An open protocol for deterministic, auditable AI-powered endurance coaching. Built for athletes who want AI coaches that follow science, not speculation.
Section 11 (S11) is a structured framework that enables AI systems (ChatGPT, Claude, Gemini, Grok (web/app), Mistral, etc.) to provide evidence-based endurance training advice with full auditability and deterministic reasoning.
- Deterministic: Same inputs produce same outputs
- Auditable: Every recommendation cites specific data and frameworks
- Evidence-based: Grounded in 15+ peer-reviewed endurance science models
- Athlete-controlled: Your data, your thresholds, your goals
| File | Description |
|---|---|
| SECTION_11.md | Complete protocol: AI Coach Guidance (11 A), Training Plan Protocol (11 B), Validation Protocol (11 C) |
| examples/workout-library/ | Workout Reference Library: 26 session templates, adapted to your data |
| examples/agentic/ | Calendar writes, activity stream reads and external API reference |
| examples/agentic/fit-file-and-merger/ | Multi-source FIT fusion, identity, verification, recovery and duplicate-cleanup reference; documentation only |
| examples/json-local-sync/ | Local automated sync, no GitHub needed |
| examples/dfa_a1/NON_GARMIN.md | DFA a1 requires Garmin + AlphaHRV; other-platform verification guidance |
| DOSSIER_TEMPLATE.md | Your stable private athlete context |
| examples/ | Sync guides and fictional JSON schema examples |
| examples/reports/ | Pre-workout, post-workout, rolling 7-day and 28-day, Season and on-request weather reports |
| SETUP_ASSISTANT.md | Interactive AI-guided setup |
| manifest.json | File versions and update tracking |
| LICENSE | MIT license |
sync.py redacts metadata.athlete_id, but does not anonymize your export. Activity names and IDs, date of birth, age, sex, height, location, timezone, notes, wellness and training data remain. Saved-workout names, descriptions, folders and full structures reveal planning intent, including goal events, target adaptations and prescribed intensities. Route JSON and weather maps reveal geometry and departure times; opaque map filenames do not hide their content. Keep exports and your dossier private. A public repository exposes its files.
Section 11 operates no hosted backend. Your data passes through the services you configure, and the AI, runtime, model, connector, repository and storage providers involved set their own processing and retention terms. See Route Weather for forecast-provider data transfers, defaults and both off switches.
Configured credentials go to their corresponding service: the Intervals.icu API key in its Authorization header, and a GitHub token to GitHub for configured publishing or issue creation. They are never included in exported or published JSON.
sync.py and push.py use private sync state in <data folder>/state/ for request accounting and cached athlete responses. Local and Manual keep it between runs; syncing the whole folder to cloud storage uploads it too, and a whole-folder connector can technically read it. data_overview.state_note and READ_THIS_FIRST.state_note in latest.json explicitly tell AI tools to ignore it. The supplied publishing and workflow artifact paths never commit or upload state. GitHub and On-demand retain only profile, curves and messages in a separate Actions cache, outside the repository and training-data download; raw core responses and accounting stay runner-only. Use latest.json.source_freshness for status.
Before you start: The documented data setups use Intervals.icu. You need an account containing your training data, plus your athlete ID and API key.
Intervals.icu is a free training platform. In its Settings, connect the service you record with, such as Garmin or Wahoo. Use a direct connection: activities imported only through Strava can appear in Intervals.icu but reach Section 11 with missing details.
Section 11 is free to use. Your chosen AI platform or hosting may have its own costs.
Recommended: Give SETUP_ASSISTANT.md to your AI for guided setup, whether you use a chat app such as ChatGPT or Claude, or OpenClaw and similar always-on agentic setups. Or use the Section 11 skill, if supported.
You can also follow the step-by-step guides below.
Create DOSSIER.md at the top of your data folder (or of your repository on GitHub) from the dossier template and fill in the context your AI cannot read from your data: background, equipment, stable constraints, health and injury history, tested fueling, and how you prefer to be coached. Current thresholds and load always come from your JSON. You can start coaching before the dossier exists; a missing dossier limits personalisation, not safety. No data folder or GitHub repository yet? Create it in step 2, then come back here.
Choose how your data reaches the AI:
- Local sync: Best for a computer you fully control and keep on often. A timer checks for new data every 1, 2, 5 or 10 minutes (you choose) while it is on, and slows or pauses if your request allowance runs low. Your AI can read the data in two ways: an AI agent on that computer, such as OpenClaw or Claude Code, opens the folder directly, or you keep the data folder in Google Drive or similar and connect a chat AI like ChatGPT or Gemini to it.
- GitHub sync: Best for automatic refreshes while your computer is off. Your files live in a private repository (an online project folder). The default schedule is every 30 minutes, but runs can be delayed or skipped. Sync Now lets you request a run, which may queue. Private repositories use your GitHub Actions minutes; a public repository runs free but exposes your training data (trade-offs). See schedule options and costs before changing the frequency.
- On-demand sync: Best for refreshing on request through GitHub. Start a run from your phone or browser. Once it finishes, download your repository ZIP for your AI or refresh its connection. No schedule and no local Python needed.
- Manual export: Best for occasional exports from your computer. Choose how many recent days to export, then upload the files to your AI. Use
--no-timerduring installation for manual-only use; otherwise an automatic timer is installed. Export and upload again when your data changes.
Raw file links work without signing in only for public repositories. For private files, connect your AI to the repository or upload the files.
We recommend one setup per user. Before switching setups or computers, stop the old timer or workflow to protect Intervals.icu. Different data folders do not share a request count, even on the same computer. Multiple setups using the same API key can exhaust its allowance.
Refreshes use your Intervals.icu request allowance, so choose a slower schedule if other tools also use your key. See local refresh limits and GitHub request limits for details.
Section 11 checks for updates during sync runs and asks for your acceptance by default; automatic installation is optional. See updates and rollback for how to install or undo an update.
Keep your own files in custom/ or outside the data folder: updates replace section11/ and can remove unprotected same-name files elsewhere. Read the removal rules before accepting an update.
Agentic: you use an AI agent such as OpenClaw, usually running on a computer or server that is on a lot or always. Web chat: you use a chat app such as ChatGPT, Claude or Gemini.
- Agentic: open PROJECT_INSTRUCTIONS_AGENTIC.md, copy the text in its grey box (GitHub shows a copy button at the top right) and add it to your agent's persistent configuration. See Agentic Platforms for your platform (OpenClaw, Claude Code, ChatGPT Codex, Grok Bot, Hermes Agent, etc.).
- Web chat: open PROJECT_INSTRUCTIONS_WEB.md, copy the text in its grey box and paste it into your Project, Space, Gem or custom instructions. See Web Chat Platforms for your platform (ChatGPT, Claude, Gemini, Grok, Mistral Vibe, etc.) and Platform Setup for connection details.
If your AI agent can read your data folder directly, give it access to the installed section11/ folder and current exported data. For web chat:
- Local sync: use a supported cloud connector and check it can read your current data
and the installed
section11/folder. Without working connector access, use the Manual export upload steps below. - GitHub sync or On-demand: use a connector or a link to your own repository first. If the chat can use neither, open your repository's Code > Download ZIP and give that ZIP to the chat or project. It includes the installed protocol and templates.
- Manual export: upload the exported data files,
DOSSIER.mdif used, and a ZIP of thesection11/folder only. Never zip the whole data folder, which can contain credentials and rollback state.
An uploaded ZIP is a snapshot. Re-upload for fresh data and after updates, replacing
the previous upload. Keep the optional dossier as DOSSIER.md at the top of your data folder or repository; replace uploaded copies when it changes. Keep any repository containing it private; Git keeps old versions.
Connectors may need refresh or re-import; check Platform Setup.
Check it works. Ask your AI:
Read my latest.json. When was it generated, and what is my most recent activity? Can you open the Section 11 protocol and the report templates? If a file is missing, tell me which.
Check that the date is recent and the activity is your latest one in Intervals.icu.
- Activities are missing or incomplete, including Strava imports
- Your data looks out of date
- Your AI cannot read the data files
For sessions with a runtime filesystem, whether on your own machine or a hosted computer. Filesystem access, code execution and write access are separate capabilities, verified per target. Reading data does not authorize calendar writes.
Use Local sync when the runtime can reach your data folder, or GitHub sync for repository access and backup. Hosted runtimes may need a connector or authenticated repository to reach your data.
Copy the block in PROJECT_INSTRUCTIONS_AGENTIC.md into the agent's persistent configuration. Point it at your files using Connect Your Agent.
This is the reference setup. Section 11 works well with OpenClaw: persistent memory, heartbeat scheduling, autonomous execution, and structured validation.
- Clone or copy the Section 11 skill folder into your OpenClaw skills directory (skills are just directories with a
SKILL.md) - Authenticate:
gh auth login, or for limited access use a fine-grained personal access token scoped to your data repo only
- Install: see claude.ai/download or
npm install -g @anthropic-ai/claude-code - Authenticate:
gh auth loginfor GitHub repo access, or clone your data repo locally
Cowork is a desktop app that can read files directly from your filesystem.
- Clone your data repo locally and grant Cowork access to that folder
- Or add GitHub from Cowork's Connectors directory.
- Connect your GitHub account at chatgpt.com/codex and authorize your data repo; web and desktop app connect directly
- Or install the CLI:
npm install -g @openai/codex(reads from local filesystem)
- Install:
npm install -g @google/gemini-cli(ornpx @google/gemini-cli) - Clone your data repo locally and verify that Gemini CLI can read the required files
Runs on a user-scoped shared computer and authenticates through Cursor. It has the agentic capability class, but it is experimental: it is not validated end to end against the Section 11 pipeline, and capability class is not a support promise. Any write must be verified against the specific target before it is used or assumed.
All Bots on your account share one cloud computer: files, browser sessions and command-line credentials are not isolated. Deleting a Bot does not clear shared files or browser sessions. Cloud storage is required and Legacy Privacy Mode is unavailable. Check your xAI and Cursor privacy settings before adding a sensitive dossier.
Hermes reads its runtime host's filesystem, which may differ from your machine. Use a pointer file for your data, set paths explicitly, and verify the resolved working directory before relying on relative paths.
An open-source coding agent for the terminal, IDE and desktop.
Control interfaces. T3 Code provides an interface for supported agents such as Claude Code, ChatGPT Codex and OpenCode. Section 11 access depends on the underlying agent's configured filesystem and tool permissions; T3 Code is not a separate coaching runtime.
With verified execution, credentials and access, use examples/agentic/:
push.py: manage planned workouts, update sport-specific thresholds and annotate activities. Preview writes through the CLI; confirm only after approving the exact change.pull.py: read raw per-second GPS, altitude, power and HR when the synced summaries are insufficient.EXTERNAL_APIS.md: endpoints for Strava, MET Norway, Open-Meteo and Intervals.icu. Read sync's briefs for covered route forecasts; direct weather calls are for uncovered locations or times.
The FIT File & Merger reference needs no execution to read. It describes a cycling-tested implementation; other sports and sources need their own validation.
For AI platforms reached through a browser or app, with no runtime-accessible filesystem. These sessions reach your data through a connector or authenticated repository, an upload or attachment, or a URL fetch.
Copy the block between the fences in PROJECT_INSTRUCTIONS_WEB.md into your Project, Space, Gem or custom instructions. If the AI has a runtime filesystem, use PROJECT_INSTRUCTIONS_AGENTIC.md.
For URL fetch, replace [USERNAME]/[REPO] with your public data repository path. Connectors need no URL editing and grant no write authority by themselves.
Create a project, add the instructions above, then connect your data or upload the files listed in Quick Start. Check metadata.last_updated before relying on a report. Access does not guarantee freshness.
GitHub connector status for web-chat platforms. Runtime repository access is covered in Agentic Setup.
| Platform | How to Connect | Plans | Private Repos | Refresh | Permissions / Caveats |
|---|---|---|---|---|---|
| ChatGPT | Apps / Plugins → GitHub | Varies: plan, workspace, surface | Selected repositories | On demand; no index | Read-only; Codex for writes; chat availability varies |
| Claude | + → Add from GitHub | All, including Free; organization policy | Yes | Manual Sync now | Files only; no history, PRs or metadata |
| Gemini | Add files → More uploads → Import code | Personal/eligible Workspace; 18+; Keep Activity on | Linked GitHub account | Frozen; re-import to refresh | Computer only; read-only; one repo/chat; 5,000 files / 100 MB; not Gem Knowledge |
| Grok (web/app) | grok.com/connectors → GitHub |
All; Business/Enterprise admin provisioning | Authorized repositories | On demand | Repo/issues/PRs/code; writes undocumented; review OAuth grant |
| Mistral Vibe | Work → Connectors → GitHub App | All, including Free; organization policy | Authorized repositories | Real-time / on demand | Search/issues/PR actions; approval required |
| Perplexity | Settings → Connectors → GitHub | Pro, Max, Enterprise Pro/Max | Yes | On demand | Actions; broad grant includes repository deletion/workflow updates; review carefully |
Google Drive access for .json files. ✅ Supported;
| Platform | .json via Drive |
Plans | Refresh | Notes |
|---|---|---|---|---|
| ChatGPT | Plus/Pro; enabled Business/Enterprise/Edu | On-demand search; sync on Pro/eligible workspaces | App capability, region and policy limits | |
| Claude | ✅ Supported | All; organization policy | On demand; re-select changes | JSON attachments; only Google Docs stay synced |
| Gemini | ✅ Supported | Signed-in; Workspace policy | Re-select changes | Add from Drive; supported-file uploads; plan limits |
| Grok (web/app) | ✅ Supported | All; Business/Enterprise admin provisioning | Real-time / on demand | JSON supported; Docs/Sheets/Slides/other files |
| Mistral Vibe | Legacy: Team/Enterprise; MCP: admin setup | Legacy: scheduled; MCP: tool-dependent | JSON/files; legacy connector being replaced; check Mistral's docs | |
| Perplexity | ✅ Supported | Pro, Max, Enterprise Pro/Max | Real-time search; selected-file sync | Enterprise: My Files sync/high-precision search |
Note: Check the linked vendor documentation before setup. Connector availability and behavior change often, and a stale connector can silently serve old training data.
Create a Project and add instructions in its settings. Install the GitHub app for the repository owner and select your repository. Name the repository and file path in requests. Otherwise upload to Project Files.
Create GPT → Configure. Paste instructions, upload files under Knowledge, and enable Web Browsing for URL fetch.
Create a Project and add Project Instructions. Select repository files through the GitHub integration, or upload to Project Knowledge. Enable Web search for public URL fetch.
Create a Gem and put the project contract in Instructions. Follow the delivery steps in Quick Start for the installed files and data. Gemini limits ZIP uploads. If a Gem rejects or truncates the ZIP, use repository import in an ordinary Gemini chat; it is not a Gem Knowledge source.
Create a Project and add instructions. Connect your data using the table above, or upload to Sources.
Create a Project and add instructions. Connect your data, or upload during project creation.
Create a Space or use chat, then add instructions. Connect your data, or upload to the Space.
After configuration, test with:
"How was today's workout?"
Good response includes:
- ✅ Reads current
latest.json; loads companion detail only when needed - ✅ Available session facts: name/type/start, moving/total/paired planned time, distance/ascent, sport-specific speed/pace, terrain, temperature, wind, climbs, power avg/NP/W/kg, max power/peaks, HR/zones, cadence, work, calories, Carbs used and TSS
- ✅ Executed structure, gated diagnostics and subjective fields per the post-workout template; peaks describe this activity only
- ✅ Weekly totals, brief Interpretation and every planned Tomorrow session
- ✅ Plain formatting without source-marker clutter or excessive emojis
Bad response:
- ❌ Requests accessible data, invents fields or omits available session data
- ❌ Treats peaks as FTP/readiness evidence or whole-session power as main-set power
- ❌ Recommends recovery for normal TSB (-10 to -30) without other triggers
"Good morning, what's my workout today?"
Good response includes:
- ✅ Readiness assessment (HRV, RHR, Sleep vs 7d/28d baselines)
- ✅ Load context (TSB, ACWR, Monotony if > 2.3)
- ✅ Capability snapshot (Durability (cycling) 7d + trend, TID drift if not consistent)
- ✅ Today's planned workout details from planned_workouts
- ✅ Go / Modify / Skip recommendation with reasoning
Bad response:
- ❌ Skipping readiness check and jumping straight to workout description
- ❌ Missing capability snapshot (Durability (cycling), TID drift)
- ❌ Generic "listen to your body" without referencing actual metrics
"weather for my rides"
Good response includes:
- ✅ Route names/dates, labelled weather, moving/total time, and matching readable map pages
- ✅ An explicit assumption for unset start times, and one linked attribution naming the actual provider
- ✅ Missing/stale/disabled status, or the no-route message and GPX/TCX hint; Weather at home only with a resolved Place/source and actual fetch capability
Bad response:
- ❌ Invented weather, unseen maps or precise section forecasts from an outlook
- ❌ Home weather substituted for a stale route brief
See the Weather Report template.
If you use a dossier, ask about something it records:
"What's in my regular training bottle?"
Good response includes:
- ✅ Your recorded bottle/equipment, or a clear statement that the official copy or entry is unavailable
- ✅ On "Which dossier did you read?", the matching authority statement, data-folder root, revision and last-reviewed date, identifying uploads as delivery copies
- ✅ On "How many should I take on my next long ride?", a current JSON/planned-session read before prescribing numbers, or a statement that no session is planned
Bad response:
- ❌ A generic answer or earlier-chat value presented as yours
- ❌ A stale or unofficial dossier treated as current
Check the source file, the delivered copy, and the AI's current read. Your web or agentic instructions define access and freshness rules.
Ask which configured paths failed. Check the runtime's actual data directory, connector scope, attached files, or public raw URL and browsing capability. A connected account may lack access to the particular repository.
Reload the project instructions, then ask the AI to identify the file it read and its timestamp/version. If it still sees an old copy, try a fresh conversation.
Plain raw.githubusercontent.com URLs return 404 for private repositories as well as wrong paths. Check the repository, branch and root filename. For private data, select the repository in an authorized connector or authenticate the runtime; organization approval may be needed. If access still fails, upload fresh files as described in Quick Start.
Compare the source and the AI's latest.json using export_freshness.generated_at and a trusted current time, falling back to metadata.last_updated. Both are UTC (Z); only older exports with a naive last_updated need the sync machine's known timezone. If time or timezone is unknown, say age cannot be verified when relevant. After expected_update_by, mention late delivery once when relevant and look for a newer accessible copy. Manual, On-demand and unknown schedules have no deadline: expect under 12 hours and mention age at 12-24 hours when relevant. Beyond 24 hours, request refresh for advice needing current data, even during a permitted pause. Continue qualified coaching; lateness alone is never a global stop.
For ordinary coaching, read source_freshness.warnings: mention source_stale only when its named sources affect the answer. Other warning codes describe sync operation, not bad data. Sync owns cached-source validity; do not rejudge per-source acquisition ages or statuses. Empty warnings and a recent as_of do not certify a live feed. Metric dates and native validity rules still apply. See the field reference.
If the source is old, check sync below. Otherwise replace uploads, refresh/re-import connectors, or let your cloud folder finish syncing. Raw GitHub content can lag a few minutes. Ask for a fresh read in this response; use a new conversation if old values persist.
Check the API key, athlete ID and run log; no recent activities can legitimately mean no new activity data.
- Local: check the timer,
sync.logandsource_freshnessfor a step-down, pause or provider hold before treating the timer as broken. Keep.sync_config.jsonat the data root and retainstate/there unless you set a state override. See request allowance and pauses for the exact pause message and recovery steps. - GitHub: inspect Actions → Auto-Sync Intervals.icu Data. Check secrets, Actions minutes and workflow permissions; request Sync Now when needed. See workflow troubleshooting.
- On-demand: trigger the workflow and download its completed
training-dataartifact within seven days, or refresh your connector. - Manual: export again and replace uploads.
For script/workflow replacement, see Update Notifications.
Missing device measurements remain unavailable. terrain_status and weather_status can explain absent data.
Strava-sourced activities may appear in Intervals.icu but reach its API only as stubs under Strava's API terms. Connect Garmin, Wahoo or your device platform directly to Intervals.icu, or upload original activity files so new records arrive from a non-Strava source.
Readiness needs rMSSD in Intervals.icu's hrv field. Apple Health commonly supplies SDNN, which passes through as hrv_sdnn for context and never substitutes for rMSSD. SDNN-only records report rmssd_missing_sdnn_available.
An upstream tool must supply rMSSD. Community projects are not verified or supported by Section 11; allow about a week for a new seven-day baseline. See Readiness Decision.
Use Knowledge files for Gems and repository import for ordinary chats. Re-import after changes; check account eligibility and GitHub authorization in Platform Setup. Large archives may exceed import limits: see Archive. If the AI ignores data, ask it to read the root latest.json and report metadata.last_updated.
Check Platform Setup, including administrator provisioning in Business/Enterprise. If the connector remains unavailable, upload fresh files; public URL fetch exposes the data described in Privacy & Security.
Root DOSSIER.md in the data-folder root (repository root on GitHub) is authoritative. Ask for its authority statement, revision and last-reviewed date; compare uploads with that root file. Resolve conflicting delivery copies yourself, then replace or remove superseded ones. If the dossier is configured or kept elsewhere, including an old Official dossier location field, the AI tells you once that the standard location is the data-folder root and asks you to move it. It never moves or copies it itself; until moved, it uses the copy you point to. Keep any repository containing it private; Git keeps old versions.
Dossier changes require your exact approval and verified write access, otherwise the AI returns a revised artifact for you to save. Check that its revision increased. The dossier template explains preference precedence and maintenance.
Reload the project instructions and ask for the protocol version, current latest.json timestamp and, if used, dossier revision. Compare them with the source; a confident summary is insufficient. Resolve stale, missing or duplicate copies before continuing.
For a fresh protocol copy, use the local updater, or download the repository in a format your platform accepts. It contains no private athlete data: supply your own current JSON. Files under examples/json-examples/ and all .example.json files are fictional schema examples, never coaching data.
Detailed causes, companion markers and refresh checks: source-copy checks.
Defines behavioral rules for AI coaches:
- No virtual math: use current logged values
- Missing data: read configured sources, then ask if required data is unavailable
- Tolerances and auditability: apply the protocol's numeric checks and evidence rules
- Self-validation: complete Steps 0–10, including 5b and 6b
See the AI Self-Validation Checklist.
Defines rules for AI systems generating or modifying training plans:
- Phase alignment with macro-cycle
- Volume ceiling validation (±10% of baseline)
- Intensity distribution control (80/20 polarization)
- Session composition rules
- Audit metadata requirements
Records data access, source timestamps, protocol version, passed/failed checks, confidence and missing inputs in structured audit metadata. See the schema and field definitions.
The protocol integrates 19+ validated endurance science frameworks:
| Framework | Application |
|---|---|
| Seiler's 80/20 Polarized Training | Intensity distribution |
| Gabbett's ACWR (2016) | Retrospective load context, not standalone injury prediction |
| Banister's Impulse-Response | CTL/ATL/TSB dynamics |
| Foster's Monotony & Strain | Overuse detection |
| Issurin's Block Periodization | Phase structure |
| Coggan's Power-Duration Model | Efficiency tracking |
| San Millán's Zone 2 Model | Metabolic health |
| Skiba's Critical Power Model | Fatigue prediction |
| And more... | See Section 11 for full list |
Training blocks adapt dynamically using dual-stream phase detection:
Build ↔ Base ↔ Deload ↔ Peak → Taper → Recovery
↑ Race calendar
Overreached (safety gate — triggers from any state)
Phase classification combines retrospective history (4-week CTL/ACWR/hard-day trends) with prospective calendar data (planned workouts + race proximity). Confidence scoring (high/medium/low) and reason codes provide full auditability.
Examples from the protocol's readiness rules:
| Trigger | Response |
|---|---|
| HRV ↓ >20% | Consider easy day / deload |
| RHR ↑ ≥5 bpm | Flag fatigue/illness |
| Current Feel ≥4/5 (solicited) | Adjust volume 30–40% for 3–4 days |
| RI <0.6 | Mandatory deload |
Green-light criteria for safe load increases:
- Durability Index ≥0.97 for 3+ long rides (≥2 h)
- HR drift <3% in aerobic sessions
- Recovery Index ≥0.85 (7-day mean)
- ACWR within 0.8–1.3
- Monotony <2.5
- Current Feel ≤3/5, when solicited
The protocol is designed to work with Intervals.icu as the primary data source. Set up a JSON mirror that syncs your fitness metrics, recent activities, wellness data, zone distributions, and planned workouts.
The sync script pre-calculates Section 11-compliant metrics so AI doesn't need to compute them. Key metrics include ACWR, Recovery Index, Monotony/Strain, Grey Zone %, Quality Intensity %, Easy Time Ratio, Benchmark Index, Phase Detection, Seiler TID, Durability (cycling), per-sport curve deltas and main-sport TID Drift; strain alarms use strain_main_sports while legacy strain remains all-sport context.
Zone aggregations (TID, polarization, grey zone %) default to power zones with HR fallback. Configure ZONE_PREFERENCE to override per sport, e.g. run:hr,cycling:power for runners who prefer HR-based zone analysis. See auto-sync setup or local sync setup for configuration.
See examples/README.md for the full derived metrics table and data output structure.
Report intensity selects main sports independently per period and shows separate power/HR lines with covered weeks and sessions. Pace-only sports can qualify, with zones unavailable until pace mapping is supported. Legacy fields and activity zone_reference retain their meanings. New history blocks appear on normal refresh; missing blocks are unavailable. Optional report_main_sports is an explicit producer configuration, absent by default. See Report intensity source.
The script generates history.json with tiered granularity: daily (90 days), weekly (180 days), and monthly (up to 3 years). Includes period summaries, FTP timeline, and data gap detection. latest.json covers current coaching. Provide history.json for trend, phase or longitudinal questions, and intervals.json or routes.json (if present) when a task needs them.
The script maintains ftp_history.json to track indoor and outdoor FTP changes over time, enabling Benchmark Index calculation. See examples/json-auto-sync/SETUP.md for details.
latest.json includes interval_data per recent activity: timeline plan match, scoped compliance evidence and descriptive Intervals.icu detected intervals side by side. has_structure points to 14-day detail; only plan-match context survives through 28 days. See Which interval data to use.
Whole-session average power and NP are not main-set power. Unknown, no-plan and descriptive fallback differ; broad plan match does not prove formal compliance. See Completed Session Structure for interpretation, freshness and limits.
intervals.json holds per-interval power, HR, cadence, zones, timing and W'bal, plus session-level DFA a1 when AlphaHRV recorded. Follow the independent has_intervals, has_dfa and has_structure flags in latest.json; an entry alone may be a whole-session RECOVERY placeholder.
Interval/DFA data covers cycling, run, ski, rowing and swim, with a 72-hour scan window and 14-day retention. Structure has separate sport rules. See Interval Data Mirror for fields and interpretation.
The script writes routes.json on every sync, with terrain analysis for planned events that have GPX/TCX file attachments: distance, elevation, course character, climbs (position, length, average and max gradient), descents and a downsampled polyline. Planned-route climbs support pre-workout reports; completed-activity terrain comes from latest.json for post-workout reports. Events with terrain are flagged has_terrain: true. Attachments are cached after download and parsing.
latest.json starts with data_overview, immediately after READ_THIS_FIRST when present: companion files, locations and route/weather/map availability. The AI reads it first; file presence does not guarantee access.
Keep the export folder between runs, including routes.json and route_eta_history.json. GitHub/On-demand persist these and maps in the repository; their six-JSON download artifact excludes maps and ETA history. Manual uploads need no extra map step. Run one sync process at a time; automated local timers should use --lockfile.
For planned outdoor rides with routes in the next five days, sync forecasts wind relative to travel, gusts, rain and temperature, timed with an estimate calibrated on your outdoor rides when history allows. Pre-workout reports include weather; ask "weather for my rides" anytime. PNG maps appear in maps/weather/ (example map). Without a usable route, a fetch-capable AI can give a home forecast.
On by default. Disable with "route_weather": false in sync config or SECTION11_ROUTE_WEATHER=0 (GitHub repository variable). Sampled route coordinates/elevations go to Open-Meteo, with MET Norway fallback; providers see the requesting IP, but no credentials, event names or activity history. Open-Meteo's free service is for non-commercial use. Maps reveal your route. Forecasts are planning aids, not safety assessments. See SECTION_11 for details.
Without a start time, 10:00 local is assumed and the AI asks when you leave. Set the time in Intervals.icu or route_weather_default_start.
saved_workouts.json mirrors your Intervals.icu folders and complete saved-workout structures for convenient reads through any data path. It is read-only; edits belong in Intervals.icu. Refresh runs on a six-hour throttle and retains the last good copy on failure, so check refresh.status before use.
It is inventory, not the Workout Reference Library or proof of historical prescriptions. See Saved Workouts Mirror.
GitHub/On-demand: select Update in the sync workflow to accept a verified replacement after its notice; Status shows the plan without installing. Local: run python3 section11/examples/sync.py --update after the daily advisory. Automatic installation is opt-in for both. See Updating Section 11.
When sources conflict, trust order is:
- Intervals.icu (primary)
- JSON Mirror (Tier-1 verified)
- Athlete-provided values (<7 days old)
The athlete dossier is not a rung in this hierarchy. It holds stable private context, never a current metric.
Also compatible with any platform that exports structured training data.
"How was today's workout?"
"Analyze my last 7 days against my targets. What's my compliance rate? Any red flags?"
"Have I met the green-light criteria for extending my Friday long ride to 5 hours?"
"Here's my workout file. Did I hit my intervals within tolerance? What does the HR drift tell us?"
- AI still makes mistakes: This protocol reduces errors but doesn't eliminate them
- Not a replacement for human coaches: Best used alongside professional guidance for serious athletes
- Requires honest data: Garbage in, garbage out
- No medical advice: Consult professionals for health concerns
This is an open protocol. Contributions welcome:
- Bug reports: Found an inconsistency? Open an issue
- Framework additions: Know a validated model that should be included? Propose it
- Translation: Help make this accessible in other languages
- Integration guides: Built a tool that uses this? Share it
This project is licensed under the MIT License; see LICENSE for details.
You can:
- Use it for personal or commercial projects
- Copy, modify, and distribute it
- Include it in closed-source or open-source tools
You must:
- Include the copyright notice
- Include the MIT license text in copies or substantial portions of the software
The software is provided "as is", without warranty of any kind.
- David Tinker: Creator of Intervals.icu
- Clive King: Pioneer of GPT-based endurance coaching and URF
- Intervals.icu Forum community
- Researchers behind the scientific frameworks cited in Section 11
- Website: section11.net
- Protocol: SECTION_11.md
- Template: DOSSIER_TEMPLATE.md
- Examples: examples/
- Report Templates: examples/reports/
- Intervals.icu: intervals.icu
- Discussion: Intervals.icu Forum Thread