Buzz · proof

Evidence

What Buzz has shown on real Bee data and how to check it yourself. Every claim names the receipt it comes from; this page is built from files in the repository.

For scripts and agents: evidence.json, the same claims, criteria, captions, receipts and word-decision numbers as data. Each claim and section has an anchor, for example #claim-delineate-recurrence.

What counted

delineate counted on 16 Oct in Bee conversation 11319904, then again on 20 Oct in conversation 11437353: the same word used on a later day, synced to the beta and read back.
Receipts: FIRST_REAL_EMBED 20 Oct at 4386915 · SYNC_PUSH_11437353_HOSTED_EMBED 20 Oct at 3a338fc
A 14 Oct session that used delineate did not count. The builder read a word list aloud to test the pipeline; it was held as not assessable at the outset, then recorded as practice on 15 Oct, and practice never counts as use.
Receipts: REHEARSAL_AS_PRACTICE 15 Oct at 9584091 · SYNC_PUSH_2 16 Oct at 74d9a09 · JUDGE_WALK_LIVE 20 Oct at 36677bc

Checked uses in the builder's own conversations, not a measured learning gain. Three kinds of data stay apart here: real (the builder's own Bee conversations, checked and read back), example data(synthetic and labelled; it shows the mechanism only) and the word brain's model run on authored examples (19 Oct, no measured advantage over the deterministic pick), plus its live pick from the builder's stored data on his own computer (22 Oct).

Four ways to check it

A. No sign-in: real week and evidence The builder's own Bee results as a dated record: each conversation, each check and why it counted or not.
B. Try the app (invited judge copy) A copy of the same real results in the app. The email and password are in the Devpost testing instructions.
C. Reproduce on example data No Bee or AWS needed: npm run demo:judge, then npm run reproduce:fixture sends five labelled example conversations through the real checking rules and store and prints each outcome: counted, used again, practice refused, judge unsure, and a replay that changes nothing.
D. Wear a Bee? Try it on your own data npm run try:own-bee on your computer, or type /try-buzz in Claude Code. Needs an invite and the Bedrock key from the testing instructions.

Quality of the Idea: Idea

A word learned in an app rarely makes it into a real conversation, and an app cannot hear your day. Buzz uses Bee's record of your own conversations to show when a word you chose turned up in your speech, checked and dated, and when it came back on another day. Practice is optional and never counts as use.

Check it: The lead result above is the idea working once on real speech; the routes below show it on your own terms. Source and tests: README.md, submission/DEVPOST_DRAFT.md.

Design: Experience

Pick a word and hear it. Practise if you like. Bee captures your day; Buzz checks only your own speech and shows each counted use with its conversation, day and reason, then marks it Used again when it returns on another day.

On 21 Oct the builder ran try:own-bee end to end on his own computer: wearer and time zone, the Nova 2 Lite check, a preview of ids and counts, an approved send, and a read-back matching 16 of 16 conversations. The app then showed the 20 Oct delineate use as counted, with its checks.
Receipts: OWN_BEE_RUN 21 Oct at b1bc485 · OWN_BEE_VISIBLE 21 Oct at edd4c0b

Example data: the mitigate word page lists two uses, Used again on 21 Oct and Counted on 19 Oct. — **Example data** · The word page for mitigate, Your uses tab, phone width, light. Two labelled example conversations used mitigate on 19 and 21 Oct, then the word page was opened. It shows: two uses: Used again, in another conversation on 21 Oct (shown as Today); Counted, in a checked conversation on Mon 19 Oct; practice offered as optional and never counted as use. Why it matters: One screen tells the wearer where a word was used and when it came back, with practice kept apart from use. Captured 21 Oct 2026, local browser suite on 21 Oct, pinned copy in docs/evidence/2026-10-21-fixture; made by `apps/web/e2e/v2-onescreen.spec.ts`.

Check it: Use route C (Reproduce on example data) and open a word's Your uses tab, or use route B (Try the app) and tap See both days. Source and tests: apps/web/src/v2/screens/Words.tsx, apps/web/e2e/v2-onescreen.spec.ts.

Potential Impact: Value

For professionals building words for work: seeing that a word reached real conversations, dated, is something a quiz or a streak cannot show. Observed so far on one wearer's Bee:

One wearer over three checked days. Whether seeing this helps a word last is not yet measured.

substantiate, delineate and preclude each qualified on 16 Oct in conversation 11319904, checked by the model and read back from the beta.
Receipts: REJUDGE_EXECUTE 16 Oct at 6ff0d23 · BETA_READBACK_COUNTED 16 Oct at 6ff0d23
Across three real conversations on three days (16, 19 and 20 Oct), Buzz counted nine uses, 3 + 3 + 3, for one wearer. Checked uses, not a measured learning gain.
Receipts: REJUDGE_EXECUTE 16 Oct at 6ff0d23 · NEW_REAL_CONVERSATION_JUDGE_INGEST_SYNC 19 Oct at 7e22127 · REJUDGE_11437353 20 Oct at 4386915
On 19 Oct a new conversation gave three qualified uses (material, tenable, implication), two refused and one that could not be assessed.
Receipts: NEW_REAL_CONVERSATION_JUDGE_INGEST_SYNC 19 Oct at 7e22127
The wearer reviewed all 24 of Buzz's real decisions. Owner-reported (the review did not record which window was shown): 22 of 23 judged agreed, and all 9 counted uses were agreed as counted. The one miss: a 19 Oct use of mitigate that Buzz could not assess was the wrong sense. The reviewer also built Buzz, so this is not independent.
Receipts: DECISION_REVIEW 21 Oct at 6107b13

The hosted beta's 19 Oct day record shows three counted uses from one real conversation, one not yet judged and two refused. — **Real data** · The hosted beta's day record for Monday 19 Oct, phone width, dark. After a real Bee conversation on 19 Oct (11398956) was checked, synced to the beta and read back, the day record was opened. It shows: material, tenable and implication counted from conversation 11398956; mitigate could not be judged yet; constraint and feasible did not count for lack of context; no day score yet. Why it matters: Real speech produced dated, checked uses, and the refusals are shown beside them rather than hidden. Captured 19 Oct 2026, processed by source 7e22127; deployed app identity not recorded; receipt `NEW_REAL_CONVERSATION_JUDGE_INGEST_SYNC` (2026-10-19T08:39:07Z).

The same 19 Oct day record at 360 px shows the same three counted uses. — **Real data** · The same day record at 360 px, light. The same record opened at a narrower phone width, light theme. It shows: the same three counted uses; the run's browser record (browser.json) also counted three after a reload at both widths. Why it matters: The same server record reads the same at a second width and theme; the run also recorded it surviving a reload. Captured 19 Oct 2026, processed by source 7e22127; deployed app identity not recorded; receipt `NEW_REAL_CONVERSATION_JUDGE_INGEST_SYNC` (2026-10-19T08:39:07Z).

Check it: Open route A (the real week): each real conversation, each check and why it counted or not. Source and tests: docs/weekend/LIVE_EVIDENCE.jsonl, docs/evidence/2026-10-19-real-use/readback.json.

Technical Implementation: Technical proof

Bee is read through the official CLI on the wearer's computer; only the wearer's own speech is checked, by Amazon Nova 2 Lite on Bedrock; decisions are stored, synced to the beta and read back. Refusals are recorded as carefully as passes: practice never counts, and a use without enough context is refused. The one write back to Bee is an approved fact, read back byte for byte. A claim on this page fails closed: if its receipt is missing, duplicated or a later run of its step did not pass, it shows as unavailable. The architecture shows each path and its proof.

GitHub links point at the private repository and need reviewer access; the same rows are in the repository's docs/weekend ledgers and in evidence.json.

A short speech window (240 characters each side of a word) goes from the paired computer to Amazon Nova 2 Lite on Bedrock for checking. The hosted beta receives only ids, days, statuses, reason codes and revisions, and refuses other fields.
Receipts: OWN_BEE_RUN 21 Oct at b1bc485 · SYNC_PUSH_3 16 Oct at 6ff0d23
Buzz wrote one approved training fact to Bee on 20 Oct and read it back byte for byte. On 22 Oct it updated the same fact to the approved closing text, which ends it, and read that back too.
Receipts: FACT_CREATE_READBACK 20 Oct at 4386915 · FACT_END_READBACK 22 Oct at d79dd85
The judge account holds a copy of these real results and was walked signed in on the live beta on 20 Oct.
Receipts: SEED_JUDGE_EXAMPLE 20 Oct at cfbe381 · JUDGE_WALK_LIVE 20 Oct at 36677bc
On 23 Oct, read-only checks matched the existing a414e63 release across CloudFormation, the Api Lambda commit, the public build stamp, evidence.json and the no-JavaScript evidence page. The stack completed its update on 22 Oct at 17:42:58 UTC; the Api is Active and configured for Nova 2 Lite. Submission-source promotion and deployment are tracked separately.
Receipts: RECOVER_DEPLOY_A414E63_1023 23 Oct at a414e63398eaffcb54bf21a2844b513fb7d180d1
On the builder's computer the Bee CLI passed all five checks: identity, changes, recall, the live stream and the approved fact read-back. A Bedrock key for judges reaches only Nova 2 Lite in us-east-2 and stops working on 21 Nov.
Receipts: BEE_DOCTOR_FULL 21 Oct at 293236c · JUDGE_KEY_CHECK 21 Oct at 12f9803
On 22 Oct, in the local paired-computer app, a Strands agent made three Nova 2 Lite calls and selected feasible from 18 candidates ranked by goal and word-use history. The saved pick was read back after a model-free restart and remained visible after reload.
Receipts: BRAIN_THINK_REAL_STORE 22 Oct at ceafde6 · BRAIN_PICK_VISIBLE_RELOAD 22 Oct at d79dd85
On 22 Oct, a local paired-computer builder and critic made two live Nova 2 Lite calls and produced six requested vocabulary cards: contingency, consensus, escalate, prioritise, viable and caveat. All six were accepted, read back after a model-free restart and visible in Words after reload.
Receipts: HONEY_PACK_ACCEPTED_1022 22 Oct at 4ec7101

Buzz's connection check on 1 Oct: Bee identity, changes and conversations answered OK; stream and fact write not run. — **Real data** · Buzz's connection check on the paired desktop. The Bee CLI was run on the builder's paired computer: bee me, bee changed, bee conversations list and get. It shows: Identity, Reconcile and Recall answered OK; Capture (bee stream) and Write (bee facts) were not run that day. Why it matters: Buzz reads Bee through the official CLI on the wearer's own computer; the one write to Bee came later, on 20 Oct (claim above). Captured 1 Oct 2026, the paired desktop's connection check; app identity not recorded; recorded in `docs/evidence/local-connection-proof/README.md`.

Amazon DynamoDB: the LEDGER item in the BuzzPilot table, with its key and body masked. — **AWS Console** · Amazon DynamoDB, the BuzzPilot table, one item open. After the 20 Oct sync of conversation 11437353, the wearer's LEDGER item was opened in the table. It shows: the item with sort key LEDGER in the BuzzPilot table; its key and body are masked because they hold Bee source ids. Why it matters: Checked decisions live in DynamoDB; the receipt records the sync written, read back to match, at ledger revision 28, with delineate's recurrence. Captured 20 Oct 2026, VB in the AWS Console, US East (Ohio), after the 11:32Z f12f037 deploy; public copy masked; receipt `SYNC_PUSH_11437353_HOSTED_EMBED` (2026-10-20T08:52:00Z).

Amazon CloudWatch: token usage charts for us.amazon.nova-2-lite-v1:0, 28 Sep to 19 Oct. — **AWS Console** · Amazon CloudWatch GenAI Observability, model invocations. The token usage view was filtered to the one model Buzz calls. It shows: input, output and total token charts for us.amazon.nova-2-lite-v1:0 from 28 Sep to 19 Oct. Why it matters: Checks ran on Amazon Bedrock; the receipt names one run, 6 live judge calls on 19 Oct with 0 failures. A token chart shows use, not quality. Captured 20 Oct 2026, VB in the AWS Console, US East (Ohio), after the 11:32Z f12f037 deploy; public copy masked; receipt `NEW_REAL_CONVERSATION_JUDGE_INGEST_SYNC` (2026-10-19T08:39:07Z).

AWS CloudFormation: stack BuzzPilot UPDATE_COMPLETE, with output Commit f12f037 and the beta URL. — **AWS Console** · AWS CloudFormation, stack BuzzPilot, Outputs tab. VB deployed the beta with AWS CDK from a clean checkout and opened the stack's outputs. It shows: UPDATE_COMPLETE; outputs Commit f12f037 and the beta URL; the Forager function and user pool id are masked. Why it matters: The hosted beta is one CDK stack that records which commit it was built from, so a deploy can be checked against its source. Captured 20 Oct 2026, the deploy shown, f12f037; later releases are read back in the release claim; receipt `DEPLOY_F12F037` (2026-10-20T11:32:52Z).

AWS Lambda: the BuzzPilot API function with 2,178 invocations and 0 errors. — **AWS Console** · AWS Lambda, the BuzzPilot API function, CloudWatch metrics. The API function's monitoring tab was opened after the deploy and its smoke check. It shows: 2,178 invocations; average duration 619 ms; errors at most 0 and success rate at least 100% over the period shown. Why it matters: The beta's API has been serving real requests without errors; the smoke receipt shows a signed-out read refused with 401. Captured 20 Oct 2026, VB in the AWS Console, US East (Ohio), after the 11:32Z f12f037 deploy; public copy masked; receipt `SMOKE_F12F037` (2026-10-20T11:40:00Z).

Check it: npm run reproduce:fixture sends five labelled conversations through the real checking path, including a refused practice and an unsure judge; every claim here fails closed if its receipt is missing, duplicated or later failed. Source and tests: apps/web/src/evidence/claims.ts, apps/web/tests/evidence-claims.test.ts, packages/buzz_domain/buzz_domain/adoption/ambient.py, packages/buzz_worker/buzz_worker/brain.py.

At a glance

Ran live

Real Bee read REAL_VERIFIED · 2026-10-16
AWS Dream Cycle run REAL_VERIFIED · 2026-10-22
Bee write-back REAL_VERIFIED · 2026-10-22

Checked locally

Browser check LOCAL_VERIFIED · 2026-10-01
Privacy boundaries LOCAL_VERIFIED · 2026-09-30

Not run yet

v11 pillars evaluation (not the word decisions) NOT_RUN · 2026-10-01

Each line is the status word in that receipt’s header. Not run is neither passed nor failed.

Latest receipts

From docs/weekend/LIVE_EVIDENCE.jsonl (live runs) and docs/weekend/EVIDENCE.jsonl (builds). A live row shows the word its result starts with; a build row shows its maturity, or Unknown when it records none. Fixture verified means tests and served checks on example data, not a live run. The same records and claims, for scripts and agents: evidence.json.

PASS 2026-10-23 · live · JUDGE_WALK_A414E63_1023 · a414e63398eaffcb54bf21a2844b513fb7d180d1
all three: Home tour Skip restores focus; See both days opens delineate; both 16 and 20 October records opened and conversation IDs 11319904 and 11437353 checked; 14 October practice remains excluded.
Limit: Agent-operated browser acceptance with the prepared judge account, not an unfamiliar-person test or a physical-device test.
PASS 2026-10-23 · live · RECOVER_DEPLOY_A414E63_1023 · a414e63398eaffcb54bf21a2844b513fb7d180d1
recovered existing deployment identity a414e63.
Limit: Read-back of an already performed release, not a deploy executed in this session.
PASS 2026-10-22 · live · SIGNIN_JUDGE_NOTICE_LIVE · e9bc780
the notice 'Judging Buzz?
Limit: the GitHub repository is private, so the README link opens only for accounts granted access; the same instructions go in Devpost's private testing field
PASS 2026-10-22 · live · DEPLOY_E9BC780 · e9bc780
UPDATE_COMPLETE 2026-10-22T12:12:38Z, output Commit e9bc780 (was ddfb7a9); diff was Api and Forager code, BUZZ_COMMIT ddfb7a9->e9bc780, site files and the Commit output; /BUILD_COMMIT and evidence.json source_commit e9bc7801060717ad5524dd4e2527186e09352391; smoke PASS (config 200, unauthenticated 401, fixture 404, test sign-in 403, /v2.html 200)
Limit: Api Lambda BUZZ_COMMIT not read (needs VB in the Console).
PASS 2026-10-22 · live · VERIFY_SCOPED_E9BC780 · e9bc780
both: web types clean; web unit 674 of 674 in 97 files; package PASS
Limit: scoped: since the full gate VERIFY_4EC7101 only apps/web (SignIn.tsx, app.css, one new test), TRY_WITH_YOUR_AGENT.md and docs changed; no Python, script or infra file changed, so pytest, mypy and import steps were not rerun
PASS 2026-10-22 · live · JUDGE_WALK_DDFB7A9 · ddfb7a9
in 2 of 3 runs, one finding: one Skip at both widths with focus on the 'Open “get to the heart of”' link; See both days opens delineate with Tue 20 Oct Used again, Fri 16 Oct Counted, Wed 14 Oct practice session Didn't count, kept after reload; Replay onboarding opens #/start?from=you and keeps replay mode on reload; hosted Next word shows the judge copy's selector pick 'clarify'; zero POST/PUT/PATCH/DELETE, no page errors, no horizontal overflow; evidence.html and architecture.html readable without JS.
Limit: mobile-width headless browser, not a physical phone; the tour finding was seen once in three runs and not diagnosed; the replay Skip link was checked on ceafde6, not separately here
Unknown 2026-10-22 · live · JUDGE_WALK_DDFB7A9_TIME_CORRECTION · 6e7f4ab
the walk's three runs finished by about 2026-10-22T11:47Z and the row was committed in 6e7f4ab at about 11:48Z; its at value 11:50:00Z was entered by hand and is later than both.
Limit: the original row is kept (append-only)
PASS 2026-10-22 · live · IDENTITY_DDFB7A9 · ddfb7a9
4 of 5, one pending: (a) UPDATE_COMPLETE, Commit ddfb7a9; (b) Api Lambda BUZZ_COMMIT NOT READ (needs VB in the Console; the diff set it to ddfb7a9); (c) /BUILD_COMMIT ddfb7a9e80deb447cd7f4ecaf6d4147e33b3e80e; (d) evidence.json source_commit ddfb7a9e80deb447cd7f4ecaf6d4147e33b3e80e, 12 claims; (e) evidence.html without JS contains 'delineate counted on 16 Oct' and cites FACT_END_READBACK, and no longer says the fact's ending has not run; architecture.html no longer says the word brain has not run on real recommendations or that five reads agree on b4ba775.
Limit: the buzz profile has no lambda:GetFunctionConfiguration; (b) waits for VB
PASS 2026-10-22 · live · SMOKE_DDFB7A9 · ddfb7a9
config 200 private-pilot; unauthenticated read 401; fixture capture 404; test sign-in 403; /v2.html 200
Limit: served routes and auth guards only
PASS 2026-10-22 · live · DEPLOY_DDFB7A9 · ddfb7a9
BuzzPilot UPDATE_COMPLETE 2026-10-22T11:42:29Z, output Commit ddfb7a9 (was ceafde6).
Limit: ddfb7a9 differs from the gated 4ec7101 only in docs (README status line, AWS_LIVE.md, ledger rows); the gate is VERIFY_4EC7101.

Receipts

Word decisions: owner-reviewed

Owner-reported: 24 reviewed of 24 real word decisions. 22 of 23 judged agreed (96%); 1 disagreed, 1 unsure and left out. All 9 counted uses were agreed as counted (100%).

Evidence-backed: none recorded. The review did not record which window the wearer saw, so none of it is checked against its source.

One wearer, one paired Bee, and the wearer built Buzz: not independent. This is not the v11 evaluation report below.

File: docs/evidence/DECISION_AGREEMENT.md
Status: MEASURED
Report date: 2026-10-21
Decision set: 6adc57adff3f (24 decisions)
Judge: judge ambient-judge-0.3, catalogue cb51af1d20aa, model RUN

Bee live read

REAL_VERIFIED

File: docs/evidence/BEE_LIVE.md
Produced by: scripts/bee_read_receipt.py and scripts/serve_real.py --receipt, on the paired machine
Last run: 2026-10-16 08:54 UTC

Full receipt

Written by scripts/bee_read_receipt.py on the paired machine. Counts, states and hash prefixes only: no transcript text, speaker names or Bee ids. revision is the first 12 hex characters of the canonical payload hash (buzz_bridge.canonical.revision_hash), so a later ingest can be matched to this read without exposing speech.

Field	Value
Read at	2026-09-29 11:50 UTC
CLI	@beeai/cli 0.7.3
Account alias	`bee_94bdb121cc9e` (hash of the Bee account id)
Identity (`bee me`)	OK
Reconcile (`bee changed`, bootstrap window)	OK: 2 changed conversations, cursor returned
Sources read (`bee conversations list/get`)	8 of 8
Cloud ingest (slice 08 task 9.1)	NOT_RUN
Write and read-back	NOT_RUN

#	State	Capture day (UTC)	Utterances	Timestamped	Speaker labels	Revision
1	COMPLETE	2026-09-28	46	46/46	2	`c0a9f5b3db3b`
2	COMPLETE	2026-09-28	6	6/6	2	`53ab6b5d919f`
3	COMPLETE	2026-09-26	4	4/4	1	`0722fe005213`
4	COMPLETE	2026-09-25	22	22/22	2	`84021c35ac4c`
5	COMPLETE	2026-09-24	4	4/4	2	`4a14689bb72e`
6	COMPLETE	2026-09-24	3	3/3	2	`f1bd3c83afa7`
7	COMPLETE	2026-09-21	93	93/93	2	`e8fec05ed27d`
8	COMPLETE	2026-09-21	7	7/7	2	`b28caf82c726`