← All workshops · Slides of workshop 5
For the person who runs the workshop. The participants' page is the README.md of the repository
of the workshop, yal-workshop-wiki (source:
workshops/5-wiki/ of the facilitator's repository), and their one-page sheet
sheet.md.
⚠️ Important. In mode E, as in modes C and D, the model can read every file of the repository, including its
README.md, andreplaydoes not flag it. The participants' pages therefore never quote an expected answer. Keep it that way when you adapt them, and keep this sheet out of the participants' repository: it talks about answers and contradictions. It is also on the public site of the slides, so it is not a secret: anyone, a participant too, can find it. What protects the test is that it is not in the participants' repository, where the model would read it.
Question of the day: how do you organise knowledge so that a person and a model can explore it? Message to leave with: the work is done once, before the question, in pages linked to each other that a person and a model can explore; each fact keeps its way back to a source; the model writes and reports, the lint is the check before the wiki feeds a chatbot used by others, and a person checks, arbitrates and signs. On a small wiki, the graph does not help the model yet, but it already helps the people who explore. This is also the last workshop on the pair's chatbot, so it answers the question of the whole series: what enters the model's context, when, and who decides?
| Time | What | How you check it |
|---|---|---|
| 15 min | One Anthropic API key per pair, all in one workspace for the room, capped at $3 × the number of pairs (see Who pays). | workshop-login says Key OK |
| 30 min | Your own run, in a Codespace that you keep for the session: a new Codespace from https://codespaces.new/Youth-AI-Lab/yal-workshop-wiki, workshop-start 5, openwebui start in terminal 2; workshop E, ingest corpus/80-benevento-aib-2026.pdf into the wiki; wiki-graph --highlight 80; replay E; a new workshop E, lint the wiki, and one finding checked by hand; wiki-explore, then the Montoro check of Part 4; score E 3 10 11; chatbot-export --name "Feurisson"; openwebui chatbot --name "Feurisson" |
each step as the page says; about $1.05 in all (see below) |
| 5 min | Prepare what you project in Part 4: in your Codespace, wiki-graph, open wiki/graph.md with Ctrl+Shift+V, and wiki-explore in a tab next to it. Zoom both out until no label can be read |
both maps are open, side by side, and no label is readable |
| 5 min | Remind each pair, a few days before, to bring my-work.zip (the newest one: it may be called my-work (1).zip), with a summary it signed in workshop 3 |
the pair can name its number, usually 81 |
| 5 min | Print the one-page sheet (one per pair). For the exit cards, one blank card or half sheet per person: the three questions are on the page (Part 5, Step 6), on the deck and below (Exit card) | - |
| 5 min | Self-assessment form ("After"): open it, check that its QR code opens it on a phone, and that it collects no names, e-mail addresses or ages (the self-assessment; form link TO BE COMPLETED) | the form opens on a phone |
| - | Translate the one-page sheet (sheet.md) into the language of the group if needed, before you print it: keep the commands and what the terminal prints in English, as they appear on the screen |
- |
How to create a key and set a limit: Open WebUI guide, Your Anthropic API key.
Or OpenRouter keys (sk-or-): same commands, no Claude model, and the whole room on one model so
that totals compare (OpenRouter guide). The
figures of this sheet come from Claude and do not hold for OpenRouter
(its measures). Cost per pair with the default model,
DeepSeek V4 Flash: about $0.04 for the terminal commands of this workshop and about $0.12 for workshops
1 to 5, played once on 2026-11-05 in a container of the image 2026.11.4 (OpenCode 2.1.4), not in a
Codespace and not with a group (measures). Not counted:
replay RAG and the Open WebUI chatbots. A pair that asks more questions spends more: take it as a
minimum.
The workshops are for adults: no young person types in the Codespace or holds a key, and OpenRouter
asks for users of 18 or older. Young people only use a chatbot in Open WebUI, in their browser. When what
they wrote goes through OpenRouter, leave out names and personal details, and type
workshop-login --europe first
(SETUP.md, Young people and OpenRouter;
OpenRouter guide, Part 8).
With --europe's chatbot model (Mistral Small 2603), the lint is not reliable: on 2026-11-02, it read only index.md
and log.md, then answered "Everything is clean!", while the pages had 6 contradictions and one page
was missing from the index (measures). One run only.
If a pair gets a clean report, show the proposals of your own report from the Claude path.
ℹ️ Note. Who pays. Each partner organisation pays for its own sessions: its own Anthropic account, workspace and keys, on its project budget. One workspace for the room, capped at $3 × the number of pairs: 8 pairs, $24. At Antwerp, the most advanced pairs used less than $2 in a 2-hour workshop. The limit is on the workspace, not on each key, so one pair can use up the share of another: keep the Anthropic console open during the workshop and watch the spending. Not checked: if the console limit is per month, the workshops of the same month share it. Then count $3 × pairs × workshops of the month, or raise the limit before each workshop. Indicative cost per pair for the whole series: TO BE COMPLETED at the rehearsal; this workshop's measured cost is under "Cost per pair".
⚠️ Important. Run this workshop only with image
2026.11.4or a later one: Part 5 needs it (mapwiki__index.md, the pair's skills,fire_dangerin Open WebUI). The repositoryyal-workshop-wikipins a later image: check.devcontainer/devcontainer.json.
🚫 Warning. Do not project these things, even for a moment:
answers/;wiki/log.mdand the Benevento page of the wiki (questions 10 and 11); the top of the lint report, which lists the contradictions already flagged (same questions);GRAPH_REPORT.mdbeyond itsCorpus Checkline: many node names of this report hold expected answers; the labels of the graphify map (questions 1, 2, 3, 4, 6 and 10), and the page titles of the fullwiki/graph.md(question 6). Show both maps only zoomed out, until no label can be read.
| Time | Part | What you do | What the room should have at the end |
|---|---|---|---|
| 0:00 | 1. Welcome (20) | Setup (SETUP.md, steps 1 to 5), workshop-start 5 --from my-work.zip, openwebui start in terminal 2. |
Ready for workshop 5; openwebui start running |
| 0:20 | 2. The wiki (30) | Read one page and the log; predict; workshop E, ingest; wiki-graph --highlight; replay E in terminal 2. |
In orange: … for each pair; one Contradiction line read |
| 0:50 | 3. The lint (15) | A new workshop E, lint the wiki; each pair checks one finding and decides. |
each pair can name one finding and its decision |
| 1:05 | 4. The graph (25) | Project the two maps side by side, then wiki-explore, the report, path, explain, and the check by hand. |
one verdict in one sentence per pair |
| 1:30 | 5. The chatbot you built (20) | score E 3 10 11 12, scores.md row by row, the table of the series, a local model and fine-tuning, chatbot-export, openwebui chatbot, workshop-save, self-assessment "After" (5 min, before the exit cards: the same form, the same personal code; QR code of the form, link TO BE COMPLETED; no QR slide in the deck yet; see the self-assessment), exit cards, what you leave with, the tools, WikiYouthBot. |
<name>.zip and my-work.zip downloaded |
| 1:50 | Margin (10) | Kept for Part 5: chatbot-export, Open WebUI, workshop-save, and the self-assessment "After" (5), which the 20 minutes of Part 5 do not include (timing to check at the rehearsal). |
- |
⚠️ Important. Part 3 is the tightest: 15 minutes, of which more than 2 for the lint, and a long report to sort. At 10 minutes, ask everyone to stop reading and to decide. If time is short in Part 5, keep
chatbot-exportandworkshop-save: everything else can wait, but a Codespace does not. The real durations of the five parts and of the margin are TO BE COMPLETED.
Without a key (2026-10-30, image 2026.10.34, digest
sha256:f257703bb8dca85006bfab6b0a085a64437e694ee0849223699bbd44d57fce69, Claude Code 2.1.303,
OpenCode 2.1.4, graphify 0.9.92; throwaway containers with no network at all, on a laptop, not a
Codespace):
workshop-start 5: 0.32 s without a save, 0.11 s with one. It already runs wiki-start, and
it does not start Open WebUI (only workshop 2 does): openwebui start is typed by hand.corpus/ only, with no summary and no signature: the pair ingests the official document,
ingest corpus/80-benevento-aib-2026.pdf into the wiki. The hint printed by wiki-graph speaks of
summaries/: the pair follows the page, not the hint.ingest does not check status: it would take a draft summary. The page says so.Contradiction lines (6 on 2026-10-30, 3 added on 2026-11-01), no orphan page.EXTRACTED, 22 INFERRED (11%), confidence 0.65 to 0.95. The 9
contradictions of log.md are nodes. 158 edges (77%) stay inside a community (counted in
graph.json). The "about 16%" of the page is what random lines give if each node keeps its number
of lines (16.3%; 14.1% if all pairs of nodes are equally likely). The gap, 0.61, is the modularity
that the calculation raises. Graph of 2026-11-02: 108 nodes, 192 edges, 147 inside a community
(77%), modularity 0.63. First graph (2026-10-21): 82 nodes, 132 edges, 7 communities,
17 INFERRED (13%), confidence 0.65 to 0.85, 103 edges (78%) inside a community.Corpus Check in tokens (the page says about 3 tokens per word, so a quarter of the window, and
half of the 100,700 of mode B): measured from the Context: lines of
measures/2026-10-30-wiki/answers/E-11h39-complet.md, with the pages of that day (commit
10495b7). Question 3 minus question 4 gives 2.82 tokens per word, question 6 minus question 4
gives 3.15. This is the gauge, tool calls included, not a tokenizer. The 4 characters of workshops
1 and 4 would give about 2 tokens per word: on our documents the gauge counts more, and we did not
measure why (3-facilitator.md, Token estimates).wiki-graph, wiki-explore path and explain, chatbot-export: under 0.25 s each; wiki-explore
1.01 s, port 8765. They all ran with no network. Only the map page needs the internet in the
browser: it loads vis-network from unpkg.wiki-explore path "Benevento" "Montoro": No node matching 'Benevento' found. (checked again
on the graph of 2026-11-06, with path and explain as printed on the page). Since that graph, the positions L… printed by explain point to
lines of the right page (they were wrong in the graph of 2026-10-21, facilitator/wiki-pack/GRAPHIFY.md):
the part of the page the model pointed to when it built the node or link, often wider than the fact
itself; for an INFERRED link, where the model looked, not a sentence that states the link.
The node log (the wiki log, 21 lines) is a crossroads: a path between two unrelated topics can
go through it.chatbot-export --name "Feurisson", no save, no ingest: 56 knowledge files (17 summaries, 18 wiki
pages, 21 documents), 152 KB. After the real ingest of 2026-10-30 (pages of the measures folder,
no save): 57 (17, 19, 21), the line printed on the page. With a save and one page written by
hand (not a measure of ingest): 59 knowledge files (18, 19, 22).workshop-save keeps the changes of the wiki, and not graphify-out.With a key (2026-10-30, image 2026.10.34, spending cap $2, about $1.16 spent; Sonnet 5.5, judge
Haiku 5.5; details in measures/2026-10-30-wiki/README.md).
AGENTS.md was the Feurisson example, before its protection rules (2026-10-31).
It ran in a throwaway container, not a Codespace, without a save, and in claude -p: nobody
accepted a prompt by hand, so the durations do not count the time of a pair.
| What | Measure |
|---|---|
ingest corpus/80-benevento-aib-2026.pdf into the wiki |
42 s, $0.14, context up to 25,863 tokens; 1 page created, 1 updated, index.md, log.md; a pair would have 6 permissions to accept (1 Write, 5 Edit). Checked by hand against corpus/text/80: no invention found. |
wiki-graph --highlight 80 after it |
19 pages, 21 source documents, the document and 2 pages in orange. The file was produced; its drawing was not looked at. |
lint the wiki |
2 min 10, $0.55, context up to 136,663 tokens (68% of the window); changed nothing (same SHA-256 sums before and after). In a real session, expect two more prompts: two for loops were refused in -p, and the model read the files one by one instead. |
replay E (11 questions, no question 12) |
1 min 57, $0.36; context 7,469 to 21,489 tokens; no answer touched the test material |
replay E 3 10 11 |
35 s, $0.11 |
Judge, score E --auto |
7 s, $0.0028; 21 out of 22 |
The context of three questions: question 3 used 16,603 tokens and only wiki/index.md and one page,
against about 23,000 in mode C (image 2026.10.33) and 100,700 in mode B (gauge after question 10,
rehearsal of 2026-10-21; every question there resends about 100,000).
Questions 10 and 11 used 13,117 and 13,484. One run each, on different images: a direction, not a
comparison to the token.
The one point lost is question 8: the judge objected to "unverified season dates", which do come
from mcp/fire-danger.json. That file was not dated of the day, so the tool gave no level, and the
model invented none. Today's page does not score question 8.
Older measures, kept with their date:
| What | Measure | When |
|---|---|---|
ingest of summary 80 |
6 Write/Edit permissions, gauge 28,400 | rehearsal of 2026-10-21, older image |
lint the wiki |
1 min 41, 114,400 tokens (57%). How it was counted was not found again: do not read 137,000 as an established rise. | same |
replay E |
1 min 59, $0.43; Q11 from the wiki, context 14,200 | rehearsal, older image |
| Building the graph | about 10 min, 163,000 tokens, one agent | 2026-10-21 |
| Building the graph again | about 7 min 30, Opus with no sub-agent, |
2026-11-02 |
| Building the graph after the fixes of the pages | 206 s, Opus with no sub-agent, starting from the scripts of 2026-11-02, $1.40 (total_cost_usd of Claude Code, public prices); 111 nodes, 204 edges (measures) |
2026-11-06 |
| Modes E and F, questions 1, 3, 5, 10 | E $0.46, 1 min 26; F $0.51, 1 min 38 | 2026-10-21, local, Claude Code 2.1.240, Sonnet 5.5 |
| Modes E and F, the distances question | E $0.20, 42 s; F $0.20, 33 s | same |
| Block F of the old single workshop, in a real Codespace | path and explain as printed; the distances checked in mode E in 15 s |
image 2026.10.10 |
openwebui start, first time |
3 min 21 in a Codespace, about 2 min 15 of download | 2026-10-26, image 2026.10.17 |
workshop-start 2 (with Open WebUI), first time |
3 min 51 and 4 min 13 in two new Codespaces, about 2 min 15 of download | 2026-10-30, image 2026.10.32 |
openwebui chatbot |
50 s for 58 files (local); 58 files also in a Codespace | 2026-10-22 and 2026-10-23, images 2026.10.15 and .16 |
| Memory | Open WebUI and one workshop E: 4.1 GiB left of 7.8 |
2026-10-26, Codespace, image not noted |
| Memory | Open WebUI alone: 2.1 to 2.2 GiB used of 7.8 | 2026-10-30, Codespace, image 2026.10.32 |
Second test, 2026-11-02 (image 2026.10.45, container, claude -p, ingest and lint only):
ingest 46 s, $0.17, the same contradiction plus a new one inside the Benevento document (plan
"2025-2027" on the title page, "2026-2028" in the text). The "same contradiction" (15 October in
document 80, 30 September in document 71) compares two different periods: the campaign's activities,
set by a circular of the minister that decree 212 "takes note of", and the period of high danger
that the decree made public. Every ingest so far wrote it (four runs, Haiku included, 2026-11-05):
the page now asks pairs to check it; lint 72 s, $0.34, changed nothing, 5
proposals, no new contradiction (facilitator/workshops/measures/2026-11-02-wiki/). The lint cost
$0.34 to $0.55 in two runs.
Cost per pair: about $1.05 for what was measured on 2026-10-30, more than half of it for the lint (ingest
$0.14, lint $0.55, replay E $0.36, judge $0.003). Not measured: applying a proposal of the lint, a
question in Open WebUI, question 12, and the paid "to go further" ($0.20 on 2026-10-21 for the
distances). Each pair's $3 share of the room's limit leaves a margin.
The chatbot of Part 5 in Open WebUI (image 2026.11.4). chatbot-export now keeps wiki/index.md
as wiki__index.md (the map) and, in the package only, rewrites each link between wiki pages to the name
of its file (wiki__<page>.md): 67 of 67 links with the kit's wiki, 0 before; wiki/ itself is not
changed. openwebui chatbot adds to the instructions one sentence on the map and on list_knowledge
with count 200 (without it, the list stops at 50 files: 50 of 56 listed, not always the same ones).
The chatbot also gets the pair's own skills (not ingest, lint, summarize, validate-summary,
source-note) and the live tool of workshop 4. Measured only with a fake model, without network, on
2026-11-04 (facilitator/openwebui-prototypes/next-image/, skills-rag/): the fake model follows a link
by the file name and opens the page. With a small local model (qwen3:8b through Ollama, 2026-11-05,
facilitator/openwebui-prototypes/real-model-local/results-thinkpad/): it used count 200 each time it
listed the knowledge, but never opened wiki__index.md. Asked "Which pages of your wiki are about Italy?",
it answered from the file names alone, mixed wiki pages with summaries, and missed the Vesuvius page (one
question). With the skill rumour-check attached, a rumour got an empty answer 3 times out of 3: the
model took the skill for a tool (same folder, section 5; see the facilitator sheet of workshop 3). Its first fire_danger call after the start was not run by Open WebUI (empty answer, cause
unknown), the next two were. With Claude Sonnet 5.5 (one try each, 2026-11-05,
facilitator/openwebui-prototypes/real-model-local/results-claude/): it searched for wiki__index, listed with
count 200, opened the map and named the 4 pages about Italy, Vesuvius included ($0.055); fire_danger
ran at the first call and gave no level, as why_no_level says ($0.021). The wiki goes into the knowledge as it is: the lint of Part 3 before a group
uses it stays the rule.
TO BE COMPLETED.
2026.11.4, container, each prompt accepted by a script, one try;
facilitator/workshops/measures/2026-11-05-wiki-sonnet/): Skill(ingest) and Skill(lint) show up
in a conversation; ingest asked 10 times (create, edits of the new page and of the Montoro page,
index, log), 35 s, about 33,500 tokens, about $0.17; the lint asked once (a shell command that reads
the pages: "A variable in this command can't be checked before it runs", then Do you want to proceed?), 1 min 55, about 149,000 tokens (74%), about $0.80 (estimates from the tokens, list
prices). The Open WebUI chatbot answering from the wiki with Claude: see above. Haiku 5.5 for
ingest and lint (2026-11-05, …/2026-11-05-wiki-haiku/): $0.009 and $0.16, but a wrong count
(15 municipalities for 16 names), no link from the existing pages, and two of the lint's eight
findings wrong: keep Sonnet.2026.11.5, container, Sonnet, one try, all in -p: in a conversation,
workshop E was refused with "Credit balance too low" four times out of four while -p worked,
cause not found; facilitator/workshops/measures/2026-11-06-wiki-suite/): a first lint ($0.38,
8 contradictions, nothing new), one proposal applied (the "30%" of ECHO for Ukraine, $0.044), a
second lint ($0.58, 12 contradictions, 3 new; 4 new findings checked against the sources, all
right). Both lints still take the 15 October / 30 September dates for a contradiction (see the box
of the page). Then replay E 21/22 ($0.37) and replay F 22/22 ($0.39) from the judge.…/2026-11-06-ingest-all/): ingest of the 20 documents into an empty wiki,
without Benevento, one call each, nobody checking in between: 13 minutes, $2.56 ($0.06 to $0.29
per document), 28 pages (18 in the pack). It names 5 of the 9 contradictions of the pack, the one of the 2.2 million hectares (documents 10 and
90) included; it writes the others side by side without naming them, and reconciles one ("27 June …
in application from 1 July"), which rule 3 forbids. The wiki of the pack was corrected by hand after
its own ingestion: a person reads every page before a group uses it.…/2026-11-06-lint-ingest-all/, image 2026.11.6): lint the wiki on these 28 pages,
$0.81, 2 min 11, changed nothing. It names one more contradiction (gas barbecues, articles 2.1 and 5.2), drops
the date of 24EB441 after reading only document 33, and misses the August share and the mayor or the préfet; it
says it did not read document 12. Then replay E on this wiki: 21/22 from the judge (question 6: 1), $0.38. One
run each. A lint does not catch what the ingestion missed: a clean report does not prove that everything was read.--from
with ingest summaries/8N-…; question 11 for a pair that ingested only its own document (it should
come from corpus/text/80); question 12; the reference columns of the facilitator; ingest, lint and replay E with OpenCode and a Claude key.2026.10.34 (last seen on 2026.10.10); the drawing of wiki/graph.md in the preview;
openwebui start in yal-workshop-wiki; openwebui chatbot with 56 to 59 files on Open WebUI
0.12.5; memory with Open WebUI, replay E and a conversation; the creation time of
yal-workshop-wiki.During Part 2. Watch four moments:
--highlight 81 is only an example. A pair that ingested Benevento types 80.After Part 2. The "Think about it": "Why must the wiki never choose one figure on its own?" Expected answer: the model does not know which source is right. If it picks one, the other disappears without a trace, and nobody can decide any more. A young person who quotes the chatbot then meets a listener who has read the other source.
The test of 2026-10-30 gives a good story: ingesting Benevento raised a contradiction that nobody had
placed there. Montoro and Benevento give two different end dates for a period, and both refer to
Decreto Dirigenziale 212. These may be two different things (a period of grave danger, and the
fire-fighting activities): the skill tells the model not to settle it, so the log keeps both, and a
person decides. The answer to question 11 mentions them without changing the right date. Already
seen at the rehearsal of 2026-10-21: expect it for Benevento. The summarize of workshop 3 also
pointed out a contradiction with summary 71 (Montoro), probably the same one. Tell the room that it
happened. The page prints both dates (box "Check the contradiction it found") as two different
periods: 15 October for the campaign's activities, 30 September for high danger. Let pairs read the
two sentences and decide. Remind the room that neither date is the answer to question 11 (6 July to
27 September 2026).
Then the difference between a wiki page and a summary: a page has a generated line and no
verified line. Nobody signed it.
After Part 3: reading the report. On 2026-10-30, with Benevento in the wiki: 19 pages, 0 problem of
header, footnote or link, no orphan, 4 new contradictions, about ten omissions or rewordings, 8
proposals. Another run may differ: we have only this report on the image of the day. Only one
remark was checked by hand (the plan DGR 368 is at line 53 of corpus/text/71, and is missing from
the Montoro page).
| Proposals of that report | How to read them |
|---|---|
| 2, 3, and in 5 the date of 30/04/2026, the Flemish measures and the "30%" | Gaps of the prebuilt wiki, left on purpose as material for the lint, and already listed at the rehearsal (facilitator/SETUP.md, row 10 bis). Seen in both reports so far: expect them again. |
| 1 | Comes from Benevento: only for pairs that ingested it. |
| In 5, the Montoro plan and clauses | A gap of the prebuilt wiki that is not in the rehearsal's list. The plan DGR 368 was checked by hand (corpus/text/71, line 53); the clauses were not. |
| 4, 6 | Not checked. They look closer to a matter of taste: notes about the page, or the scope of a word. |
7, the generated.by line, the headings of the index |
Cosmetic. A good example of what to refuse. |
| 8 | Reading three more sources: the lint says itself that it did not read them. Honest, and a reason not to trust a clean report. |
The easiest finding to check in a pair: the date of 30/04/2026, one Ctrl+F in corpus/text/32
(line 40). Not all pairs will get the same report.
The "Think about it": "Why does the skill forbid the model to fix anything by itself?" Expected answer: a fix can remove or smooth a figure, and nobody would notice. In an early test, before the rules were hardened, Haiku rewrote a contradiction and Sonnet left it out, and the lint saw nothing. Hence the rules of the skills: compare each figure and its scope, never omit, never reword.
During Part 4. Open with the two maps side by side. This is the first time the room sees the full
wiki-graph: this was kept for today on purpose (decision of 2026-10-22). Say: in the left map, each
arrow is written in a page; in the right map, a tool drew ideas, and guessed some links.
If a pair asks what Cohesion: 0.23 means in the report. It is the lines drawn inside the group,
divided by the lines possible. Between n nodes, n(n-1)/2 lines are possible. Community 3 has 13 nodes,
so 78 possible lines. 18 are drawn (counted in graph.json; the report does not print this number):
18 / 78 = 0.23. The score comes after the grouping: it is not what makes the groups. Do not rank
groups with it, because a small group scores high more easily: in community 7 (0.50) a node has 2
inside lines on average, in community 3 (0.23) it has 2.8. Use community 3 for the sum: for
communities 0, 2 and 4, the report prints Nodes (24), Nodes (17) and Nodes (10) where
graph.json has 25, 18 and 11 (the report does not count the nodes of type rationale), so the sum
does not come out.
After Part 4: the verdicts (let the pairs reach them).
semantically_similar_to, confidence 0.85; 0.75 in
the graph of 2026-10-21). The
figures of the node are right: corpus/text/71, line 113 (metri venti, for owners of
urban-rural interface areas and managers of campsites, hotels and tourist structures) and line 124
(5 metri, around uncultivated or fallow land); corpus/text/22, lines 74 and 113 (50 m around the
house and installations, 10 m along access roads) and line 111 (land less than 200 m from classified
forests, moors, maquis or garrigues). The link holds as an idea, clearing vegetation so that a fire
does not spread, and not as an equivalence: two countries, different people concerned, different
figures, and different objects (a radius around a settlement, a strip around a field, a depth around a
building). Good sentence for the room: similar is not the same.semantically_similar_to, 0.75; in the graph of 2026-10-21,
conceptually_related_to, 0.65, its lowest score), for quick pairs. Same amount, two offences: burning green waste (corpus/text/23, line 120), and ignoring the
prevention instructions of the authorities (corpus/text/20, line 154). The link says little.The "Think about it": "What does it cost to keep a graph up to date, and who checks what the model
guessed?" Expected answer: a model reads everything again (about 10 minutes and 163,000 tokens for 18
pages on 2026-10-21; about 7 min 30 with Opus on 2026-11-02, tokens not measured 7.86 million tokens, most of them read from the cache; $1.40 on 2026-11-06, because Opus started from the scripts of 2026-11-02), and all the
guesses are new, so a person has to check them again. The report says itself that
we may not need a graph. The modes E and F show no gain on 18 pages: a graph helps people who explore.
It may help a much larger wiki: we have not measured that. Do not sell it as an improvement of the
chatbot.
Part 5: reading scores.md. The columns are sparse: Bare and A were scored on all questions, B and
RAG too, C on 3, 11, 12 (and 8 when workshop 3 or 4 scored it), D on 8, 11, 12, E on 3, 10, 11, 12. So
the Total row means nothing between columns. Read the rows: 3 runs across nearly every column, 10
across Bare, A, B, RAG and E, 11 and 12 across C, D and E. The same score with a smaller gauge is a
result. Your reference columns are TO BE COMPLETED: until then, use the one run of 2026-10-30, 2
for questions 3, 10 and 11 in mode E, no save, no question 12. The proposal score E 3 10 11 12 is
ours: the plan names no questions for this workshop.
Part 5: the chatbot you built. Project the table of the series, and read the last column aloud: you,
nobody or a search engine, the model and a person, the model and you, the model and a person. The more
the model does by itself, the more it matters that a person checks. Say what does not travel in
<name>.zip: the live tool of workshop 4 (a program, not a document; with a group, the Open WebUI
guide gives it back, see "Reuse with young people") and the kit's skills (ingest,
lint, summarize, validate-summary, source-note); the pair's own skills of workshop 3 do. And what never ends: the same test questions, before and
after each change.
Then point each person to one next step:
| If you want... | Go to |
|---|---|
| A chatbot for a group | Open WebUI guide, Part 4 - One chatbot per topic, activity or project |
| Nothing leaving your organisation's machine | Local AI guide, and What we measured |
| The same pieces in a folder, with other models | OpenCode guide, Your own chatbot, and Part 8 - Test and compare models |
| The repositories of the series on your own computer | Dev containers guide, Part 3 - Open a repository (not for a group) |
| An assistant that prepares your sessions | Hermes guide, Part 3 - Memory: same question, who decides what is kept? For you, never with young people. |
EXTRACTED and INFERRED, and who checks?What a good answer contains: (1) the model does not decide which source is right: a person must see the disagreement; (2) written in a page, against guessed by a model; a person checks, in the sources; (3) the model wrote it, beforehand; a person signs the documents, arbitrates and checks.
| Problem | What to do |
|---|---|
| A pair has no save | workshop-start 5, then ingest corpus/80-benevento-aib-2026.pdf into the wiki. It has Feurisson, and the same steps. Say it plainly: nobody signed this one. |
A pair restored an old save (saved on is not the end of workshop 4) |
workshop-start 5 --from the newest zip, its exact name between quotes. |
| A pair has a save, but no signed summary | It ingests the official text of its own document (corpus/8N-…), or Benevento. Never a draft. A pair that missed workshop 3 has no document of its own: Benevento, which workshop-start 4 or 5 added to corpus/: look for corpus/8N-benevento-aib-2026.pdf. |
| The lint is too long, or a pair is stuck | Stop it. Show the section Proposals of rapport-lint.md, or of your own report, and let the pair choose from it. Project that section only. |
A pair does not get Skill(ingest) or Skill(lint) |
Check that .claude/skills/ingest/ and lint/ exist; leave and start workshop E again; ask with the exact words of the page. |
ingest gives nothing in orange |
No page cites a document matching '…' yet: check the number, then read the report of the skill. |
The drawing of wiki/graph.md does not show |
Open the preview with Ctrl+Shift+V. If it fails, read the list Pages that cite … at the bottom of the file: it gives the same answer. |
| The map does not open | Ports tab, 8765, Open in Browser. Otherwise wiki-explore path, wiki-explore explain and GRAPH_REPORT.md work in the terminal and without the internet, and the check by hand needs no map. Show the graph slide of the deck, at its small size: full screen, the image is readable and holds expected answers. |
openwebui start has not finished at the end of Part 2 |
Open a third terminal for replay E. |
| Open WebUI is not ready in Part 5 | Skip openwebui chatbot, and show your own chatbot on the big screen. chatbot-export works without it. |
| A pair has no question 12 | score E 3 10 11. |
| A pair uses OpenCode | It is asked before every shell command and every file it writes: same steps, more "allow" answers. An OpenRouter key takes the same steps, judge and Open WebUI included; openwebui chatbot with this key imported the chatbot on 2026-11-02, answers not measured (measures, section a), and was tried again on 2026-11-05, on a laptop, not in a Codespace: 4 min 07, wiki map and fire danger tool attached, and asked which wiki pages are about Italy, DeepSeek V4 Flash opened the wiki's index (wiki__index.md) and found the 4 pages (measures). Its ingest and lint were measured on 2026-11-02, with DeepSeek v4 Flash and Mistral Small 2603 (measures, section e): neither ingest followed the schema of the wiki (log, links, contradictions), and the lint of Mistral Small 2603 gave a false clean report (see the warning at the top of this sheet). A second run of DeepSeek v4 Flash, on 2026-11-05, wrote the page, a line in index.md, two lines in log.md and a link from the page on Italy; it noted no contradiction and declared one source it never cited, which its own lint then found, with 9 contradictions, each with its page (measures): one run each, so the same model does better or worse from one run to the next. If they stall, show your own report. replay runs OpenCode in a copy of the kit without the test material: no warning on 61 answers in modes bare to E, on 2026-11-01 (measures); mode F: (not tested). An answer marked "Warning: the model looked at the test material" gets 0, as test-questions.md and the troubleshooting of SETUP.md say. |
<name>.zip and my-work.zip: a Codespace stops by itself, and it
is deleted after 30 days of inactivity.If someone asks about Projects and Gems. On personal accounts, Google makes Gems read-only on 2026-11-17 and removes them on 2026-12-17 (Workspace for Education: June 2027), and claude.ai is for adults only (18+). So neither is a place to rebuild a chatbot with 12 to 15-year-olds for now.
Reuse with young people. The wiki is one way of filling the Knowledge box: a group brings
local documents, an adult checks them, and the chatbot receives clean pages, each fact with its source.
The habit is worth more than the tool: two sources that disagree are shown side by side, and a link
guessed by a model is checked by a person. A line Contradiction of the log can start a discussion
with a group, with no computer: two sources, the same question, who is right, and how would you
know? To build the chatbot: Open WebUI guide, Step 1 - Gather your documents,
then Part 3 - Test before you open.
Before any use with a group, read The participants' data.
To open it to the group: Part 5 - Open it to your group,
and never port 3000 of a Codespace. The live tool of workshop 4 is not in <name>.zip, but the
guide has it from image 2026.11.8 (not published yet): after openwebui import, you tick Fire
danger today in the chatbot (the import never does it), and you fill in the guide's own
mcp/fire-danger.json on the morning of each session
(Step 5 - Add the live tool;
measures and limits: the facilitator sheet of workshop 4, "Reuse with young people").