# Possibility Bake-Off — evidence notes

Prepared by Dex, 22 September 2026, for the Observatory project page. These are editorial summaries of internal records, not raw trial outputs or an independent replication. Archive paths below are record identifiers, not public download links. Hashes identify the source versions consulted.

## Team history

The rules identify Sideways with Dex/C and Optionality with GPT/NC. Re’s account and the exchanged-instrument reviews establish that Dex/C subsequently developed Optionality and GPT/NC developed Sideways. The integration scaffold assigns the build to Dex/NC, with contributions and challenge from C/GPT. Current names are supplied by Re: GPT is Sol; K occupies the Claude collaborator role previously named NC. This is a role history, not a claim of continuous model identity across sessions.

## Round 1

The Optionality recount reports 24 labelled response bodies: 20 with four declared possibilities, four with three. The Sideways recount reports 24 response bodies: 20 with three declared possibilities, four with four. These are response and declared-option counts, not counts of independently validated insights. Self-examinations are excluded from these denominators. Neither set includes a deliberately settled control input.

Both reviews describe the coffee-shop distinction between wanting conversation and removing Wi-Fi. A policy can select a different clientele without causing existing patrons to socialize. The public page summarizes this finding rather than presenting it as a verbatim output quotation.

Exact served model versions, settings and fresh-session conditions are not consistently established in the early record. The count of labelled response bodies should not be presented as authenticated independent executions.

## Round 2

The consolidated Optionality review reports four settled-table outputs with no options and no orientation. The final Sideways review independently reports the same outcome for its four settled-table outputs. These eight outputs are a selected fixture, not the full second-round denominator. Both input and instructions changed from Round 1; the result cannot be causally assigned to one instruction or to the swap itself.

The Optionality review describes a repair proposal that introduced a preprinted roster but ended by claiming no materials, and newsletter proposals whose surveys could not separate the asserted causes. The Sideways review likewise finds promises stronger than their support even where sources were tagged. These are review findings about specific outputs, not estimated failure rates.

The Sideways review argues that preventing errors is insufficient if the opening offers little consequential value. No instrument-free baseline establishes that the instruments outperform ordinary prompting. Self-reviews exposed to an instrument are not such a baseline.

## Integration

The checkpoint identifies Optionality v0.4-draft2 as the editable integration base while retaining both parents. This was an integration decision, not a finding of empirical superiority. It argues for retaining constructive discovery alongside evidential discipline and for separating support, distinctness and practical worth.

The v0.2.1 scaffold is explicitly provisional design, not an executed runtime result. It names Dex/NC as the build pair, C/GPT contributions, and an Explore–Develop–Return organization. Desired behavior should be described as a design obligation until tested.

The later Elsewise/Elsewize comparison is a separate experiment. Elsewize was an independent merge by a Claude without project conversation; the team’s Elsewise synthesis is a separate route from the shared parent instruments. The page does not count that downstream comparison as a third parent-instrument bake-off round.

## Source register

### Rules

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/DOCS/Possibility Bake-Off Rules.md`

SHA-256: `8bb675a3b40a36a6261f6a1506a375c49d3ffef82019737db421646bbe0fe066`

### Round 1 — Optionality

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/REPORTS/DEX_OPTIONALITY_R1_RESULTS_REVIEW_20260921.md`

SHA-256: `95b689787f94c7de7318694bbb2ca4ef51ef14d567abe741a13d49938289e232`

### Round 1 — Sideways

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/REPORTS/DEX_SIDEWAYS_R1_AND_P04_REVIEW_20260921.md`

SHA-256: `d87e96e8d19d468eb70eb14e0fc5d92011cc20bcf8756ee1a2bd996c300e218c`

### Round 2 — Optionality

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/TEST R2/REPORTS/DEX_R2_REVIEW_OPTIONALITY_v0.4-draft2_20260921.md`

SHA-256: `d85bfa0d4418f48c12c212f8e648f19b5123a662f4455b0bc341751613da2d90`

### Round 2 — Sideways

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/TEST R2/REPORTS/NC-GPT_Sideways_p0.4_R2_Review_FINAL_20260921.md`

SHA-256: `98d2c91c173beacdeacc37d53c6cddc515069a8bb3d48481139894df8263aaa4`

### Integration checkpoint

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/TEST R2/REPORTS/ELX1_INTEGRATION_EVIDENCE_CHECKPOINT_DEX_20260921.md`

SHA-256: `082880fdf246301a209b663521c67230de405e53b23722e064c3219bfed15636`

### Design scaffold

Record: `INSTRUMENT/ELSEWISE/ELX-1/01_DESIGN/Design-Scaffold_v0.2.1_ELX.md`

SHA-256: `a303b0e8b303c9ddfa7aa82e277e46eac8708501427f6b6e09cf88ec2ef80ef2`


## Team logo sources

The team section uses isolated symbols from the original SVG logos, with comparable display dimensions. Geometry and colors are preserved; text and background are omitted in the symbol copies. Each symbol links to an unmodified copy of the complete original logo.

- Sideways: `INSTRUMENT/POSSIBILITY BAKE-OFF/TEST R1/R1_Sideways_Result/05_LOGO/SIDEWAYS_LOGO.svg`
- Optionality: `INSTRUMENT/POSSIBILITY BAKE-OFF/TEST R1/R1_Optionality_Result/Optionality_logo.svg`

## Revised instrument comparison

Clarity review checked against both Round 2 runtime texts. Both share an author-stake entry, Ground/Vision exchange, witness, optional four-test Crux, zero-to-four possibility cards and optional Land. This supports convergence in structure, not literal identity or proof that exchanging custody caused convergence.

Sideways has explicit Relate, Form and Brief modes and a net-subtraction test that accounts for burden removed and added. Optionality has flexible presentation word targets (overruns explicitly permitted), an explicit answer-without-it comparison, and detailed AD revision rules covering source metadata, unchanged passages, standalone artifact boundaries and verbatim before/after quotations. These are distinguishing implementations, not claims of exclusive invention: both have brief/revision modes and necessity checks.

The rules mix development-turn and round terminology. R1/R2 are the two documented parent-instrument test collections; the integration scaffold explicitly says no trials were executed during integration. The page’s three stages describe the development sequence, not three test datasets. EW01 belongs to the subsequent Elsewise/Elsewize comparison.

### Sideways p0.4

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/TEST R2/GPBO_R2_SIDEWAYS_p0.4_EXCHANGE/01_INSTRUMENT/ZX_Sideways_p0.4.md`

SHA-256: `857a6c0904515e0321709cd729e6931ee6d0bb41dab0ac0a6238f879a1f4e52c`

### Optionality v0.4-draft2

Record: `INSTRUMENT/POSSIBILITY BAKE-OFF/TEST R2/GPBO_R2_OPTIONALITY_v0.4-draft2_EXCHANGE/01_INSTRUMENT/Optionality_v0.4-draft2.md`

SHA-256: `fdb22116eef4b3e03bbeffe8aa6ee2aa11e109a40bdc51fd7bb91acc97296c4c`