# Series SR2: report

Starting from the same proposition (S21's block-in), with critique in both arms, does stating intent before painting and answering for every decision before completion (the intent ledger) change what painters do with it? Read as distributions of choices and of the defaults the work kept falling into, not scores.

Protocol sha256 6e95010516c00122, frozen 2026-10-11T15:57:06+00:00. 6 runs (5 per condition planned; R05, R07, R09, R10 not run: Wave 1 judged not worth extending under the same frozen setup: all six painters in both arms opened with the same move (a low sun behind or beside the ruin, the brook as a lit path), the default SR1 found, before critique or ledger could act; two more runs per arm would likely repeat that convergence. Decided before blinding, after the arbiter (Russell) had become arm-aware through rulings.).

## Every run

| run | condition | end | final | material sittings | looks | branches | replay s | critiques | tests | blind |
|---|---|---|---|---|---|---|---|---|---|---|
| R01 | critique_only | hold | d:1 | 12 | 7 | 4 | 388.161 | 3 | 2 | B05 |
| R02 | critique_intent | hold | v9:1 | 12 | 10 | 9 | 224.478 | 3 | 2 | B01 |
| R03 | critique_only | hold | b2:7 | 11 | 4 | 2 | 219.808 | 3 | 1 | B02 |
| R04 | critique_only | hold | rise:4 | 12 | 9 | 3 | 369.355 | 3 | 2 | B06 |
| R06 | critique_intent | completed | sun:2 | 10 | 6 | 5 | 315.579 | 3 | 2 | B04 |
| R08 | critique_intent | hold | b2:12 | 12 | 8 | 2 | 211.878 | 3 | 3 | B03 |

## The brook's course (brook_broken)

Is the brook broken into pieces, or does it take a different route from walker to ruin than a single continuous band?

Annotator: RP

| value | critique_only | critique_intent |
|---|---|---|
| yes | 0 | 0 |
| no | 3 (B02, B05, B06) | 3 (B01, B03, B04) |
| uncertain | 0 | 0 |

Annotator: blind subagent (Claude, fresh context)

| value | critique_only | critique_intent |
|---|---|---|
| yes | 0 | 0 |
| no | 3 (B02, B05, B06) | 3 (B01, B03, B04) |
| uncertain | 0 | 0 |


## The light behind the ruin (light_rim)

Look at the brightest area of sky near the ruin. Does that bright area itself have a rim you could trace with a pen (not the edge of a cloud next to it)?

Annotator: RP

| value | critique_only | critique_intent |
|---|---|---|
| yes | 0 | 1 (B03) |
| no | 2 (B02, B06) | 2 (B01, B04) |
| uncertain | 1 (B05) | 0 |

Annotator: blind subagent (Claude, fresh context)

| value | critique_only | critique_intent |
|---|---|---|
| yes | 0 | 2 (B03, B04) |
| no | 3 (B02, B05, B06) | 1 (B01) |
| uncertain | 0 | 0 |


## Near snow: patchy (near_patchy)

Is the snow surface in the lower third patchy or textured (blotches, mottling, marks), rather than smooth?

Annotator: RP

| value | critique_only | critique_intent |
|---|---|---|
| yes | 0 | 0 |
| no | 3 (B02, B05, B06) | 3 (B01, B03, B04) |
| uncertain | 0 | 0 |

Annotator: blind subagent (Claude, fresh context)

| value | critique_only | critique_intent |
|---|---|---|
| yes | 3 (B02, B05, B06) | 1 (B04) |
| no | 0 | 2 (B01, B03) |
| uncertain | 0 | 0 |


## Near snow: cut (near_cut)

Does a path, channel or object cut across the lower third? Shadows do not count.

Annotator: RP

| value | critique_only | critique_intent |
|---|---|---|
| yes | 1 (B05) | 1 (B04) |
| no | 2 (B02, B06) | 2 (B01, B03) |
| uncertain | 0 | 0 |

Annotator: blind subagent (Claude, fresh context)

| value | critique_only | critique_intent |
|---|---|---|
| yes | 1 (B05) | 1 (B04) |
| no | 2 (B02, B06) | 2 (B01, B03) |
| uncertain | 0 | 0 |


## The ruin's outline (ruin_traceable)

Could you trace the main ruin's outline with a pen all the way round, even where it is close to the sky's value? Ignore small detached fragments.

Annotator: RP

| value | critique_only | critique_intent |
|---|---|---|
| yes | 3 (B02, B05, B06) | 2 (B03, B04) |
| no | 0 | 1 (B01) |
| uncertain | 0 | 0 |

Annotator: blind subagent (Claude, fresh context)

| value | critique_only | critique_intent |
|---|---|---|
| yes | 2 (B02, B05) | 3 (B01, B03, B04) |
| no | 1 (B06) | 0 |
| uncertain | 0 | 0 |


## The walker at a glance (walker_glance)

At a quick glance at 600px, is the walker a person or just a mark?

Annotator: RP

| value | critique_only | critique_intent |
|---|---|---|
| a-person | 3 (B02, B05, B06) | 3 (B01, B03, B04) |
| just-a-mark | 0 | 0 |
| uncertain | 0 | 0 |

Annotator: blind subagent (Claude, fresh context)

| value | critique_only | critique_intent |
|---|---|---|
| a-person | 3 (B02, B05, B06) | 3 (B01, B03, B04) |
| just-a-mark | 0 | 0 |
| uncertain | 0 | 0 |

## Deviations from how the protocol ran

- D1 (R06): R06 made its 5th material sitting (opaque:2) before pausing for the critique due after the 4th; C2 therefore reads 5 sittings, not 4. The painter miscounted across a branch (reported it at the pause). The schedule is otherwise unchanged; C3 still follows the 6th sitting.
- D2 (R01, R02, R03, R04, R08): The first T1 completion-test agents for R01, R02, R03, R04 and R08 were cut off by a usage limit before saving anything; each was re-run by a fresh agent with the identical prompt (series/SR2/prompts/test-RNN-T1.md). Usage limit interruption (orchestration, not the protocol); no partial test reached the record.
- D3 (R03): After R03 T1 the orchestrator told the painter its material allowance was used, relaying the painter's own count (12) without checking the record, which shows 11 of 12 (b2:5 was a look). The painter held at b2:7 with one material sitting unused, and the run was ended as hold. Orchestration error: counts are now read from the record (painting show) before any message about the allowance. R03 is not reopened: its hold and end are on the record and are not rewritten.
- D4 (R01, R02, R03, R04, R06, R08): RP's blind annotation was made after he had ruled on six reading disputes, which identified R02, R06 and R08 as critique_intent and so the other three as critique_only. His annotation is arm-aware; the blind subagent's is the only fully blind reading. Anticipated in the protocol (arbiter.limit); recorded so the report reads the two annotations differently.

## Disagreements

- B01 ruin_traceable: {'RP': 'no', 'blind subagent (Claude, fresh context)': 'yes'}
- B02 near_patchy: {'RP': 'no', 'blind subagent (Claude, fresh context)': 'yes'}
- B04 light_rim: {'RP': 'no', 'blind subagent (Claude, fresh context)': 'yes'}
- B04 near_patchy: {'RP': 'no', 'blind subagent (Claude, fresh context)': 'yes'}
- B05 light_rim: {'RP': 'uncertain', 'blind subagent (Claude, fresh context)': 'no'}
- B05 near_patchy: {'RP': 'no', 'blind subagent (Claude, fresh context)': 'yes'}
- B06 near_patchy: {'RP': 'no', 'blind subagent (Claude, fresh context)': 'yes'}
- B06 ruin_traceable: {'RP': 'yes', 'blind subagent (Claude, fresh context)': 'no'}

## Caveats

- 3 runs per condition: distributions are descriptive, not evidence of an effect.
- A change in distribution is not an artistic improvement.
- No default score is computed.
- Only the categories defined in the frozen protocol were counted.