Add role-based future directions agenda

FUTURE-DIRECTIONS.md

@@ -0,0 +1,73 @@
+# Future Directions
+
+Loop Engineering is still an emerging practice. This agenda turns its largest open gaps into falsifiable research questions, buildable infrastructure, and testable product work. It is a guide to useful next contributions, not a prediction that every direction will succeed.
+
+Prioritize work that makes at least one recurring agent loop **more measurable, portable, recoverable, economical, or governable**. A credible result should expose the trigger, permissions, external verification gate, durable state, budget, escalation path, and failure evidence.
+
+## For Researchers
+
+| Category | High-value question | Evidence of progress | Concrete starting point |
+| --- | --- | --- | --- |
+| **Factorized evaluation** | How much performance comes from the model, harness, verifier, and outer loop separately? | Matched model and token budgets, controlled component ablations, confidence intervals, and held-out tasks | Run one fixed model across several harness and loop policies while scoring outcome, recovery, abstention, and resource use |
+| **Verification science** | Which combinations of deterministic checks, learned judges, independent agents, and humans prevent false completion? | False-accept and false-reject rates, calibration, tamper resistance, inter-rater agreement, and adversarial challenge sets | Compare a test suite, single judge, judge ensemble, and human review on the same traces with a frozen acceptance policy |
+| **State and learning across runs** | Which state representations support cold resume and improvement without stale-memory drift or compaction errors? | Longitudinal replay, state corruption tests, retention curves, recovery accuracy, state size, and maintenance cost | Build tasks that inject missing, stale, conflicting, or poisoned state between runs and measure recovery |
+| **Control, stopping, and budgets** | When should a loop retry, change strategy, escalate, abstain, or stop under uncertainty? | Quality-cost-latency Pareto curves, regret, budget adherence, failure containment, and sensitivity analysis | Compare fixed retries, evidence-aware stopping, and learned escalation under identical budgets |
+| **Human oversight** | Which handoff designs help people intervene accurately without review fatigue or automation bias? | Preregistered user studies, intervention precision, time to decision, missed failures, workload, and trust calibration | Compare raw traces, concise summaries, and structured evidence receipts on the same review decisions |
+
+Research reports should publish negative results and separate task success from process discipline. A loop that reaches the answer through uncontrolled retries, hidden state, or verifier leakage is not equivalent to one that succeeds within its declared contract.
+
+## For Infrastructure And Reliability Engineers
+
+| Category | Engineering objective | Evidence of progress | Concrete starting point |
+| --- | --- | --- | --- |
+| **Durable execution** | Resume safely after crashes, rate limits, deploys, and partial work | Idempotency tests, deterministic replay, duplicate-work rate, recovery time, and chaos results | Implement checkpointed queue processing with stable item IDs, bounded leases, and externally verified completion |
+| **Receipts and observability** | Make every decision reconstructable from evidence rather than model narration | Correlated traces, immutable receipts, state diffs, cost attribution, and replay tooling | Define a receipt envelope for trigger, input version, action, verifier result, state transition, budget, and next action |
+| **Permissions and isolation** | Enforce least privilege across tools, files, networks, credentials, and approval boundaries | Denied-action tests, capability manifests, sandbox escape tests, secret-exposure checks, and audited human overrides | Translate each Loop Contract permission into runtime policy and test both allowed and forbidden actions |
+| **Portability and interoperability** | Move one contract between session, CI, scheduled, and durable runtimes without changing its semantics | Conformance tests, adapter compatibility, migration fixtures, and equivalent receipts across runtimes | Build adapters for one contract on a local schedule, GitHub Actions, and a recoverable worker |
+| **Reliability economics** | Allocate models, tools, retries, and reviewers to maximize verified value within budget | Cost per verified outcome, latency percentiles, retry amplification, queue stability, and quality-budget frontiers | Add policy-based routing and compare it with fixed-model baselines on a replayable workload |
+
+Operational maturity should be demonstrated with failure injection, not only happy-path demos. Useful artifacts include runbooks, service-level objectives, incident reviews, migration notes, and conformance suites.
+
+## For Application And Product Developers
+
+| Category | Product objective | Evidence of progress | Concrete starting point |
+| --- | --- | --- | --- |
+| **Loop-worthy use cases** | Select recurring work with stable intake and a defensible external definition of done | Baseline human effort, recurrence rate, verifier coverage, exception rate, and avoided rework | Start with a narrow queue such as failing checks, stale documents, support triage, or threshold monitoring |
+| **Domain verification** | Encode completion in domain evidence rather than a generic model confidence score | Golden cases, contract tests, reviewer agreement, regression suites, and unresolved-edge-case logs | Pair every action path with a domain-specific checker and keep the agent unable to edit that checker |
+| **Human handoff experience** | Show progress, uncertainty, evidence, and safe next actions at the moment judgment is needed | Time to understand, correct intervention rate, override quality, notification burden, and user-reported clarity | Design one escalation view around objective, attempted actions, receipts, remaining budget, and recommended choices |
+| **Progressive rollout** | Move from observation to bounded production use without hiding risk | Shadow-mode precision, canary outcomes, rollback rate, budget breaches, and post-launch incidents | Roll out through read-only, suggestion, approval-required, and bounded-autonomy stages with explicit promotion gates |
+| **Maintainable adoption** | Keep loop behavior understandable as models, tools, policies, and teams change | Contract review time, configuration drift, ownership coverage, change failure rate, and onboarding success | Store versioned contracts beside code, assign an owner, and test them whenever prompts, tools, or verifiers change |
+
+Strong developer contributions are complete vertical slices: a real trigger, bounded permissions, independent gate, durable receipts, escalation UX, and a before/after operating result.
+
+## Shared Infrastructure Priorities
+
+| Priority | Shared asset | Minimum useful outcome |
+| --- | --- | --- |
+| **Portable Loop Contract profile** | A small interoperable subset of objective, trigger, permissions, verification, state, budget, escalation, and exit fields | Two runtimes execute the same contract and produce semantically equivalent decisions |
+| **Receipt and replay format** | A stable event envelope for reconstructing each state transition and verifier decision | A third party can replay one run and identify why it retried, escalated, or exited |
+| **Factorized benchmark harness** | Evaluation that varies model, harness, verifier, and loop policy independently | Results attribute gains and costs to the component that caused them |
+| **Failure and safety challenge set** | Reusable cases for infinite retries, stale state, judge bias, prompt injection, permission escape, and budget exhaustion | Implementations report comparable containment and recovery metrics |
+| **Case-study commons** | Public or safely anonymized operating reports with contracts, receipts, budgets, and lessons | Builders can compare real failure modes and operating economics across domains |
+
+## Three Starter Projects
+
+1. **Research:** create a matched-budget benchmark that compares fixed retry, evidence-aware retry, and human escalation while publishing every trace and negative result.
+2. **Engineering:** implement the same validated contract in a local scheduler, CI workflow, and durable worker, then publish a conformance and crash-recovery report.
+3. **Development:** ship one read-only production loop with a measurable baseline, independent checker, evidence-rich handoff, and staged rollout report.
+
+## Qualification Checklist
+
+Before proposing a new direction, record:
+
+- the recurring task and user or operator it serves;
+- the baseline and falsifiable claim;
+- the external evidence gate and who controls it;
+- the durable state and receipt format;
+- the time, cost, retry, and concurrency budgets;
+- the failure injection or adversarial cases;
+- the human escalation and rollback path;
+- the reproducibility, privacy, and licensing constraints;
+- the negative result that would cause the approach to be rejected.
+
+Use [GitHub Discussions](https://github.com/ChaoYue0307/awesome-loop-engineering/discussions) to propose a study, reference implementation, or case study. Use the [gallery checklist](gallery/README.md#minimum-useful-case-study) when reporting a loop that has actually run.

README.md

@@ -102,6 +102,18 @@ Use `url` as the durable join key. `row_id` and `source_line` are positional and
 
 Do not use `signal_strength`, GitHub stars, forks, or inclusion in this collection as a quality label, endorsement, or automated ranking of scientific validity.
 
+## Future Directions
+
+The project maintains a role-based agenda for moving Loop Engineering forward:
+
+| Audience | Priority categories |
+| --- | --- |
+| Researchers | Factorized evaluation, verification science, state across runs, stopping policies, and human oversight |
+| Infrastructure and reliability engineers | Durable execution, receipts and replay, least privilege, portability, and reliability economics |
+| Application and product developers | Loop-worthy use cases, domain verification, escalation UX, progressive rollout, and maintainable adoption |
+
+The complete [Future Directions agenda](https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/FUTURE-DIRECTIONS.md) defines measurable outputs, concrete starter projects, shared infrastructure priorities, and a qualification checklist.
+
 ## Dataset Structure
 
 The primary configuration is `resources`, with one `train` split backed by `data/resources.jsonl`. `data/resources.csv` contains the same rows for spreadsheet and dataframe workflows.

ROADMAP.md

@@ -2,18 +2,20 @@
 
 The roadmap prioritizes stronger evidence, runnable implementations, and real operating results.
 
+The role-based [Future Directions agenda](FUTURE-DIRECTIONS.md) complements this project roadmap with research questions, infrastructure priorities, application opportunities, measurable outputs, and starter projects.
+
 ## Near Term
 
 - Collect more direct Loop Engineering sources as the term stabilizes.
 - Add real or anonymized gallery entries from practitioners running recurring agent loops.
-- Grow the runnable loop directory beyond the test-repair reference loop, including scheduled-trigger variants per runtime.
+- Turn more validated contracts into runtime-specific executables with crash recovery, timeouts, replayable receipts, and failure-injection tests.
 - Add more translations for the introduction, mental model, Loop Contract, and contribution guide.
 - Audit contextual sources in small batches; replace weak summaries and secondary links with canonical evidence.
 - Replace unstable links with primary sources, official docs, papers, or implementation-heavy write-ups.
 
 ## Pattern Library
 
-The library now contains 15 reference patterns: PR babysitting, CI repair, docs drift, deploy verification, feedback clustering, dependency triage, evaluation regression, security review, cost control, bug hunting, enterprise approval, incident response, data quality, release notes, and model routing. Every pattern ships a schema-validated loop contract in `examples/`.
+The library now contains 20 reference patterns: PR babysitting, CI repair, docs drift, deploy verification, feedback clustering, dependency triage, evaluation regression, security review, cost control, bug hunting, enterprise approval, incident response, data quality, release notes, model routing, benchmark optimization, knowledge freshness, performance regression, accessibility regression, and adversarial red teaming. Every pattern ships a schema-validated loop contract in `examples/`.
 
 Next pattern-library work should prioritize variants backed by operational evidence rather than adding names for coverage. Useful additions include runtime-specific implementations, before/after receipts, measured retry and cost budgets, failure cases, and human-escalation outcomes.
 
@@ -45,7 +47,8 @@ The gallery should grow from reference examples into public or anonymized case s
 
 ## Open Questions
 
-- Which loop primitives become common across Codex, Claude Code, GitHub Agentic Workflows, and custom runtimes?
-- What is the right schema shape for portable loop contracts?
-- Which verification gates are strong enough for unattended or semi-attended loops?
-- How should maintainers evaluate submitted real-world loop examples without exposing private data?
+- Which Loop Contract fields preserve their semantics across session, CI, scheduled, and durable runtimes?
+- How should benchmarks separate model, harness, verifier, and outer-loop effects under matched budgets?
+- Which verification gates remain calibrated under distribution shift and adversarial pressure?
+- Which receipt format is sufficient for replay, incident review, and human escalation without retaining sensitive data?
+- How should maintainers evaluate real-world loop examples without exposing private inputs, state, or traces?

design-qa.md

@@ -14,7 +14,7 @@
 
 - Desktop first viewport: 1440 x 1000, page top, sticky navigation and the complete six-station agent loop visible.
 - Mobile first viewport: 390 x 844, page top, compact six-station scene, stage selector, and next-section cue visible without horizontal overflow.
-- Mobile navigation: eight links visible when expanded; Escape closes the menu, updates `aria-expanded`, and restores focus.
+- Mobile navigation: nine links visible when expanded; Escape closes the menu, updates `aria-expanded`, and restores focus.
 - Pattern library: 20 symptom-first entries; mobile query `accessibility` returns exactly one matching pattern.
 - Contract and lifecycle: purpose-built mobile summaries replace desktop-wide figures at 560 px and below.
 - Runtime map: desktop figure plus eight labeled starters in a balanced 4 x 2 grid; mobile uses the explanatory copy and one-column starter sequence instead of shrinking the wide figure.
@@ -24,6 +24,7 @@
 - Interactive hero, mobile: 390 x 844, one ready WebGL canvas, no horizontal overflow, and all six stations inside the frame.
 - Interactive hero motion: the frame counter advanced by 47 frames during a 450 ms pointer-parallax check.
 - Interactive hero efficiency: the frame counter advanced by zero during a 700 ms offscreen check at the Resources section.
+- Future directions: three role-based tracks remain side by side at 1440 x 1000 and become a rule-separated single column at 390 x 844 without horizontal overflow.
 
 ## Comparison History
 
@@ -39,6 +40,7 @@
 10. Verified progressive enhancement behavior: the semantic hero remains complete without WebGL, forced-colors and print hide the scene, reduced-motion renders a static frame, and the animation stops when the hero leaves the viewport.
 11. Expanded the implementation kit to 20 patterns, 20 validated contracts, and eight runtime starters. The runtime section now labels all five adaptable templates and all three executables, with use-case copy and an even desktop grid.
 12. Rechecked the expanded pattern and runtime sections at 1440 x 1000 and 390 x 844. Anchor offsets clear the sticky header, the page width stays equal to the viewport, and headings, icons, labels, and descriptions remain aligned.
+13. Added the role-based future-directions section and rechecked it at both target viewports. Resource Atlas expansion initially displaced direct links to later sections; post-load hash restoration now leaves `#future` 84 px below the desktop viewport top and 74 px below it on mobile.
 
 ## Findings
 
@@ -57,21 +59,24 @@
 - Mobile pattern library: `/private/tmp/ale-site-v07-patterns-mobile.png`
 - Desktop runtime map and starter grid: `/private/tmp/ale-site-v07-runtime-desktop-final.png`
 - Mobile runtime starters: `/private/tmp/ale-site-v07-runtime-mobile-final.png`
+- Desktop future directions: `/private/tmp/future-viewport-desktop.png`
+- Mobile future directions: `/private/tmp/future-viewport-mobile.png`
 - Three.js pointer-motion frame: `/private/tmp/awesome-loop-hero-three-motion-final.png`
 - Social preview: `/Users/chaoyue/Library/CloudStorage/Dropbox/Loop Engineering/awesome-loop-engineering/assets/social-preview.png`
 
 ## 2026-07-18 Repository Audit
 
-- Desktop: 1440 x 1000, no horizontal overflow, no hero text/scene overlap, and all eight section headings render without requiring scroll-triggered opacity changes.
+- Desktop: 1440 x 1000, no horizontal overflow, no hero text/scene overlap, and all nine section headings render without requiring scroll-triggered opacity changes.
 - Tablet: 768 x 900. The first pass clipped the intake and decision stations. The responsive camera now fits an 11.2-unit horizontal view between 720 px and 900 px, keeping all six stations, the retry path, and the outcome node inside the canvas.
 - Phone: 390 x 844, no horizontal overflow. The compact loop keeps all six stations, the moving work packet, legend, outcome key, and stage selector visible without crossing the hero copy.
-- Mobile navigation: opens all eight project links, reports `aria-expanded="true"`, closes on Escape, and restores focus to the menu button.
+- Mobile navigation: opens all nine project links, reports `aria-expanded="true"`, closes on Escape, and restores focus to the menu button.
 - Pattern filter: `accessibility` returns exactly one of the 20 patterns.
 - Resource Atlas: remains idle above the fold, loads on approach or a direct resource anchor, and reports `545 of 545 resources | showing 8` on mobile.
 - Resource interactions: `SWE-bench` returns 9 matches, the paper filter narrows those to 6, Reset restores all 545, and Show more expands the rendered set from 8 to 16.
 - Scene motion: the canvas frame counter advanced from 640 to 662 over 900 ms. Two canvas crops taken 1.2 seconds apart changed 17.79% of sampled RGB channels.
 - Scene pixels: the sampled desktop crop has 15.71% active pixels with RGB standard deviations of 35.83, 29.77, and 23.01; mobile has 27.84% active pixels with deviations of 41.27, 30.61, and 19.52. Both renders are nonblank and well contrasted.
 - Runtime starters: desktop cards form two equal-height rows of four; mobile cards are 358 px wide within a 390 px document and preserve label, icon, heading, and copy alignment.
+- Future directions: direct-anchor restoration waits for the Resource Atlas reflow, all three audience tracks remain readable, and the agenda CTA is fully visible at desktop and mobile widths.
 - Console: no page-origin errors or warnings during mobile navigation, pattern filtering, and Resource Atlas interactions.
 - Source audit: 545 rows checked; 496 public sources reachable, 1 access-restricted, 48 repository-native, and 0 broken or unreachable as of 2026-07-17 UTC.
 - Data and copy: the website, README, translations, social preview source, release copy, and Hugging Face card all report 545 resources, 20 patterns, 20 contracts, 8 runtime starters, and version 0.7.0.

docs/index.html

@@ -479,6 +479,36 @@
         text-transform: uppercase;
       }
       .runtime-kind.executable { color: var(--green); }
+      .future-grid {
+        display: grid;
+        grid-template-columns: repeat(3, minmax(0, 1fr));
+        border-top: 1px solid var(--line-strong);
+        border-bottom: 1px solid var(--line-strong);
+      }
+      .future-track {
+        display: flex;
+        min-width: 0;
+        flex-direction: column;
+        padding: 24px 24px 24px 0;
+      }
+      .future-track + .future-track { border-left: 1px solid var(--line); padding-left: 24px; }
+      .future-track h3 { display: flex; align-items: center; gap: 9px; margin: 0; font-size: 1rem; }
+      .future-track h3 .icon { width: 18px; height: 18px; color: var(--blue); }
+      .future-track ul { margin: 16px 0 22px; padding-left: 19px; }
+      .future-track li { margin-top: 9px; color: var(--muted); font-size: 0.86rem; line-height: 1.5; }
+      .future-track li::marker { color: var(--blue); }
+      .future-output { margin: auto 0 0; border-top: 1px solid var(--line); padding-top: 14px; color: var(--quiet); font-size: 0.76rem; line-height: 1.5; }
+      .future-output strong { display: block; margin-bottom: 3px; color: var(--ink); }
+      .future-shared {
+        display: grid;
+        grid-template-columns: 180px minmax(0, 1fr) auto;
+        gap: 22px;
+        align-items: center;
+        border-bottom: 1px solid var(--line-strong);
+        padding: 20px 0;
+      }
+      .future-shared strong { font-size: 0.92rem; }
+      .future-shared p { margin: 0; color: var(--muted); font-size: 0.84rem; line-height: 1.55; }
       code {
         border: 1px solid var(--line);
         border-radius: 4px;
@@ -674,6 +704,10 @@
         .stat:nth-child(n+3) { border-top: 1px solid var(--line); }
         .stat:last-child { grid-column: 1 / -1; border-left: 0; }
         .principles, .link-grid, .pattern-grid { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+        .future-grid { grid-template-columns: 1fr; }
+        .future-track { padding: 20px 0; }
+        .future-track + .future-track { border-top: 1px solid var(--line); border-left: 0; padding-left: 0; }
+        .future-shared { grid-template-columns: 1fr; gap: 10px; align-items: start; }
         .level { grid-template-columns: 54px 1fr; }
         .level .detail { grid-column: 2; }
         .resource-toolbar { grid-template-columns: repeat(2, minmax(0, 1fr)); }
@@ -1057,6 +1091,7 @@
           <a href="#runtimes">Runtimes</a>
           <a href="#maturity">Maturity</a>
           <a href="#resources">Resources</a>
+          <a href="#future">Future</a>
           <a href="#faq">FAQ</a>
           <a class="mobile-nav-only" href="https://huggingface.co/datasets/cy0307/awesome-loop-engineering"><i class="icon" data-lucide="database" aria-hidden="true"></i>Dataset</a>
         </nav>
@@ -1417,11 +1452,57 @@
         </div>
       </section>
 
+      <section id="future">
+        <div class="wrap">
+          <div class="section-head">
+            <div>
+              <p class="section-label"><i class="icon" data-lucide="compass" aria-hidden="true"></i><span>07 / Future directions</span></p>
+              <h2>Decide what the field should prove next.</h2>
+            </div>
+            <p class="section-intro">A credible agenda separates claims to test, infrastructure to build, and product loops to validate. Each direction should end in external evidence, reproducible artifacts, or measured operating results.</p>
+          </div>
+          <div class="future-grid">
+            <article class="future-track">
+              <h3><i class="icon" data-lucide="book-open" aria-hidden="true"></i>For researchers</h3>
+              <ul>
+                <li>Separate model, harness, verifier, and loop effects under matched budgets.</li>
+                <li>Measure false completion, calibration, tamper resistance, and recovery.</li>
+                <li>Study state drift, stopping policies, and human oversight across runs.</li>
+              </ul>
+              <p class="future-output"><strong>Useful outputs</strong>Benchmarks, ablations, longitudinal datasets, preregistered studies, and negative results.</p>
+            </article>
+            <article class="future-track">
+              <h3><i class="icon" data-lucide="wrench" aria-hidden="true"></i>For infrastructure engineers</h3>
+              <ul>
+                <li>Build crash-safe, idempotent, replayable, and observable runtimes.</li>
+                <li>Turn contract permissions into enforceable least-privilege policy.</li>
+                <li>Standardize receipts and preserve semantics across runtime adapters.</li>
+              </ul>
+              <p class="future-output"><strong>Useful outputs</strong>Reference workers, conformance suites, receipt schemas, replay tools, SLOs, and incident reports.</p>
+            </article>
+            <article class="future-track">
+              <h3><i class="icon" data-lucide="code-2" aria-hidden="true"></i>For application developers</h3>
+              <ul>
+                <li>Choose recurring jobs with stable intake and a defensible definition of done.</li>
+                <li>Design domain checkers and evidence-rich human handoffs.</li>
+                <li>Progress from read-only shadowing to bounded autonomy with promotion gates.</li>
+              </ul>
+              <p class="future-output"><strong>Useful outputs</strong>Runnable vertical slices, domain evals, escalation UX, rollout studies, and failure write-ups.</p>
+            </article>
+          </div>
+          <div class="future-shared">
+            <strong>Shared milestones</strong>
+            <p>Portable contracts and receipts, factorized benchmarks, replayable state, failure-injection suites, and public or safely anonymized case studies.</p>
+            <a class="button" href="https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/FUTURE-DIRECTIONS.md"><i class="icon" data-lucide="compass" aria-hidden="true"></i>Open the agenda</a>
+          </div>
+        </div>
+      </section>
+
       <section id="contribute">
         <div class="wrap">
           <div class="section-head">
             <div>
-              <p class="section-label"><i class="icon" data-lucide="handshake" aria-hidden="true"></i><span>07 / Contribute</span></p>
+              <p class="section-label"><i class="icon" data-lucide="handshake" aria-hidden="true"></i><span>08 / Contribute</span></p>
               <h2>Improve the evidence or share a loop.</h2>
             </div>
             <p class="section-intro">Submit one canonical source, correct an existing record with primary evidence, or document a loop you have run.</p>
@@ -1439,7 +1520,7 @@
         <div class="wrap">
           <div class="section-head">
             <div>
-              <p class="section-label"><i class="icon" data-lucide="message-circle-question-mark" aria-hidden="true"></i><span>08 / FAQ</span></p>
+              <p class="section-label"><i class="icon" data-lucide="message-circle-question-mark" aria-hidden="true"></i><span>09 / FAQ</span></p>
               <h2>Questions that define the boundary.</h2>
             </div>
             <p class="section-intro">Use these tests to decide whether a recurring agent system needs a loop.</p>
@@ -1490,6 +1571,7 @@
             <a href="https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/MANIFESTO.md">Manifesto</a>
             <a href="https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/TAXONOMY.md">Taxonomy</a>
             <a href="https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/ANTI-PATTERNS.md">Anti-patterns</a>
+            <a href="https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/FUTURE-DIRECTIONS.md">Future directions</a>
             <a href="https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/meta/CURATION.md">Curation + provenance</a>
             <a href="https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/data/resource_source_audit.csv">Source audit</a>
           </div>
@@ -1888,6 +1970,7 @@
               });
               explorer.dataset.indexState = 'ready';
               render();
+              restorePostResourceHash();
             })
             .catch(function () {
               loading = false;
@@ -1898,6 +1981,17 @@
             });
         }
 
+        function restorePostResourceHash() {
+          var targetId = window.location.hash.slice(1);
+          if (!targetId || targetId === 'resources') return;
+          var target = document.getElementById(targetId);
+          if (!target || !resourceSection) return;
+          if (!(resourceSection.compareDocumentPosition(target) & Node.DOCUMENT_POSITION_FOLLOWING)) return;
+          requestAnimationFrame(function () {
+            target.scrollIntoView({ behavior: 'instant', block: 'start' });
+          });
+        }
+
         var resourceSection = document.getElementById('resources');
         if (window.location.hash === '#resources') {
           loadResources();

meta/hf_card_body.md

@@ -69,6 +69,18 @@ Use `url` as the durable join key. `row_id` and `source_line` are positional and
 
 Do not use `signal_strength`, GitHub stars, forks, or inclusion in this collection as a quality label, endorsement, or automated ranking of scientific validity.
 
+## Future Directions
+
+The project maintains a role-based agenda for moving Loop Engineering forward:
+
+| Audience | Priority categories |
+| --- | --- |
+| Researchers | Factorized evaluation, verification science, state across runs, stopping policies, and human oversight |
+| Infrastructure and reliability engineers | Durable execution, receipts and replay, least privilege, portability, and reliability economics |
+| Application and product developers | Loop-worthy use cases, domain verification, escalation UX, progressive rollout, and maintainable adoption |
+
+The complete [Future Directions agenda](https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/FUTURE-DIRECTIONS.md) defines measurable outputs, concrete starter projects, shared infrastructure priorities, and a qualification checklist.
+
 ## Dataset Structure
 
 The primary configuration is `resources`, with one `train` split backed by `data/resources.jsonl`. `data/resources.csv` contains the same rows for spreadsheet and dataframe workflows.