes alongside old history.
None of these hand examples demonstrates prospective validity.
