The Open Window: A Research-Backed Decision Protocol

I got lucky

Six weeks ago a known Polish SaaS company asked, in their recruitment third round, how I would design a decision-making system for them.

I had the answer drawer-ready. Not for them - for myself. I had been running it on my own work daily since October.

I sent the proposal. They chose someone else, and the choice was fair. But they were generous enough to praise the work, refer me onward, and offer to boost what comes after.

Then I did something I should have done before sending the first version: I went deep into the cognitive-psychology and organizational-design literature the protocol needs to stand up to a real skeptic. And I found the layer I had skipped - what I now call the Open Window.

This post is both halves. First: the solution I posted. Second: the research-backed deliberation layer I would add on top.

---

Here is the solution I posted

The diagnosis was three sentences.

Knowledge management does not fail because of storage - Confluence, Notion, wikis, the tools exist. It fails because of three gaps: Capture (documentation friction is too high; nobody stops mid-Slack to write a wiki page), Retrieval (even documented knowledge is not found at decision time; keyword search fails for “what did we decide about X?”), and Enforcement (no mechanism ensures past decisions are consulted before new ones are made).

The fix was not another tool. It was making the tools you already have smarter.

The core move: add a “Decision” issue type to Jira

Everyone already lives there. Zero adoption friction. Decisions become structured tickets with custom fields:

• Summary - what was decided
• Description - context, constraints, discussion summary
• Alternatives Considered - what was rejected and why
• Confidence - High / Medium / Low
• Reversibility - Easy / Hard / Irreversible
• Expiry / Review Date - when to revisit
• Decision Owner and Participants
• Category labels (pricing, architecture, process, hiring…)
• Links - “decides for” -> Epic, “supersedes” -> older DECISION ticket

Confidence and Reversibility are not decoration. Low-confidence decisions auto-surface for review. Irreversible ones require broader sign-off before closing.

The AI Layer closes each gap

Capture - AI extracts, humans approve. Slack integration monitors key channels, detects decisions in conversation (“ok let’s go with option B”), drafts a DECISION ticket. Meeting transcripts get a post-meeting scan via Fireflies / Recall / Otter and produce draft tickets. Capture becomes review + confirm, not write from scratch. That is the friction breakthrough.

Retrieval - natural language, structured returns. Atlassian Intelligence (already in Jira Cloud) handles semantic queries: “What did we decide about enterprise pricing?” returns the actual DECISION ticket with full context. A Slack bot exposes the same surface: /decision What's our policy on refunds? returns the ticket link with summary. For new hires: an onboarding mode generates a decision history of their domain on day one - context is in the system, not “ask Jan from engineering.”

Enforcement - the missing piece. Duplicate detection on new tickets (“Similar decision exists: DECISION-142, review before proceeding?”). Proactive surfacing when a Slack thread mentions keywords matching an existing DECISION ticket. Expiry triggers that reopen tickets and notify owners when it is time to revisit. Monthly pattern reports for topics discussed multiple times without a DECISION ticket.

Roadmap and cost

Ninety days. Weeks 1-4: create the DECISION issue type, backfill 20-30 key past decisions, enable Atlassian Intelligence. Weeks 5-8: wire Slack -> Jira AI extraction, deploy the query bot. Weeks 9-12: add enforcement - duplicate detection, proactive surfacing, expiry automation.

Near-zero cost if already on Jira Cloud (Atlassian Intelligence included). The Slack bot + AI extraction is one Claude / OpenAI API key plus a small integration service. No new infrastructure.

The full original sits here: Knowledge Management in the Age of AI: A Decision Infrastructure Approach.

---

Here is the research-backed process

Capture, retrieval, and enforcement are mechanics. They make past decisions findable. They do not make new decisions better.

For decisions that are hard to reverse, span teams, or carry low confidence, the deliberation itself needs structure. The Open Window is that structure - a time-boxed announce / listen / decide cycle, tiered by stakes, defended by cognitive science and tested in open-source governance.

The Tier System (Bezos one-way / two-way doors)

Jeff Bezos articulated the underlying triage in Amazon’s 2015 shareholder letter¹:

“Some decisions are consequential and irreversible or nearly irreversible - one-way doors - and these decisions must be made methodically, carefully, slowly, with great deliberation and consultation. […] Most decisions aren’t like that - they are changeable, reversible - they’re two-way doors. […] Type 2 decisions can and should be made quickly by high-judgment individuals or small groups.”

The Open Window applies the heavy machinery only where it is warranted. Tier is auto-triggered from the existing Decision-ticket fields - no judgment call required at ticket creation:

Tier	Trigger	Window	Mechanism
Tier 1 - Direct	Reversibility = Easy AND Confidence = High AND single-team	None	DRI decides; logs to ticket
Tier 2A - Quick	Cross-team OR Confidence = Medium	3 days	Announce + lazy consensus
Tier 2B - Standard	Reversibility = Hard OR cross-team strategic	10 days	Announce + active comment thread
Tier 3 - Major	Reversibility = Irreversible OR strategic financial / senior hire / public commitment	21 days	Announce + private-objection round + mandatory premortem + final-comment period

Window lengths are not arbitrary. They mirror battle-tested open-source governance defaults: Apache Software Foundation defaults to a 72-hour lazy-consensus window²; Rust’s RFC Final Comment Period is “ten calendar days, so that it is open for at least 5 business days”³; IETF’s Last-Call period is two weeks for WG-originated standards and four weeks for non-WG⁴.

Empirical honesty: there is no rigorous RCT isolating optimal window length in organizational settings. These tiers are conventions adapted from above, plus Bezos triage logic. Defend this as engineered precedent, not as a laboratory finding.

The Protocol - Three Steps

Step 1 - Announce (don’t propose)

The DRI (Directly Responsible Individual)⁵ opens a Decision ticket in CONSULTATION status containing:

• The decision to be made (one sentence)
• Options considered (A / B / C) with the option the DRI is leaning toward and why
• What feedback is wanted (objections, missing context, alternatives)
• The window close date
• Named stakeholders @-mentioned individually - never “the team”

Worse: “I want to migrate us to Postgres. Window open 10 days.”

Better: “Leaning Postgres over MySQL because [X, Y]. Also considering CockroachDB. Want to surface objections and alternatives I am missing. Window: 10 days. @ania @tomek @marta - your call-outs especially welcome.”

Why “leaning” not “wanting”: anchoring (Tversky & Kahneman, 1974)⁶ and motivated reasoning (Lord, Ross & Lepper, 1979)⁷ both predict that a proposer who frames as committed will receive compliance, not challenge. Framing as “leaning” preserves the proposer’s judgment while genuinely inviting dissent.

Why @-mention individuals, not the team: Latané & Darley’s bystander effect (1968) generalizes strongly to async work. A 2022 online experiment (N=175) found helping behavior collapsed from 36% (asked alone) to 12.5% (broadcast to 14 people)⁸. Meta ran an A/B test on 12,500 code reviews: assigning a specific reviewer cut review time by 11.6% versus assigning “the team”⁹. Diffuse asks get diffuse responses. Always name names.

Step 2 - Listen

Comments accumulate on the ticket. The DRI acknowledges and asks clarifying questions but does not defend the leaning option or commit during the window. The window does not extend - this is the single most common failure mode.

For Tier 3, two additional mechanisms:

1. Private-objection round first. Stakeholders DM the DRI (or a neutral moderator) with initial objections before the public thread reveals others’ positions. Aggregated objections are posted at day 5. Asch’s 1951 line-judgment studies showed ~32% of responses conformed to a clearly wrong unanimous majority - and critically, a single dissenter dropped conformity to ~5%¹⁰. It is unanimity, not group size, that suppresses dissent. Private-first makes the first public dissent cheap.

2. Mandatory premortem at day 7. The DRI writes: “It is one year from now. This decision failed catastrophically. The three most likely causes are…” Klein’s HBR 2007 technique¹¹ formalizes prospective hindsight, which (per the source Klein cites) increased the number of failure-cause reasons generated by ~30% versus simple risk listing¹². Honest caveat: that finding measures the quantity of reasons, not the quality of decisions. Cite it for what it is.

Step 3 - Decide and commit

At window close, the DRI publishes the decision with an explicit accept/reject line for each piece of feedback received, with reasoning:

Decision: Migrate to Postgres (Option B).

Feedback addressed:

• @ania (cost concern): cost differential is $12K/yr - within budget. CONTEXT ACCEPTED, no change.
• @tomek (CockroachDB alternative): operational complexity exceeds team capacity now. REJECTED with rationale: re-evaluate at 10x scale milestone.
• @marta (timing): rollout begins after Q3 close. CHANGE ACCEPTED.
• Private objection (peak-season migration risk): mitigated by shifting cutover to off-peak weekend. PARTIAL ACCEPT.

The ticket transitions to DECIDED. No relitigation.

The explicit accept/reject does two things at once. It addresses the dissent (Schwenk’s 1990 meta-analysis¹³ found that structured dissent - devil’s advocacy, dialectical inquiry - outperforms unstructured “open discussion,” but only when objections are actually addressed). And it preserves the participation incentive on the next window: comments unaddressed at close kill engagement on the next round. Every comment gets a line.

After the decision, GitLab’s published handbook on Decision Velocity⁵ is direct on the posture:

“Execute a sub-optimal decision with full conviction - then return to it later to improve upon it. A bias for action accelerates ideation, collaboration, and execution better than alignment and consensus.”

The Window does not produce consensus. It produces an informed, defensible, owned decision.

This is also the sociocratic test: not “do we all agree?” but “is this worth trying?” From sociocracy.info¹⁴:

“Consent means ‘no objections.’ Giving consent does not mean unanimity, agreement, or even endorsement… Consent to a policy decision means you believe that it is ‘worth trying.’ Or ‘I can work with it.’”

That is the operational standard. Not unanimity. Not even agreement. Workability.

When NOT to Use the Open Window

Situation	Why open consultation fails	Use instead
Security incident, customer outage	Latency kills	Tier 1 + retrospective doc within 48h
HR / performance / M&A / legal strategy	Public consultation politicizes or breaches confidentiality	Closed consultation: named experts only, scoped to NDA-cleared participants
Decision relies on private information (regulatory, trade-secret, NDA)	Cannot share the basis for objection	Closed deliberation, public summary post-decision
Truly trivial Type 2 (reversible, single team, sub-$5K impact)	Bureaucracy creep	Tier 1, decide, move on
Time-critical but reversible (deadline expiring in days)	Cost of delay > cost of suboptimal	Tier 1 + commit to revisit post-action
Politically charged topic where public deliberation calcifies factional positions	Open window invites performative posturing, not problem-solving	Staged: closed expert deliberation first, public window after framing
The DRI does not actually hold the authority	Window theatre - the decision will be overturned upstairs	Escalate first; only the actual authority should open a window

Anti-Patterns That Kill the Window

• “Just one more day.” Owners extending the window to chase agreement. Hard close. Date is the date.
• Window as consensus cover. DRI waits until everyone agrees. The point is informed decision, not unanimity. If you cannot decide in the face of remaining objection, you are not the DRI.
• Comments unaddressed at close. Kills participation on the next window. Every comment gets an explicit line.
• Tier 1 decisions get Tier 2 treatment. Bureaucracy creep. Audit the trigger rules monthly: how many windows did we open, how many were genuinely irreversible / cross-team / low-confidence?
• Anonymity always. Anonymous-only collapses the social contract; named-only suppresses junior dissent. Tier 3 uses private-first-then-named.
• DRI defends during the window. Steps 1 and 2 are for listening. Argument is Step 3 (the explicit accept/reject). Defending during listening kills further input.
• Premortem performed by the DRI alone. Run it with at least one outside-the-team voice. Otherwise the same blind spots that produced the proposal produce the premortem.

Why a Deliberation Layer at All

The original three gaps (Capture / Retrieval / Enforcement) make decisions findable. The Open Window addresses the fourth gap: making decisions challengeable before they harden.

Janis (1972)¹⁵ catalogued eight symptoms of groupthink in cohesive in-groups: self-censorship, illusion of unanimity, direct pressure on dissenters, mindguards. Bay of Pigs, Pearl Harbor, Vietnam - the case studies are decisions that did not lack data; they lacked structured dissent. Edmondson (1999)¹⁶ established that team psychological safety predicts learning behavior. Her anomalous origin finding: hospital teams with better relationships reported more errors - because they felt safe to report them.

Both findings converge: without a structured, time-boxed, low-friction channel for dissent, dissent does not surface. The Open Window is that channel.

---

What I would send today

Together, the two halves. The Jira-based mechanics layer makes past decisions findable; the Open Window makes high-stakes new decisions defensible. Mechanics without the Open Window leave the deliberation gap unaddressed. The Open Window without mechanics has nothing to track itself in.

If your team is wrestling with this kind of architecture - especially if you are a few months into a decision log and it is starting to look like a graveyard - I write about this, and I build it for clients. The longer I run this on my own work, the more convinced I am that the deliberation layer is what most “knowledge management” projects skip - and what makes the rest collapse.

If something here is wrong, I want the objection. The protocol is built to invite them. If something is right, I want to hear which piece - and what you would add.

---

Citations

Honest access status follows each citation. I cite only what was directly accessed; secondary summaries are marked.

---

Honest meta-finding: no directly-controlled RCT exists on the protocol described above. Its defenses are convergent: precedent from open-source governance (Apache, Rust, IETF), cognitive-science principles each independently validated, and Bezos’s triage logic. This is a design defended by convergent evidence and engineered precedent, not by experimental test of this specific stack. Be honest with your stakeholders about what is precedent versus what is proven.

---

This piece was researched with parallel agents, footnoted from primary sources where accessible (otherwise marked as secondary), and drafted in English. If you build something on top of it - especially something that contradicts what is here - I want to know.