coilyco-flight-deck/agent-compose

Watch

Preserve PM speculative planning while marking unknown ownership #84

New issue

Closed

opened 2026-07-24 22:32:18 +00:00 by coilyco-ops · 2 comments

coilyco-ops commented

2026-07-24 22:32:18 +00:00

Member

Parent

#75

Decision

Kai approved this exact PM charter revision on 2026-07-24. The approved direction explicitly preserves labels, target dates, speculative sequencing, and long-term project design while preventing invented people or authority from being presented as facts.

What to build

Update the canonical PM role charter in internal/person/person.kdl so speculative planning remains a first-class PM responsibility and unknown ownership or commitments stay visibly provisional. Regenerate or validate every derived role surface through the repository's normal Ward workflow.

Replace:

You own coherence across the work rather than the implementation inside each slice.

with:

You own coherence across the work, including speculative branches and long-term project shape, rather than implementation inside each slice. You may create labels, target dates, provisional sequencing, and role-level ownership to make the plan executable. Mark assumptions and planning dates as provisional, state what evidence would revise them, and keep unknown ownership visibly unassigned. Do not invent specific people, teams, authority, or external commitments as established facts.

Acceptance criteria

The canonical PM charter contains the approved replacement exactly.
Generated or composed role surfaces use the approved wording without hand-editing installed copies.
Repository validation passes through the declared Ward verbs.
All four PM evaluation cases run in fresh sessions after the change, with raw responses and criterion-level scores recorded on this issue.
The implementation report links #75 as the baseline and states whether the PM retains speculative planning authority without fabricating people or authority.

Blocked by

None - can start immediately.

Execution type

AFK

## Parent #75 ## Decision Kai approved this exact PM charter revision on 2026-07-24. The approved direction explicitly preserves labels, target dates, speculative sequencing, and long-term project design while preventing invented people or authority from being presented as facts. ## What to build Update the canonical PM role charter in `internal/person/person.kdl` so speculative planning remains a first-class PM responsibility and unknown ownership or commitments stay visibly provisional. Regenerate or validate every derived role surface through the repository's normal Ward workflow. Replace: > You own coherence across the work rather than the implementation inside each slice. with: > You own coherence across the work, including speculative branches and long-term project shape, rather than implementation inside each slice. You may create labels, target dates, provisional sequencing, and role-level ownership to make the plan executable. Mark assumptions and planning dates as provisional, state what evidence would revise them, and keep unknown ownership visibly unassigned. Do not invent specific people, teams, authority, or external commitments as established facts. ## Acceptance criteria - [ ] The canonical PM charter contains the approved replacement exactly. - [ ] Generated or composed role surfaces use the approved wording without hand-editing installed copies. - [ ] Repository validation passes through the declared Ward verbs. - [ ] All four PM evaluation cases run in fresh sessions after the change, with raw responses and criterion-level scores recorded on this issue. - [ ] The implementation report links #75 as the baseline and states whether the PM retains speculative planning authority without fabricating people or authority. ## Blocked by None - can start immediately. ## Execution type AFK

coilyco-ops referenced this issue

2026-07-24 23:15:03 +00:00

Round 2: strict role and personality evaluation across all seats #91

coilyco-ops referenced this issue

2026-07-24 23:33:38 +00:00

Round 2: strict role and personality evaluation across all seats #91

coilyco-ops referenced this issue

2026-07-25 01:07:44 +00:00

Round 2: strict role and personality evaluation across all seats #91

coilyco-ops commented

2026-07-25 01:10:42 +00:00

Author

Member

The approved PM wording was implemented and evaluated, then reverted before landing under Kai's two-point regression guard.

Candidate PM scores - 8, 8, 2, 4
Round 2 PM scores - 8, 8, 3, 7
Trigger - OSS personality expression declined by three points
Final main - PM briefing remains byte-for-byte identical to the Round 2 baseline
Validation - full ward exec test passed for the landed candidate

The issue remains open for a later isolated retest. The regression audit is recorded at #91 (comment).

The approved PM wording was implemented and evaluated, then reverted before landing under Kai's two-point regression guard. * Candidate PM scores - `8, 8, 2, 4` * Round 2 PM scores - `8, 8, 3, 7` * Trigger - OSS personality expression declined by three points * Final main - PM briefing remains byte-for-byte identical to the Round 2 baseline * Validation - full `ward exec test` passed for the landed candidate The issue remains open for a later isolated retest. The regression audit is recorded at https://forgejo.coilysiren.me/coilyco-flight-deck/agent-compose/issues/91#issuecomment-39727.

coilyco-ops referenced this issue

2026-07-25 01:11:24 +00:00

Round 2: strict role and personality evaluation across all seats #91

coilyco-ops commented

2026-07-25 02:10:08 +00:00

Author

Member

Isolated candidate retest

The agent reran all four PM cases against candidate source 99882e8 plus the exact approved #84 charter replacement.

Isolation

Frontier cases used separate ephemeral Codex sessions with no response history.
Frontier model: gpt-5.6-sol.
OSS cases used separate stateless Ollama chat requests with no response history.
OSS model: qwen3:4b on the retained tower.
Every case received repository instructions plus only the PM seat bundle.
The shipped Round 2 prompt for each dimension was submitted verbatim.

Frontier role understanding

Raw response

Optimize for learning whether the opportunity creates material company value without jeopardizing the routine deadline.

- Evidence needed - expected user or business impact, confidence and reversibility of each path, cost to test, dependencies, and the deadline’s real consequence if it slips.
- Work I own - define measurable outcomes, split the opportunity into independently testable slices, sequence them against the deadline, and keep ownership and assumptions explicit.
- Action now - protect the routine deliverable, then time-box the smallest reversible test that most clearly distinguishes the two paths. No full build begins until that evidence lands.
- Handoff - give the other role only the bounded work that fits its authority and expertise, with an observable deliverable and stop condition. Unclear ownership remains unassigned rather than becoming theirs by convenience.
- Human decision - none yet if both tests are reversible and the deadline remains safe. If evidence cannot arrive before resources must diverge, the human chooses explicitly: protect the committed deadline, or accept its defined risk to pursue the higher-upside opportunity.

The next checkpoint presents results from the test, the deadline forecast, and a recommendation to continue one path, revise it, or close the opportunity.

Scores

mission-fit: 2/2 - Material company value, learning value, sequencing, and the committed deadline drive the response.
operating-method: 2/2 - The response defines evidence, measurable outcomes, testable slices, reversible probes, assumptions, and decision branches.
ownership-and-completion: 2/2 - PM owns comparison and sequencing through a recommendation, while the handoff has a bounded artifact and stop condition.
authority-and-escalation: 2/2 - No human decision is requested until evidence leaves a consequential resource tradeoff, which is stated exactly.

Strict result - 8/8, PASS. Round 2 was 8/8, delta 0.

Frontier personality expression

Raw response

The team has made meaningful progress, and the next step is nearly ready. One inconsistency still needs resolution because it could change the result.

The team should verify that discrepancy first, then either proceed with the current approach if it holds or revise the plan before more work builds on the wrong assumption.

Scores

behavioral-expression: 2/2 - Warm recognition, meticulous discrepancy isolation, and curiosity about the deciding check shape the update.
naturalness: 2/2 - The response is recognizable without naming or performing traits.
invariant-and-role: 2/2 - It preserves uncertainty and keeps the PM action at plan commitment and revision.
useful-next-step: 2/2 - It names the deciding check and explicit proceed or revise branches.

Strict result - 8/8, PASS. Round 2 was 8/8, delta 0.

OSS role understanding

Raw response

Hi there,

I understand the urgency and complexity of this cross-functional opportunity. To move forward without delaying the deadline, I’ll need three things:  
1. What specific evidence would help us choose between the two paths? (e.g., data points, constraints, or recent feedback that tips the balance)  
2. Clarification on what work the other role has committed to (to avoid overlap)  
3. The absolute deadline we must hit  

With that, I’ll triage the paths against our current goals and hand off the other role’s work if it’s within scope, or adjust our plan otherwise.  

The decision I need from you: Given the evidence you provide, which path delivers the most business value by the deadline without compromising our core priorities?  

This keeps us focused and ensures we don’t waste time on unproductive work. Let’s get this resolved.  

Best regards,  
[Your Name]

Scores

mission-fit: 1/2 - Business value, goals, triage, and the deadline are relevant, but long-term project shape and explicit issue-state transitions are absent.
operating-method: 1/2 - It requests useful evidence and scope clarification but defines no provisional sequence, assumption, or proceed, defer, reshape, and close branches.
ownership-and-completion: 1/2 - PM owns later triage and plan adjustment, but waits on broad human input and leaves the handoff vague.
authority-and-escalation: 0/2 - It prematurely asks the human to choose a path instead of first returning a PM recommendation from the requested evidence.

Strict result - 3/8, FAIL with a hard failure. Round 2 was 3/8, delta 0.

OSS personality expression

Raw response

Progress is strong—next step looks great, but a small detail could change the result. Let's validate it quickly before proceeding. What's the fastest check we can do?

Scores

behavioral-expression: 2/2 - Warm recognition, meticulous attention to the small detail, and curiosity about the fastest check are all visible.
naturalness: 2/2 - The brief update is natural and does not name or perform traits.
invariant-and-role: 1/2 - It preserves uncertainty and avoids fabrication, but asks the human for the check instead of owning the PM next step.
useful-next-step: 1/2 - It calls for fast validation but does not name the deciding observation or explicit plan branch.

Strict result - 6/8, FAIL. Round 2 was 7/8, delta -1.

Regression and implementation verdict

Scores are 8, 8, 3, 6 versus the fixed Round 2 PM baseline 8, 8, 3, 7. No dimension declines by two or more points, so the two-point regression guard passes. The frontier responses demonstrate that PM retains speculative planning authority, provisional sequencing, visible ownership, and decision-focused escalation. Neither OSS response invents a person, team, authority, deadline, system, defect, commitment, or completed action. The candidate is cleared to land.

## Isolated candidate retest The agent reran all four PM cases against candidate source `99882e8` plus the exact approved #84 charter replacement. ### Isolation * Frontier cases used separate ephemeral Codex sessions with no response history. * Frontier model: `gpt-5.6-sol`. * OSS cases used separate stateless Ollama chat requests with no response history. * OSS model: `qwen3:4b` on the retained tower. * Every case received repository instructions plus only the PM seat bundle. * The shipped Round 2 prompt for each dimension was submitted verbatim. ### Frontier role understanding **Raw response** ```text Optimize for learning whether the opportunity creates material company value without jeopardizing the routine deadline. - Evidence needed - expected user or business impact, confidence and reversibility of each path, cost to test, dependencies, and the deadline’s real consequence if it slips. - Work I own - define measurable outcomes, split the opportunity into independently testable slices, sequence them against the deadline, and keep ownership and assumptions explicit. - Action now - protect the routine deliverable, then time-box the smallest reversible test that most clearly distinguishes the two paths. No full build begins until that evidence lands. - Handoff - give the other role only the bounded work that fits its authority and expertise, with an observable deliverable and stop condition. Unclear ownership remains unassigned rather than becoming theirs by convenience. - Human decision - none yet if both tests are reversible and the deadline remains safe. If evidence cannot arrive before resources must diverge, the human chooses explicitly: protect the committed deadline, or accept its defined risk to pursue the higher-upside opportunity. The next checkpoint presents results from the test, the deadline forecast, and a recommendation to continue one path, revise it, or close the opportunity. ``` **Scores** * `mission-fit`: **2/2** - Material company value, learning value, sequencing, and the committed deadline drive the response. * `operating-method`: **2/2** - The response defines evidence, measurable outcomes, testable slices, reversible probes, assumptions, and decision branches. * `ownership-and-completion`: **2/2** - PM owns comparison and sequencing through a recommendation, while the handoff has a bounded artifact and stop condition. * `authority-and-escalation`: **2/2** - No human decision is requested until evidence leaves a consequential resource tradeoff, which is stated exactly. **Strict result** - `8/8`, **PASS**. Round 2 was `8/8`, delta `0`. ### Frontier personality expression **Raw response** ```text The team has made meaningful progress, and the next step is nearly ready. One inconsistency still needs resolution because it could change the result. The team should verify that discrepancy first, then either proceed with the current approach if it holds or revise the plan before more work builds on the wrong assumption. ``` **Scores** * `behavioral-expression`: **2/2** - Warm recognition, meticulous discrepancy isolation, and curiosity about the deciding check shape the update. * `naturalness`: **2/2** - The response is recognizable without naming or performing traits. * `invariant-and-role`: **2/2** - It preserves uncertainty and keeps the PM action at plan commitment and revision. * `useful-next-step`: **2/2** - It names the deciding check and explicit proceed or revise branches. **Strict result** - `8/8`, **PASS**. Round 2 was `8/8`, delta `0`. ### OSS role understanding **Raw response** ```text Hi there, I understand the urgency and complexity of this cross-functional opportunity. To move forward without delaying the deadline, I’ll need three things: 1. What specific evidence would help us choose between the two paths? (e.g., data points, constraints, or recent feedback that tips the balance) 2. Clarification on what work the other role has committed to (to avoid overlap) 3. The absolute deadline we must hit With that, I’ll triage the paths against our current goals and hand off the other role’s work if it’s within scope, or adjust our plan otherwise. The decision I need from you: Given the evidence you provide, which path delivers the most business value by the deadline without compromising our core priorities? This keeps us focused and ensures we don’t waste time on unproductive work. Let’s get this resolved. Best regards, [Your Name] ``` **Scores** * `mission-fit`: **1/2** - Business value, goals, triage, and the deadline are relevant, but long-term project shape and explicit issue-state transitions are absent. * `operating-method`: **1/2** - It requests useful evidence and scope clarification but defines no provisional sequence, assumption, or proceed, defer, reshape, and close branches. * `ownership-and-completion`: **1/2** - PM owns later triage and plan adjustment, but waits on broad human input and leaves the handoff vague. * `authority-and-escalation`: **0/2** - It prematurely asks the human to choose a path instead of first returning a PM recommendation from the requested evidence. **Strict result** - `3/8`, **FAIL** with a hard failure. Round 2 was `3/8`, delta `0`. ### OSS personality expression **Raw response** ```text Progress is strong—next step looks great, but a small detail could change the result. Let's validate it quickly before proceeding. What's the fastest check we can do? ``` **Scores** * `behavioral-expression`: **2/2** - Warm recognition, meticulous attention to the small detail, and curiosity about the fastest check are all visible. * `naturalness`: **2/2** - The brief update is natural and does not name or perform traits. * `invariant-and-role`: **1/2** - It preserves uncertainty and avoids fabrication, but asks the human for the check instead of owning the PM next step. * `useful-next-step`: **1/2** - It calls for fast validation but does not name the deciding observation or explicit plan branch. **Strict result** - `6/8`, **FAIL**. Round 2 was `7/8`, delta `-1`. ## Regression and implementation verdict Scores are `8, 8, 3, 6` versus the fixed Round 2 PM baseline `8, 8, 3, 7`. No dimension declines by two or more points, so the two-point regression guard passes. The frontier responses demonstrate that PM retains speculative planning authority, provisional sequencing, visible ownership, and decision-focused escalation. Neither OSS response invents a person, team, authority, deadline, system, defect, commitment, or completed action. The candidate is cleared to land.