JOB 132 - METHODS AUDIT AND RECOMMENDATIONS
Status: design advice only; no readers run and no study results assessed.
Audited local precedents: 131/receiver/START-HERE.txt, grammar.json,
resources.cjs, decode-score.cjs; 131/analysis/dialogue.txt,
exchange-verification.json, development-note.txt; viewer/disclosure.json
and the opening of viewer/first-interpretation.txt. Frozen job 131 was not edited.

1. QUESTION AND CLAIM BOUNDARY
Ask whether three supplied, assisted workflows let fresh-context instances of
the same model inspect one new conceptual composition, produce a grounded
interpretation, and construct an answer. The conditions are Carrier artwork
PNG plus reference resources, generic structured notation plus equivalent
resources, and explicit controlled prose plus equivalent resources. This is
a comparison of representation-and-tool packages, not a pure test of colour,
unaided perception, intelligence, consciousness, or semantic compression.
The shared 7 x 9 grammar itself constrains all three conditions. One artwork
does not support generalization to artworks, languages, readers, or models.

2. FREEZE BEFORE READERS
Compose the new score first. Freeze its canonical 63-cell selections, accents,
links, grammar, full checked 6,257-entry dictionary, candidate vocabulary and
all input assets. Derive both controls and the factual answer key from this
one score. Verify semantic equivalence before any reader receives material.
Locally hash and timestamp the protocol, score, controls, utilities, question
file, answer key, rubric and nine-run schedule. Call this local prespecification,
not externally registered or independently attested preregistration. Keep any
author reading, anticipated themes, comparative hypotheses and the 131 dialogue
outside reader inputs. The author reading must never become the grading key.

3. SEMANTIC PARITY AND TOOL PARITY
Give every condition verbatim copies of the same role definitions, relation
meanings, blank/accent semantics, truth boundary and full checked dictionary.
The prose control states every selected concept, address/role, accent, blank,
edge endpoint and edge direction explicitly in ordinary controlled sentences.
It adds no motivation, thematic transition, praise, inference or author essay.
Labels and names must identify the same entries as numeric IDs. Treat accent
as visual emphasis, even in prose. Do not omit awkward or redundant details
to make one control sound better. Normalized semantic checks should compare
all 63 cells, all 63 accents and the complete relation multiset.

Provide identical dictionary lookup/search, neutral reply validation and
available resource budgets. Do not give Carrier a rich resolved reading while
leaving controls to build their own equivalent. A common source-inspection
interface can expose the same resolved facts through condition-specific
adapters. If it outputs the same normalized table, explicitly report that this
tests assisted workflows with a common semantic endpoint; successful readings
cannot be attributed to visually reading the colours. Preserve original
presentation exposure and log whether the reader inspected it. Avoid requiring
Carrier-only reply encoding during the timed/common creative task: everyone
returns the same neutral reply-score schema, and all replies may subsequently
be encoded into Carrier by the parent. Report encoding as a separate check.
Do not covertly narrow one condition's reply vocabulary. If any curated
candidate list is used, the same list and selection rule apply to all nine,
with the full dictionary equally accessible; otherwise use all 6,257 entries.

4. NINE RUNS, COMMON SEQUENCE
Use exactly three fresh-context runs per condition, with the same model and
reasoning configuration, same session mechanism and tool permissions, and
one common instruction template except representation-specific filenames and
necessary loading directions. Record settings actually known; do not invent
seeds, temperatures, token counts or model build identifiers the host hides.
Use a prespecified randomized order balanced in three blocks, one condition
per block, with the randomization record saved before dispatch. The blocks
help manage order/host drift but do not create matched independent subjects.
Use fresh input-only paths, no inherited conversation, no other jobs/readers,
no network and no author text. These are procedural boundaries unless the
host truly enforces them. Report actual access and any violation.

Stage A: inspect source and shared resources; write first interpretation to
a file, hash it and retain it unchanged. Use identical length guidance, for
example 600-900 words. Require grounded reading plus explicitly marked
interpretive additions and uncertainties; permit criticism and disagreement.
Stage B: without seeing factual probe questions, create a reply score with
the same common 7 x 9 schema, then write its explanation (e.g. 300-500 words).
Freeze both. Allow deliberate blanks and a documented refusal to compose;
do not quietly reclassify refusal as a technical crash or erase it.
Stage C: reveal the same fixed factual probes and response schema to every
run. Readers may consult their inputs. No revisions of A or B follow probes.
This ordering prevents probes from teaching a preferred initial reading or
steering the reply. A probe-first design would answer a different question.

Set one common finite run budget and save it before launch. Do not stop early
after a striking result or add runs until one condition looks favorable. Keep
all nine attempted runs in the report, including incomplete runs, tool errors,
invalid outputs and access deviations. Distinguish a provisioning failure
before source exposure from an observed workflow failure; document any restart
under a prewritten rule. No performance-driven reruns or quiet replacements.
Protocol/tool changes after launch require an amendment and visible version
boundary; do not pool changed runs as if conditions remained unchanged.

5. PRESPECIFIED OBSERVABLES
Keep several separate outcomes rather than a single superiority score:
(a) Workflow: reached each frozen stage, explicit declines, required help,
    tool errors, validation attempts, unavailable outputs and access violations.
    Parent coaching would alter a run; log it instead of silently rescuing it.
(b) Structural: recovered/represented cells and accents exact; relations exact
    as a multiset with directed endpoints preserved; blanks explicit; IDs all
    checked; reply schema valid. Structural correctness is not philosophical
    quality. Supplied code validation is not independent implementation.
(c) Factual probes: a suggested fixed 16-item set is four cell/role facts,
    two blank/emphasis facts, six relation facts (one per operator), two exact
    definition/scope distinctions, and two grammar limitations. Choose actual
    addresses/entries using a recorded rule after score freeze, before readers.
    Freeze wording, item answer keys, acceptable variants and item weights.
    Avoid subjective themes or agreement with the author as factual answers.
    Report every item's outcome plus each run's count; unanswered remains
    unanswered/incorrect in the attempted-run denominator, not silently dropped.
(d) Interpretation/reply: use masked qualitative appraisal below. Report
    concrete divergences, productive challenges, omissions and overclaims.
(e) Effort: observed tool calls, lookups, revisions and active elapsed time if
    reliably captured. Host/orchestrator waits and model latency are confounds;
    do not turn rough wall time into a cognitive-efficiency claim.

6. QUALITATIVE APPRAISAL
Use at least one fresh appraiser, preferably two, not the composer or a reader.
Give appraisers the common canonical semantic reference and frozen dictionary,
but no author essay, condition labels, timings, condition hypothesis or results
summary. Randomize anonymized texts. Retain substantive text verbatim, including
format references; remove only identifying headers/file paths. Complete masking
is impossible when a text mentions colour or tools. Record appraisers' guessed
condition and confidence only after their assessment, and disclose this limit.
If two appraisers are unavailable, label a single appraisal accordingly.

For each interpretation, separately assess: (i) factual grounding, (ii) respect
for unresolved roles/relations and definitions, (iii) separation of supplied
content from added inference, and (iv) coherent development of a reading.
For each reply, assess: (i) traceable response to source features, (ii) agreement
between its explanation and actual reply score, and (iii) intelligible changes
or challenges. New-concept count and distance are descriptors, not creativity
or quality rewards. Mere praise, thematic convergence, greater verbosity and
agreement with author intent receive no credit by themselves.
Use an ordinal rubric per dimension: 0 = substantial documented failure;
1 = mixed/limited with specific gap; 2 = well-supported with concrete evidence;
NA = insufficient material. Require quotations, coordinates and counterexamples
for every judgment. Do not average this into an authoritative artistic rank.
Keep independent appraisals and disagreements visible; any later adjudication
is separately labeled. Explain what the rubric cannot capture aesthetically.

7. UNITS, ANALYSIS AND BYTES
The run is the analysis unit, n = 3 per condition; 16 probes within a run are
not 16 independent readers. Runs share a model, orchestrator, tools, vocabulary
and artwork, so call them fresh-context repeated runs, not independent human
subjects or cross-provider replication. Show all nine run rows and failure
denominators; use counts, medians/ranges and case examples. No significance
tests, population confidence claims, causal superiority or model-general
ranking are warranted. Ceiling performance is an informative limitation.

Report bytes as a separate engineering inventory: payload, condition-specific
assets/tools, common frozen dependencies, complete cold-start kit, and defined
warm-cache incremental transfer. State exactly which items each number counts.
Compare compression only with the same algorithm/settings and hash preserved
inputs. Avoid multiplying shared dependencies for one condition while omitting
them for another. PNG packet bytes are not the cost of the full usable artwork
and shared dictionary. Byte size cannot establish meaning preservation or
interpretation quality; those are separately observed outcomes.

8. PUBLIC REPORT
Publish frozen inputs, controls, tools, prompts, schedule, item keys, raw first
interpretations, replies, probe responses, failures/deviations, appraisal rubric
and appraisals. Separate technical verification, descriptive pilot findings and
the author's later personal reflection. The public headline should describe a
nine-run local exploratory comparison and its actual outcome, not a benchmark
victory. A later Carrier-word arm or different grammar is a new study, not a
post hoc way to rescue this one.
